Magnetic Tape Storage Technology

In state-of-the-art tape systems such as the latest generation of Linear Tape Open (LTO) format, LTO-9, the $12.65\mathrm{m}\mathrm{m}$ wide tape is organized in four data bands (DBs) which span the whole length of the tape, i.e., from the beginning of tape (BOT) to the end of tape (EOT). The four DBs are sandwiched between five dedicated servo bands (SBs), which contain a chevron-like TBS pattern [10] as shown in Figure 12. The factory pre-formatted TBS patterns are primarily used to estimate the lateral position (y-position) of the tape head relative to the tape [29], but also provide other essential servo parameters such as tape velocity, tape-to-head skew, tape width (TDS), SB identification, and tape longitudinal position (LPOS). The latter is encoded by means of position modulation of 4 out of 18 servo stripes that define a servo frame.

To write/read data to/from tape, the tape drive uses a stepper motor to laterally coarse position a three module tape head over the desired DB, such that the two servo readers are located on the two adjacent TBS patterns. Once the track-following control system has fine-positioned the tape head relative to the reference trajectory of the TBS pattern, and a desired tape LPOS has been reached by the tape transport control system, data can be written/read by means of the 32 active data transducers located between the servo readers. A set of 32 data tracks that are written in parallel along the full length of tape is called a wrap. Note that the 32 parallel tracks that define a wrap are equally distributed across the full width of the DB. Each track is located in one of 32 sub-DBs, each having a width of $83.25\mu \mathrm{m}$ corresponding to the pitch of the data transducers. The width of the servo patterns is nominally $93\mu \mathrm{m}$ and thus provides sufficient range to place the tracks at any desired lateral offset in each sub-DB.

LTO tape drives employ linear serpentine recording, where a large number of wraps are sequentially written to each DB, see Figure 12 (Right). Writing to a blank tape starts in DB 0. First, wrap 0 is written in the forward direction (BOT to EOT), placing 32 parallel tracks at the top of each sub-DB. Second, wrap 1 is written in backward direction (EOT to BOT), writing the tracks at the bottom of each sub-DB. Next, wrap 2 is written in forward direction, laterally offset by a track pitch relative to wrap 0 towards the sub-DB center, thereby shingling (partially overwriting) the previously written wrap 0 tracks to its final track width. Similarly, backward wrap 3 shingles wrap 1 to its final shingled track width, again offset towards the sub-DB center. This process of writing forward and reverse wraps in a spiraling-in manner continues until DB 0 is filled. Writing proceeds in the same manner subsequently in DBs 1, 2, and 3. The LTO-9 format specifies 70 wraps per DB and thus a total of 280 wraps per cartridge, offering a total of $18\mathrm{T}\mathrm{B}$ native cartridge capacity.

The use of shingled magnetic recording enables backward write compatibility. LTO-9 drives can read and write to both LTO-9 and LTO-8 media types. This backward-compatibility allows tape administrators to upgrade to the newest formats immediately without fully migrating all archives. To write the previous generation’s (wider) track width, the writer width significantly exceeds the current generation track pitch. On the other hand, the shingled recording nature prevents in-place overwriting of existing data because this would inevitably overwrite adjacent tracks/wraps. Tape systems are designed as append-only storage, where new data can be appended to where data writing previously ended, but the existing data is immutable. It is possible to partition the tape longitudinally or into groups of wraps, thereby creating independent storage pools and append points. Longitudinal partitioning requires a guard gap between partitions. For wrap-based partitioning, depending on the choice of partition sizes, guard wraps may be necessary between partitions. Partitioning using DB boundaries does not require guard wraps.

The host organizes data to be written to tape as host records of up to $16\mathrm{M}\mathrm{B}$ in size and host file marks. To map a host record to magnetic transitions on tape, the tape drive applies a number of data processing and formatting steps highlighted in Figure 13. The goal of an efficient tape format is to introduce as little redundancy and overhead as possible while ensuring highly reliable data storage and retrieval.

When receiving a write request for a host record, the drive first appends a 4-byte cyclic redundancy check (CRC) creating a protected record, followed by data compression using an algorithm known as Streaming Lossless Data Compression (SLDC) based on Standard ECMA-321 with extended history buffer. Optionally, the compressed protected record (CPR) is encrypted using AES encryption, leading to an encrypted compressed record (ECR). Next, the stream of formatted records, file marks, and control symbols are broken into datasets (DS) of 9,804,912 user bytes to which a 912-byte dataset information table (DSIT) is appended. A dataset represents the smallest complete unit of information that is written to, or received from, the tape.

Each DS is further broken into 64 sub datasets (SDS). The construction of a SDS including error correction code (ECC) encoding is detailed in Figure 14. The process starts with four matrices of size $K_2\times K_1$ bytes, where $K_2=168$ and $K_1=228$ bytes, from the stream of formatted records, labeled as “user data” in Figure 14. Each matrix is subsequently ECC encoded into a product code word (PCW) by means of two orthogonal encoding steps. First, a systematic column encoding using a RS$(N_2=192,K_2=168)$ Reed-Solomon code appends $N_2-K_2=24$ rows of so called C2 parity bytes. Then a systematic row encoding with a RS$(N_1=240,K_1=224)$ code adds $N_1-K_1=12$ columns of C1 parity bytes. The resulting ECC PCWs have a total size of $N_2\times N_1=192\times 240$ data and parity bytes.

The encoded SDS is constructed by means of 4-way column interleaving of the four PCWs as shown in Figure 14. In preparation for the tape layout mapping, a 12-byte header field is prepended to each row of the SDS, header protection is applied without overhead, turning each row of the SDS into a 972-byte headerized code word interleave (CWI-4), as highlighted in Figure 14. There’s a total of 12,288 CWI-4s per dataset.

Referring back to Figure 13, ECC encoding and CWI-4 header insertion is complete. The tape layout mapping step defines how the CWI-4s are ordered, assigned to the 32 parallel tracks, and eventually mapped to tape. The chosen mapping provides deep interleaving [36] to mitigate spatial bursts errors on the surface of the magnetic tape. Deep interleaving refers to spreading the information both laterally across all parallel recording tracks as well as longitudinally far along the tracks. A first property of deep interleaving aims at rendering the error symbols at the input of the C1 and C2 component decoders to be uncorrelated/independent of each other. A second property is that data can still be decoded correctly when a lateral stripe error affects many millimeters of tape surface in down track direction, for example due to instantaneous tape speed variations. Last but not least, the interleaving ensures that data can be decoded correctly even if up to four of the 32 tape tracks have simultaneous read errors, for example, due to disfunctional readers. As a result of the deep interleaving, the CWI-4s from the same SDS are separated by about $1\mathrm{m}\mathrm{m}$ on tape. Similarly, the bytes of each C2 column code word are uniformly spread across the full dataset.

Referring to Figure 13 again, now that the CWI-4s have been assigned to tracks, each track’s data stream passes through a data randomizer, followed by a rate-$32/33$ modulation encoder. The modulation code, often referred to as a runlength-limited (RLL) code, satisfies various run-length constraints on the maximum length of zeros [114], transition-runs [150], but also synchronization patterns used for timing acquisition [34] and twins pattern to limit the path memory length of the sequence detector [35]. The modulation encoded bit streams are extended with synchronization information, and passed to the write driver ASIC, which generates the write currents for the individual inductive tape write heads. The write clock frequency is selected in a tape-speed dependent fashion to achieve a written bit length $B=46.6\mathrm{n}\mathrm{m}$, as specified by LTO-9.

Some of the data encoding and formatting steps discussed above lead to overhead. The largest overhead is due to ECC, i.e., the linear product code with rate $(228/240)\times (168/192)=0.83$. The modulation code, a nonlinear constrained code, adds a rate $32/33=0.97$. Combined with header insertion and sync patterns, the total overhead is about $21\mathrm{\%}$.

Modern tape drives such as LTO-9 use another mechanism to further improve reliability and data integrity: read-while-write verification and rewrites. This technique uses a set of readers placed downstream of the write transducers to read back data immediately after it has been written to tape. The readback signal from each reader is processed by the read channel, modulation decoder and C1 decoder. If the number of C1 byte errors in a given CWI-4 exceed a threshold level, then the CWI-4 data is rewritten to a new location further down tape. The actual rewrites happen at the end of each dataset, i.e., after the 12,288 first-written CWI-4s have completed.

The number of rewrites per dataset is dynamic, but about $3\mathrm{\%}$ of the total cartridge capacity is reserved for rewrites. The average rewrite overhead for a new cartridge is less than $1\mathrm{\%}$. If more than $3\mathrm{\%}$ rewrites are necessary, the cartridge may not be able to achieve its advertised capacity.

It may have become clear that the combination of recording techniques discussed in this section mean that modern tape systems are designed using a write friendly architecture. For example, the tape drive applies a mechanism called speed matching to match the host data rate with the native write speed, so that frequent write stops and back-hitch operations to restart writing can be avoided. In the next section, we discuss data retrieval.

Figure 15 shows the data flow from the 32 MR readers scanning over the magnetic transitions of a data wrap on tape to the retrieved host data record provided via the host interface. Starting on the left side of Figure 15, a set of 32 shielded TMR readers are connected to a pair of analog/mixed-signal ASICs commonly referred to as the analog front- and backend, which implement the electrical biasing of the TMR readers, and act as a parallel bank of readback signal pre-amplifiers, variable gain amplifiers, filters and analog-to-digital (A/D) converters. To support a range of tape speeds for write and read operations, the analog filters have programmable cut-in and cut-off frequency settings to match the signal bandwidth and further provide variable high-frequency boosting of the readback signals. Furthermore, the programmable band pass filters avoid signal anti-aliasing during subsequent asynchronous A/D conversions at a rate of approximately $4/(5T)$, where T is the minimal bit temporal duration. This 25% oversampling is typical for tape drives and is a safety margin to cope with variations in tape velocity and linear density. The temporal bit length T and the spatial bit length B are related as $T=B/v_s$, where $v_s$ represents the tape speed.

Shielded MR readers are designed such that the readback signals respond primarily to transitions of the magnetization patterns on tape. At low bit densities, the transition responses had been clearly separated such that reading the recorded data could be accomplished by detecting the peaks of the transition responses by means of a peak detection channel. Today, because of the short bit length $B=46.6\mathrm{n}\mathrm{m}$ in LTO-9, corresponding to $545\mathrm{k}\mathrm{b}\mathrm{p}\mathrm{i}$ (kilo bits per inch) linear density, and the finite head resolution, the superposition of isolated transition responses partly cancel each other, causing a known phenomenon called intersymbol interference (ISI). To efficiently combat ISI, magnetic recording channels in tape and HDD products introduced partial response (PR) signaling with maximum likelihood (ML) sequence detection, or PRML detection [34, 64, 132]. A PRML read channel targets a controlled small amount of ISI at the output of the channel equalizer (EQ), which is left to the sequence detector to combat. This concept avoids full channel equalization and prohibitively strong noise enhancement at high bit densities. Modern tape read channels such as those in LTO-9 implement advanced versions of PRML and noise-predictive maximum-likelihood (NPML) [5, 37, 58] detection schemes, aiming at minimizing the influence of correlated and data-dependent noise in the detection process.

Read channels for tape drives require a high level of adaptivity to deal with media variations, such as fluctuations in particle dispersion and magnetic layer thickness, as well as variations in head-media spacing, wear of the head and media, head-tape interface related friction and tape velocity fluctuations. To cope with these temporal and spatial variations, the majority of signal processing and detection algorithms in tape read channels operate in a highly dynamic and adaptive manner. Examples include adaptive channel equalization, dynamic multi-channel timing recovery/synchronization, fast gain control, and noise-adaptive NPML detection [58].

Furthermore, the nature of parallel track recording in tape drives, coupled with the need for cartridge interchangeability, demands additional SNR margins and adaptability for robust operation over the lifespan of tape cartridges and drives.

At the output of the read channels, the bit streams of detected bits from the 32 parallel tracks excluding the synchronization patterns are passed to the modulation decoders and derandomizers. At that point, the raw headerized CWI-4s read from tape can be further processed, as indicated in Figure 15, by decoding the CWI-4 header, which provides the CWI-4 identifier (CWID) information necessary to subsequently de-interleave the CWI-4 payloads in memory in order to reconstruct the SDS and ECC PCW matrices discussed in the previous section. The next step, ECC decoding, will be thoroughly examined in the following subsection. However, if error detection and correction through ECC decoding successfully completes for all the data within a dataset, then the dataset is considered fully recovered and reconstructed. In case of an ECC decoding failure, i.e., uncorrectable code words, the drive switches to an error recovery procedure (ERP) mode, where additional or alternate ECC decoding steps and/or read retries are attempted.

By combining the user data from subsequent datasets into a stream of symbols, and parsing the symbol stream consisting of user bytes, file marks, and control symbols, the original host records can be derived by means of decryption (if enabled) and decompression. Data integrity is checked by means of the records CRC bytes and the requested host records are transferred back to the host.

The raw information bits and bytes read from tape, i.e., the data at the output of the read channels or modulation decoders, suffer from both random and permanent errors, typically caused by noise in the readback process and persistent media defects, respectively. Error correction coding (ECC), and specifically ECC decoding, aims to turn these unreliable raw data bytes at the ECC decoder input back into highly reliable user bytes at the output.

From LTO-1 to LTO-9, the storage capacity of LTO cartridges has increased by 180 times from 100 GB to 18 TB. This gain was achieved primarily by increasing the areal density $\sim$70x, but also by increasing the tape length by $\sim$1.7x and the format efficiency by $\sim$1.5x. Increasing the areal density in magnetic recording leads to a loss in SNR, which results in an increase in error rate and hence a reduction in reliability. To avoid such a decrease in reliability, the SNR loss due to areal density scaling needs to be compensated for with advancements in recording technologies such as e.g., improved media, read and write heads, data detection and ECC.

The detailed structure and parameters of the LTO-9 [39] ECC was already introduced in the previous section. As a reminder, Figure 16 highlights the two-dimensional LTO-9 product code with RS$(240,224)$ C1 row code and RS$(192,168)$ C2 column code again, and details the ECC decoding architecture and performance. Although the LTO ECC format has always been based on a two-dimensional interleaved dual ECC, it has evolved significantly in terms of performance. For example, LTO-9 has a new C2 format of RS$(192,168)$ compared to the LTO-8 C2 format of RS$(96,84)$, delivering superior performance at the same code rate and format overhead.

ECC decoding in the first eight generations of LTO has been based on a single pass through the C1 decoder, followed by a single pass through the C2 decoders. The C1 decoder operates in “error decoding” mode, where it detects and corrects byte errors in unknown locations. If the number of byte errors within a C1 code word exceeds the code’s error-correction capability, a decoding failure happens. The C2 decoder typically operates in “error and erasure decoding” mode: Upon a C1 decoding failure, the C1 decoder provides that failure information to the C2 decoding engine by marking the corresponding byte as erasures, which can be considered as byte errors with known location. The two-dimensional deep interleaving discussed in Section 5 spreads the data physically on tape, and therefore randomizes the effects of media defects and burst errors. This randomization results in largely uncorrelated byte errors at the inputs of both the C1 and C2 decoders and enables ECC to work effectively and efficiently.

LTO-9 tape drives, for the first time in the LTO history, introduced iterative ECC decoding in streaming mode. Although the first pass through the C1 and C2 decoder, which still employ hard-decision bounded-distance decoding, is identical to the previous generation’s technique discussed above, one or more additional iterations of C1 and/or C2 decoding steps are employed subsequently to further enhance the performance. Figure 16(b) shows the architecture of an ECC decoder with support for iterative decoding.

Figure 16(c) shows the ECC decoding performance for the LTO-9 ECC scheme, evaluated by means of hardware simulations (curves with symbols) and a probabilistic analysis (dash-dotted lines) [69]. Both the simulations and the probabilistic analysis assume uncorrelated byte errors at the input of the ECC decoder. We plot the post-decoding byte-error rate (BER) performance at the ECC decoder output, after a selected number of m decoding steps, as a function of the raw BER at the input. The “C1-C2” curve illustrates the outcome of a single pass through the C1 and C2 decoders, operating in classic mode. Remarkably, even with just one C1-C2 decoding sequence, the ECC decoder can transform a $1\mathrm{\%}$ raw input BER into an exceptionally reliable output BER that is better than ${10}^{-20}$, demonstrating the sheer power of the LTO-9 ECC. By adding an additional C1 decoding step, the curve labeled “C1-C2-C1” shows that with $m=3$ decoding steps, a $\sim 3\mathrm{\%}$ raw input BER is tolerable for a ${10}^{-20}$ output BER. With $m=4$ decoding steps, “C1-C2-C1-C2” decoding can tolerate up to $4.5\mathrm{\%}$ raw input BER. Figure 16(c) shows results for up to $m=6$ decoding steps, indicating that with increasing m, the rate of error correction improvements diminishes, and the probabilistic analysis slightly underestimates the post-decoding BER due to simplifying assumptions [69]. It should be noted that although the plots assume uncorrelated byte errors at the input of the ECC decoder, the LTO-9 ECC scheme is robust against burst errors too. A temporary dysfunctional data reader or large debris on tape can lead to long bursts errors in the C1 code words, but can be corrected by the C2 decoder at the expense of a small increase in post-decoding BER, as discussed in Arslan et al. [7].

LTO tape technology is known to have outstanding data reliability and durability, ensuring stored data remains intact and uncorrupted over time, thus guaranteeing long-term accessibility. This reliability is quantified by the End-Of-Life (EOL) uncorrectable bit error rate (UBER) of LTO cartridges [7, 115], which represents the likelihood of encountering an uncorrectable data error during readback. LTO has an innate resilience to write mode latent errors, thanks to its read-while-write architecture, where such write-related errors are promptly detected and rewritten during the writing process. The occurrence of uncorrectable error events in LTO tape is exceptionally rare, which makes it very challenging to measure the UBER experimentally. Instead, the LTO’s UBER specification is based on the theoretical performance of the ECCs implemented within the tape drive, combined with the assumption that errors are random and uncorrelated. The latest generation, LTO-9, further increased the reliability with an UBER specification of ${10}^{-20}$, which corresponds to the post-decoding BER threshold of ${10}^{-20}$ that we used in the LTO-9 ECC performance evaluation above. An UBER of ${10}^{-20}$ corresponds to one unrecoverable read error event for every 12.5 Exabytes of data read. Alternatively stated, on average, only one out of 694,444 LTO-9 tape cartridges will contain an uncorrectable error event due to ECC failure. We also note that the LTO-9 ECC scheme reserves two bytes of parity to detect and correct any one byte error that can occur with very low probability due to a C1 miscorrection. As a result, the probability of an undetected error due to miscorrection is extremely low as discussed in Arslan et al. [7].

TBS is a powerful servo technology that was developed in the late 1990’s specifically for linear tape [10]. All nine generations of LTO to date have used repeating TBS patterns with a basic geometry that is shown in Figure 18(a). A servo frame consists of four bursts of servo stripes designated as A, B, C, and D bursts. Each burst comprises four or five servo stripes having an azimuth angle $\pm \alpha$, which allows unambiguous servo frame detection and synchronization. The servo frame geometry is further defined by the servo sub-frame length $d_{\mathrm{s}}$ (or servo frame length 2$d_{\mathrm{s}}$), the stripe pitch p and the servo pattern height h. During tape drive operation, a servo reader moves over the repeating TBS patterns on tape and produces a readback signal. The readback signal of an individual servo stripe is called a dibit because of its shape [70]. A measured servo readback signal of a TBS frame is shown in Figure 18(b). Note that the timing of the dibits depends on the servo head’s trajectory over the servo pattern, which is indicated in Figure 18(a).

The servo readback signal is processed by a servo channel [29, 135] that measures the exact arrival times of the dibit pulses to calculate the time intervals between bursts of stripes with identical and opposite azimuth angle, called B-counts $B_i$ and A-counts $A_i$, respectively. The servo channel subsequently calculates the estimated lateral position $\hat{y}$ of the servo reader (trajectory) relative to the SB centerline based on the A and B-counts and the TBS pattern geometry according to

Similarly, the the tape velocity $\hat{v}$ is estimated using the B-counts aswhere $f_{\mathrm{s}}$ represents the sampling rate of the servo signal. The servo channel furthermore extracts additional information, such as the LPOS information and an SB identifier, which are embedded in the TBS pattern by means of pulse position modulation on a subset of servo stripes in the A and B-bursts.

As illustrated in Figure 12 (Section 5), LTO media contains five SBs that straddle four DBs spanning the width of tape. At media manufacturing time, repeating TBS patterns are written to all SBs, from the beginning to the end of tape. To write/read data to/from tape, as discussed in Section 5, the tape drive uses a stepper motor to laterally coarse position its head over the desired DB, such that the two pairs of servo readers on both the active writer and reader module of a terzetto head are located on the two adjacent TBS patterns straddling the DB, as depicted in Figure 19. To support lateral position estimates $\hat{y}$ from all four servo readers, indicated as ${\hat{y}}_{\mathrm{WT}}$, ${\hat{y}}_{\mathrm{WB}}$ for the writer module and ${\hat{y}}_{\mathrm{RT}}$, ${\hat{y}}_{\mathrm{RB}}$ for the reader module, the main ASIC in recent generations of LTO drives implements a servo hardware core containing four servo channels running simultaneously.

Furthermore, recent servo cores implement top-bottom skew estimators that process the exact arrival times of the dibit pulses from corresponding stripe bursts observed by the servo readers passing over the top and bottom servo patterns. Specifically, the top-bottom skew estimator calculates the distance $\mathrm{\Delta }x={\hat{x}}_{\mathrm{T}}-{\hat{x}}_{\mathrm{B}}$ between the LPOSs ${\hat{x}}_{\mathrm{T}}$ and ${\hat{x}}_{\mathrm{B}}$ of a module’s top and bottom servo reader, respectively, measured in tape travel direction x. The corresponding head-to-tape skew estimate, measured as an angle $\hat{\beta }$, can now be expressed aswhere $d_{\mathrm{TB}}$ represents the physical distance between the top and bottom servo readers. This skew estimate can be computed independently for both the active writer and the reader module.

For a tape drive to write (or read) the data tracks of a specific wrap, the tape head’s lateral position and skew do not only need to be accurately measured, but they also need to be actively controlled. We start by considering a write operation, as shown in Figure 19. Note that for simplicity, only 3 of the 32 data tracks of an LTO-9 drive are shown. The head’s lateral position is estimated by the servo channels as ${\hat{y}}_{\mathrm{WT}}$, ${\hat{y}}_{\mathrm{WB}}$ for the writer module and ${\hat{y}}_{\mathrm{RT}}$, ${\hat{y}}_{\mathrm{RB}}$ for the reader module, respectively. The servo reader trajectories are located above the servo pattern centerlines for this example of a write operation in backward direction. For optimal write track placement, the head’s lateral position $y_{\mathrm{avg}}$ is determined as the average of the top and bottom y-position estimates from the writer module asAveraging the top and bottom y-position estimates effectively determines the lateral position in the center of the writer (or reader) module, i.e., in the middle of the write (or read) transducer array. Furthermore, this dual channel averaging improves the position accuracy/resolution by reducing the noise in the position estimate by a factor of $\sqrt{2}$ in standard deviation compared to a single channel estimate [73].

Read-while-write verification is a unique functionality in tape recording, where a set of data readers placed downstream of the data writers immediately read and verify the data after it has been written to tape. A large head-to-tape skew $\beta$, as shown in Figure 19, causes the data readers to get “out of the shadow”, of the writers, and therefore off track, and consequently would break the read-while-write verification. To avoid such a situation, the measured head-to-tape skew $\hat{\beta }$ also needs to be controlled.

LTO-9 drives use two separate controllers for track following and skew following. Figure 20 shows a block diagram of the track-following control system. The control reference $y_{\mathrm{ref}}$ is the desired location of the data tracks (wrap) to be written or read. A position-error signal (PES) is generated by subtracting the estimated head position ${\hat{y}}_{\mathrm{avg}}$ from the control reference $y_{\mathrm{ref}}$ and fed to the controller. The servo controller, in combination with a current driver and a track-following actuator, adjusts the position of the head and thereby closes the track-following servo control loop. Note that instead of the exact head position $y_{\mathrm{avg}}$, only a delayed and noisy estimate ${\hat{y}}_{\mathrm{avg}}$ is available for control [57]. The controller is an H$\mathrm{\infty }$ compensator which takes into account the sensor delay D and the transfer function $G_{\mathrm{TF}}$ of the track-following actuator. The weighting functions of the controller are shaped to reduce disturbances from the tape path such as roller and reel vibrations [162, 163].

Figure 21 shows a block diagram of the skew-following control system which works in a similar fashion as the track-following control [160], but aims at minimizing the skew-error signal (SES). The controller aligns the readers and writers perpendicular to the tape by using a skew control reference ${\beta }_{\mathrm{ref}}=0$.

Tape path disturbances such as reel and roller vibrations, tape pack shifts, and tape path eigenmodes [24, 124] lead to lateral tape motion and tape tilt during tape transport. The skew- and track-following control loops discussed above aim at rejecting these disturbances by minimizing both the skew and lateral position errors. Head actuation for track- and skew-following is implemented using three actuators. The tape head is mounted in a spring guided voice coil actuator responsible for track-following (fine positioning) shown in Figure 10. For coarse positioning of the head to a desired DB and wrap location, a stepper motor controls the movement of the track-following actuator as it traverses along a guide rod. The assembly of tape head, track-following and coarse actuator is attached to a second voice coil actuator that provides rotational capability for skew following. Closing the position and skew feedback loops introduces secondary disturbances from the voice coil actuators or the servo sensor. Secondary disturbances are actuator eigenmodes, sensor noise, sensor delay, and longitudinal compression waves which can propagate through the control loop and must be properly addressed in the design of actuator, servo sensor and controller [57, 74].

To access data in a tape cartridge, the cartridge needs to be mounted to a drive, and subsequently loaded by the drive. The load operation includes the process of threading the tape from the cartridge reel over guide rollers and the tape head onto the machine reel located in the drive. At that point, the tape transport system is ready to move tape to carry out seek, read or write operations. Figure 17 depicts the tape’s path from one reel to the other via four grooved flangeless guide rollers which ensure a smooth and stable head-tape interface at the tape head. Both reels are driven by brushless DC motors to wind and unwind tape. The reel motors are regulated by the tape transport control system, which aims at moving the tape at a commanded tape speed and tape tension across the read/write head.

As discussed in Section 3, the emergence of tape featuring an ultra-thin polymer substrate, coupled with parallel-track recording at narrow track widths, has posed a growing challenge from the phenomenon known as TDS. TDS refers to changes in tape width in response to changes in temperature, humidity and tension. If data is written under one set of environmental conditions and then read back again later under different conditions, the pitch of the transducers in the read head may not match the pitch of the tracks on tape. To address these TDS effects, LTO-9 drives implement active TDS compensation leveraging tension control. By adjusting the tape tension around a nominal tension value, the width of tape can be changed, as shown in Figure 22. By increasing or decreasing the tape tension, the relative distances between data tracks on tape get smaller or larger, respectively. Figure 22(b) shows a hypothetical example of a write operation, where a (too) large tape tension during write leads to a misalignment between desired and written data tracks. An estimate of the relative width mismatch $\hat{w}$ between the head span (servo reader pitch) versus the tape span (SB pitch) can be derived from the servo position estimates of a writer or reader module asassuming that the head-to-tape skew $\beta$ is negligible small.

Figure 23 shows a simplified block diagram of the multiple-input multiple-output tape transport control system [15, 31, 32, 161]. The model based controller is fed with the width mismatch error signal $w_{\mathrm{err}}=w_{\mathrm{ref}}-\hat{w}$ and tape velocity error $v_{\mathrm{err}}=v_{\mathrm{ref}}-\hat{v}$ and calculates and applies the motor currents. The controller model takes into account the reel inertias, viscous friction values, and the amount of tape wound up on each reel [30]. Disturbances influencing the width error are tension changes due to roller and reel imbalances, temperature, humidity and tape creep. On the other hand, the tape velocity error is affected by unmodeled friction errors and unmodeled tape transport dynamic effects. The controller performance is also limited by sensor noise and delay.

At the time of writing, two families of tape drive were under active development with roadmaps for future generations: IBM TS11xx Enterprise and LTO. The latest generations of both families: the IBM TS1170 and the LTO Gen 9 full high (FH) and half high (HH) drives, are shown in Figure 24. Table 2 presents a summary of their performance characteristics. All three drives have recommended temperature and humidity ranges of 16–25C and 20%–80% relative humidity respectively. LTO-9 drives provide backwards read and write functionality to LTO-8 media. Compared to the FH drive, the HH drive has lower performance in terms of data rate and locate times and is not rated for as many load/unload cycles. Early generations of both families of tape drive are also commercially available and may provide a lower $/TB media price point with the tradeoff of lower data rates and volumetric densities.

State-of-the-art tape drives have a variety of characteristics that are distinct relative to HDD or flash. Knowledge of these characteristics may be useful to users of tape or anyone writing an application to use tape and are discussed in the following sub-sections.

Tape is an append only technology, as discussed in Section 5. As a result, data cannot be deleted or overwritten in place. It is therefore necessary for the tape management software to keep track of what data has been written on which cartridge and to also keep track of what data has been deleted. If a significant fraction of the data stored on a cartridge has been deleted, the remaining data can be migrated to a new cartridge and the initial cartridge can be reused.

LTO tape drives maintain a tape directory (TD) in the CM of each cartridge that contains, among other information, the: (1) record count, (2) file-mark count, (3) write pass number and (4) dataset number at the middle and end of each wrap. When a request to retrieve data from the tape is received, the drive uses the TD to estimate/interpolate the LPOS on tape of the requested data. The drive then high speed locates to an LPOS position several meters before the estimated position, slows down to read velocity and begins to read the DSIT information to determine if it is near the target dataset. If so, the drive continues the read operation until the data has been returned to the host. If the drive significantly over or underestimated the location, the DSIT data is used to re-estimate the target position and a back-hitch or new seek operation will be performed. If the data stored on tape was in-compressible or had a relatively constant compression rate, the initial estimate will be quite accurate and the first seek will likely be successful. However, if there is a lot of variation in the compression rate, the estimate may be inaccurate and the seek operation will take more time. IBM TS11xx drives maintain a high-resolution tape directory (HRTD) with 64 equally spaced entries per wrap. This enables a much more accurate position estimate and hence also enables a better random seek performance. The length of tape used to store user data in an LTO-9 cartridge is about 1,000 m. The average distance travelled for a random seek after load is 1/2 the length of tape and each subsequent random seek is 1/3 the length of tape on average, or about 55 s and 35 s, respectively.

Because of the serpentine recording technique used in linear tape drives and built-in compression, there is no clear relation between block number and physical location on tape. Two blocks of data separated by a large interval in block number may be physically very close but on adjacent wraps or in an adjacent DB. Recent generations of TS11xx drives implement a function called RAO which enables faster data retrieval if multiple blocks/files are to be recalled from a cartridge. An application can use the function to request the drive to generate a recommended access order in which blocks should be requested to minimize the seek time. The drive uses the HRTD to estimate the physical start location for each file and then calculates a recall order that minimizes the total seek time as illustrated in Figure 25. If the number of files is large this calculation is very computationally intensive and the drive calculates an approximate solution. The performance gain from RAO depends on the number and size of the files but is typically on the order of a 30–60% reduction in access time [207]. The latest generation of LTO drive, Gen 9, provides a similar functionality called open RAO, however the performance gains are lower because of the lower resolution of the TD in LTO.

The phenomena of TDS and storage creep were discussed in Section 3 of the article. Both the TS1170 and LTO-9 drives use tension control to actively compensate for TDS, as discussed in Section 7. This can result in a significant variation in the tension wound into a tape reel which would cause pack distortions over time due to storage creep, if left unaddressed. To minimize storage creep, TS1170 and LTO-9 drives perform a low-tension unload operation in which the tape is wound into the cartridge at a low tension immediately before it is unloaded from the drive. To do this efficiently, the drive keeps track of the furthest location down tape accessed during a given mount, and the drive high speed locates to just beyond this point, lowers tension and rewinds the tape into the cartridge reel. In the TS1170 drive this operation is performed at 18 m/s and is significantly faster than in LTO-9 drives that perform the operation at the high speed search speeds listed in Table 2.

In the early days of magnetic tape storage, tape reels were stored on shelves in large rooms adjacent to the data center and had to be manually retrieved and loaded onto a tape drive by human operators. Such rooms were referred to as tape libraries and typically housed hundreds to tens of thousands of tape reels. The introduction of tape cartridges and automated robotic tape libraries (see Section 13) reduced the access time to data stored on tape from hours to minutes and significantly improved reliability by removing error prone humans from the loop. Today, an automated tape library typically consists of an enclosure that contains tape cartridges, one or more tape drives for performing read/write operations, an optical mechanism such as a bar code reader to identify cartridges and a robotic and control mechanism used to transport cartridges between storage locations and the tape drives and to mount and unmount cartridges to and from the drives. After the robot mounts a cartridge to a drive, the drive loads the cartridge and performs a seek operation to the desired location on tape to perform a read or write operation. After read/write operations are complete, the drive must rewind and unload the cartridge before the robot can unmount it from the drive. Storage locations are typically referred to as “slots” and the robotic mechanism as an “accessor” or “robot arm”. A library also typically includes a mechanism(s) for loading and unloading cartridges to the library referred to as an I/O slot.

Automated tape libraries are available in sizes that range from a few slots and a single drive in a 1U “pizza box”, to large modular systems that are expandable to tens of thousands of slots and that support more than one hundred tape drives. Larger systems may also support more than one accessor. We take IBM’s current tape library portfolio, shown in Figure 26, as an example to illustrate the range of available systems. The smallest system, the TS2900, has 9 available slots and one HH LTO tape drive in a 1U form factor. The TS4300 is a rack mounted modular system with a starting configuration of 40 slots and up to 3 HH LTO drives or one FH LTO drive and can be expanded to up to 640 slots and 16 FH or 48 HH LTO drives. The Diamondback library is a rack sized library designed specifically for the unique requirements of hyperscale cloud companies, such as ease and speed of deployment as well as optimal reliability in erasure coded environments and self service. The library can be shipped with pre-installed media and drives and can be installed in a datacenter in less than 30 minutes. The Diamondback library fits in the same floor space as a standard open compute project (OCP) rack and with LTO-9 technology provides up to 27 PB native capacity in a 0.7 m² footprint. At the upper end, the TS4500 is a modular system consisting of 1 to 18 storage frames in a linear architecture that can be configured with up to 23,170 slots and up to 128 FH LTO or TS11xx Enterprise drives. With TS1170 media, the maximum capacity is 877 PB for a density of 51.6 PB/m² of floor space. All four library types have an ethernet port for remote library management and have options for fiber channel or SAS for the data path, with the exception of the TS2900 which only supports SAS. A variety of other companies including Dell, Fujitsu, HPE, Quantum, Spectra Logic, and Tandberg also currently offer tape library solutions.

The functionality and features provided by tape libraries vary with the size/type of library and between vendors. To provide some insight into the type of features that are available, we use the IBM TS4500 library as an example. The TS4500 has an integrated management console and provides remote management functionality via a GUI or a command line interface (CLI) and in addition supports REST API and REST over SCSI commands. Remote monitoring is supported via SNMP, e-mail or syslog. The library has separate and redundant data and control paths with path failover for both and has redundant power supplies. In addition, it supports up to two active accessors and provides automated/library managed media health checking. Mixed media and drive types are supported in the same library and a single physical library can be partitioned into multiple logical libraries each with distinct cartridge and tape drive resources. Partitioning can be used for file management purposes or to give different users or applications access to dedicated, independent resources [8]. Recent generations of tape drives provide built-in hardware-based encryption. There are typically three methods for managing encryption: system managed, library managed and application managed encryption. In all cases a key is required to encrypt and decrypt the data. Details of how this is implemented and managed vary depending on the application and/or library and vendor.

A key performance metric for tape libraries is the time to mount a cartridge, i.e., to fetch it from a slot and insert it into a drive. Performance is often specified as the number of mount cycles that can be performed per hour, where a mount cycle is defined as the steps of removing (unmounting) a cartridge from a drive and replacing it to a storage slot and then fetching a new cartridge and mounting it to the drive. Mount rates can vary considerably and typically depend on the number and speed of the accessors (robot arms) and the architecture of the library. For example, in expandable modular libraries such as the TS4500, the average time for a mount cycle increases with the number of frames in the configuration because on average the accessor needs to travel longer distances between the drives and slots. The design of the storage slots also plays a role. For example, in the TS4500 and Diamondback libraries cartridges are stored in deep slots where multiple cartridges are stored one behind the other in “tiers”. The accessors are equipped with dual grippers that can manipulate two cartridges such that cartridges in the front two tiers can be accessed quickly by the dual gripper, whereas access to cartridges in deeper tiers requires the cartridges in front to be first moved to other slots, resulting in longer access times. To improve the mount performance of such systems, algorithms have been developed to ensure that frequently accessed cartridges are stored in the front tiers and infrequently accessed cartridges in the back tiers. In Spectra Logic libraries, cartridges are stored in drawers which must first be opened by the accessor to retrieve a cartridge which adds an extra step to the mount process, but can be efficient if multiple cartridges are accessed sequentially from the same drawer.

To properly dimension tape library systems, it is imperative to develop models that evaluate the effect of the various parameters on their performance. The results obtained can subsequently be used to provision tape systems to achieve service level agreements for given workloads and data access times. Here we provide a survey of the relevant literature on performance evaluation of tape library systems.

Over the last six decades there have only been a few publications that evaluated the performance of tape library systems. This is attributed to the fact that a theoretical evaluation of their performance is an extremely challenging task and consequently a performance assessment was mainly conducted by simulations [55, 78, 138] and measurements [123]. However, simulations and measurements are time-consuming compared with analytical models that not only provide fast execution times, but also insight into the system dynamics and the effect of the various parameters.

The first analytical attempt resulted in an intractable queueing model, which led the authors to conduct their study by means of simulation [138]. Twenty years later, approximate analyses were presented using $M/M/c$ [77] and $M^X/G/c$ [122] queueing models. Another twenty years later, an improved model that captures the polling nature of operation of tape library systems was presented by Iliadis et al. [113]. This model obtained an approximate closed-form expression of the mean waiting time for independent and identically distributed (i.i.d) seek times and for an abundance of robot arms such that one is always available to mount/unmount a cartridge. System operation and performance was assessed in the light-load and the heavy-load regions using an $M/G/K$ and an $M/G/1$ polling queueing model, respectively, and subsequently in the medium-load region through a tangent interpolation.

An enhanced theoretical analytical model that considered the variability of seek times and captured the principal aspects of tape library operation including the contention for robot arms and its effect on system performance was presented by Iliadis et al. [111, 112]. Using elaborate, non-standard queueing theory models, it evaluated performance measures of interest, such as the mean waiting time and the robot mount rate, as a function of the system parameters, including the number of tape cartridges, the number of tape drives, the number of robot arms, and the tape load, seek, rewind, and unload times. It obtained results that enabled a better understanding of the design tradeoffs and provided useful insights into the behavior of the tape libraries. Also, a theoretical analysis presented by Iliadis et al. [112] yielded a condition for determining whether the robotic mechanism is a bottleneck.

We proceed to briefly review the model presented by Iliadis et al. [112] along with the main parameters and some of the performance results obtained. The tape library system contains c cartridges, d tape drives, and a robot arms with a robot transfer time R. Requests submitted to tapes are queued and subsequently served according to a first-come-first-served (FCFS) policy. The waiting time of a request is the time interval from its arrival to the tape system until the corresponding data transfer is initiated. To ensure fairness, tapes are mounted according to a cyclic (round-robin) policy. Requests to cartridges are served according to an exhaustive service discipline such that a tape is only unmounted when all its requests have been served. Subsequently, another tape with pending requests is mounted. If, however, there are no pending requests to any other non-mounted cartridge, two policies are considered depending on whether the tape remains mounted.

The workload was assumed to be symmetric with requests arriving to each of the c cartridges according to independent and identical Poisson processes at a rate of ${\lambda }_{\text{ct}}$ such that requests arrive to the tape system according to a Poisson process with a rate $\lambda$, where $\lambda =c{\lambda }_{\text{ct}}$. The request sizes were assumed to be i.i.d. The notation is summarized in Table 3. The parameters are divided according to whether they are independent or derived and are listed in the upper and the lower part of the table, respectively.

Performance was evaluated using a combination of elaborate, non-standard queueing theory models, whereby requests submitted to tapes correspond to jobs arriving to queues and tape drives correspond to servers. System operation and performance was assessed in the light-load and heavy-load regions using an $M/G/K$-based and an $M/G/1$ polling-based queueing model, respectively, based on an iterative procedure. Subsequently, the mean waiting time $\bar{W}$ in the medium-load region was derived through interpolation. Several relevant performance measures including the robot mount rate ${\lambda }_{\text{rb}}$ were also obtained as a function of the relative arrival rate $\theta$ defined as follows: $\theta \triangleq \lambda /{\lambda }_{\text{max}}$, such that $0\le \theta \le 1$, which reflects the system load.

Next, we review the theoretically obtained closed-form approximation for the robot mount rate ${\lambda }_{\text{rb}}$ as a function of the load $\theta$ [112, Equation (58)]. The maximum robotic mount rate $m_R$ for a number a of robot arms is determined by $m_R=a/(2R)$, where R denotes the robot transfer time.

In the light-load region, the robot mount rate ${\lambda }_{\text{rb}}$ increases linearly with $\theta$, as depicted by segment AB in Figure 27(a). By contrast, in the heavy-load region, ${\lambda }_{\text{rb}}$ decreases linearly, as depicted by segment BC in Figure 27(a). The maximum robot mount rate $m^{\ast }$ is achieved when $\theta ={\theta }^{\ast }$, that is, $m^{\ast }={\lambda }_{\text{rb}}({\theta }^{\ast })$, with ${\theta }^{\ast }$ and $m^{\ast }$ determined by Equations (56) and (57) of Iliadis et al. [112], respectively.

Note that Figure 27(a) is valid only when the robot mechanism is not a bottleneck, that is, when the robot mount rates do not exceed the maximum possible rate of $m_R$, or equivalently, when the number of robot arms a exceeds a value, which is determined by Equation (59) of Iliadis et al. [112]. When the number of robot arms a is less than this value, the robot mechanism is a bottleneck and the robot mount rates are truncated, as shown in Figure 27(b) by the segment DE. In this case, it holds that ${\lambda }_{\text{rb}}({\theta }_1)={\lambda }_{\text{rb}}({\theta }_2)=m_R$.

Support for tape usage at the operating system and application programming level is very dependent on the operating system being used. As potentially useful examples of tape system support we summarize tape interfaces from a selection of IBM mainframe (IBM Z), Unix, and PC-class operating systems.

Applications that support tape should consider the unique abilities and concerns of tape devices and media. Any concerns should not deter an application from using tape where appropriate, but rather should inform the implementer’s approach to using tape.

The previous sections discussed the benefits of tape storage in enterprise IT environments. Cloud storage has become one of the most important storage technologies in recent years [186] and the S3 API developed by Amazon has been widely adopted by other cloud providers [208]. To enable tape use amongst enterprises that are already familiar with the S3 protocol for cloud storage, a variety of products have been developed that provide an S3 interface to tape. Examples include: PoINT Archival Gateway [188], Grau Data XtreemStore [48], IBM S3 Deep Archive [106], Versity ScoutAM/Gateway [194, 195], FUJIFILM Object Archive [67], XenData LTO Active Archive [205], QStar Archive Storage Manager [168], and Spectra On-Prem Glacier [140].

In order to get a comprehensive understanding of how tape storage is used, it is helpful to look at the industries that use tape. According to the 2023 Tape Storage Global Market Report [173], key end users of tape storage include cloud providers, data centers and enterprises. The leading industries are IT, telecommunications, media and entertainment, healthcare, oil and gas, government and defense. A common use case that traverses many industries is data retention on tape for legal compliance reasons [88]. Another common factor motivating the use of tape highlighted in Wu [204] is tape’s low energy consumption relative to other storage technologies. Rising energy costs and the impact of global warming affect all enterprises and provide a motivation for reducing IT energy consumption. For enterprises that have a lot of cold data there is an opportunity for significant energy reduction simply by moving this data to tape.

Tape storage has undergone a remarkable evolution over its 70-year history, with data rate and capacity experiencing continuous growth. While data rate improvements have been achieved primarily through scaling of the linear density and increasing the number of parallel channels, capacity has increased mostly through scaling of the areal density, complemented by format efficiency improvements and thinner tapes allowing for greater lengths. Currently, a state-of-the-art enterprise class tape drive (IBM TS1170) operates at a native data rate of 400 MB/s, uncompressed capacity of 50TB written at an areal density of 26.1 Gb/in² [102]. Compared to the first IBM tape drive, this corresponds to an increase of more than five orders of magnitude in data rate and more than seven orders of magnitude in areal density. However, despite this impressive scaling, the areal density of the TS1170 is still about 48 times lower than the 1,260 Gb/in² areal density of a recent 22 TB HDD [178]. Considering that both tape and HDD rely on the same basic magnetic recording principles, this gap implies that from a basic recording physics perspective, tape has considerable potential to continue scaling its areal density. This gap also indicates that there is an opportunity for tape engineers to continue to leverage and adapt technologies developed for HDD to enable continued tape scaling.

The potential to continue scaling tape areal density has been explored in multiple tape areal density demonstrations [28, 69, 71, 72, 126, 137]. For example, a recent tape demo that used a prototype Strontium Ferrite (SrFe) particulate media showed the potential for tape recording at 317 Gb/in² [69], which could enable a cartridge capacity of 580 TB, assuming similar format overheads to current products. Another more recent demonstration on sputtered thin film media showed the potential for recording at 400 Gb/in² [126] which could enable even higher cartridge capacities.

The 2024 INSIC Tape Roadmap projects that tape areal density will scale at a 28% compound annual growth rate (CAGR) over the period 2024–2034 [117]. Combined with a 2.5% CAGR in tape length and small improvements in format efficiency, capacity is projected to scale at a 32% CAGR. Over the same period, data rates are projected to scale at a 15% CAGR. This scaling results in a projected native cartridge capacity of 723 TB recorded at an areal density of 315 Gb/in² and a tape drive data rate of around 1.6 GB/s by 2034. Note that the areal density projected at the end of the roadmap is lower than both the aforementioned tape demos and is well below that of current HDD products, indicating the potential to continue scaling beyond the end of the roadmap. The IEEE International Roadmap for Devices and Systems: 2023 Mass Data Storage—Tape Storage Roadmap projects similar scaling rates with predictions out to 2037 at which time native cartridge capacities are predicted to reach 1.5 PB, with an areal density of about 602 Gb/in² [33]. Both these roadmaps discuss many of the challenges and the technologies that will likely be required to enable this scaling.

The tape research demos mentioned above used a single recording channel to explore the potential for head and media technologies to support future tape areal density requirements. This aspect arises from the use of HDD heads with a single, very narrow reader to explore these future operating points. Several of these demos also reported other key technologies that will be needed in the future such as technologies to achieve tape track-following accuracies down to the nanometer scale and the potential to write ultra-narrow tracks on tape. However, the additional challenges such as TDS, that arise in tape recording due to the use of multiple parallel channels, were not addressed. Hence, to achieve such areal densities in commercial tape products, continued research and development will be needed in several areas. Foremost is the need for active TDS compensation schemes that can achieve performance levels comparable to those demonstrated for track-following. Additional future research areas include continued improvements in tape media, tape-head tribology, data channel, ECC and servo control, both in terms of refining the technologies developed for these demonstrations for use in commercial tape systems and developing new technologies to enable further scaling.

Another key research topic is adapting state-of-the-art HDD reader and writer technologies for use in tape heads. For example, HDD introduced monopole writers in combination with a media that incorporates a soft under layer (SUL) in the 2005 time frame [117] at areal densities around 150 Gb/in². At current rates of scaling, tape will likely reach this areal density range in the early 2030’s time frame. Monopole writers with a SUL enable much stronger write fields that are necessary for writing very high coercivity media. If this technology can be implemented in tape write heads it could enable the use of state-of-the-art HDD media technology. Moreover, if a SUL can be combined with a particulate mag-layer, monopole writers could enable the use of very high coercivity particle technologies such as spherical epsilon-Fe₂O₃ [156]. Another HDD head technology likely to be needed for tape is the combination of TMR readers with soft biasing and side shields that the HDD industry introduced at a reader width of about 40 nm [142]. The use of soft biasing helps to reduce sensor variability and enables the use of side shields that reduce the pickup of signal from adjacent tracks.

To scale tape beyond the limits of conventional perpendicular magnetic recording, tape will likely have to adapt one of the two so-called energy assisted magnetic recording technologies that the HDD industry has been developing. The first of these, called heat assisted magnetic recording (HAMR), uses a laser and near field transducer integrated on the HDD head to locally heat the recording layer to temporarily lower its coercivity [133]. Implementing this technology for tape would be very challenging. For example, each write transducer in the head would require a laser and near field transducer. Currently tape drives have 64 write transducers and are expected to increase this number to 256 within a decade. HAMR technology also requires the careful design of the heat sinking of the recording layer into the HDD platter to constrain the local heating to the track being written. It seems unlikely that similar heat sinking could be achieved on the 4–5 $\mu m$ thick polymer substrate used for tape. The second energy assisted recording technology under development is called microwave assisted magnetic recording (MAMR). In MAMR, a spin torque oscillator is integrated into the write head and used to assist in the recording of very high coercivity media [209]. The spin torque oscillator is fabricated using thin film microfabrication technology during the fabrication of the writer and would, therefore, be more practical to integrate into the manufacture of tape heads. In addition, MAMR does not require the heat sinking needed for HAMR again making it a better candidate for use with tape. Initial research into microwave assisted switching for patriclate based tape media has also been reported [158].

To enable significant future increases in tape data rate, it will be necessary to increase the number of parallel channels. This will most likely occur by first doubling to 64 channels and then later doubling again to 128 channels. This straightforward sounding change impacts many aspects of the drive architecture, including the tape layout, the head and flex cables, the drive card and all of the the custom ASICs needed to drive the parallel channels, i.e., write drivers, analog front and backends, and the main ASIC that hosts the data and servo channels and ECC. The continued scaling of ASIC technology has enabled tape engineers to develop new ASICs with additional parallel channels without significant increases in the ASIC die size. However, as discussed in Sections 5 and 6, operation at variable tape speeds, the removable nature of tape media and backwards compatibility all add significant complexity to the design of these ASICs. Another challenge arises from the rising costs of designing custom ASICs as the technology nodes scale to smaller transistor sizes and more expensive lithography and mask technology. Considering the relatively small volume of ASICs required to meet the demand for tape drives, the tape industry has to amortize the cost of designing and manufacturing custom ASICs by reusing them in multiple generations and across different product families. In spite of these challenges, tape has significant potential and a clear roadmap to continue scaling for multiple future generations. From a technology perspective, we believe tape has the potential to scale to cartridge capacities on the order of a petabyte and beyond. However, whether such capacities will be achieved in commercial products also depends on a variety of additional business factors such as the market demand for such products and the amount of funding available to develop them.

Recent studies have reported that data is growing exponentially and is expected to continue to do so in the coming years. Much of this data is transient and may not need to be retained, however worldwide storage capacity is nevertheless projected to also grow exponentially. For example, a recent study estimated the total installed base of storage capacity to be around 8 zettabytes in 2021 and projected it will grow to 16 zettabytes by 2025 [192].

Multiple factors are responsible for the growth in data and data retention needs, including the growth in smartphone users (currently estimated at 2.87 billion [187]), the growth in connected devices (projected to grow to 41.6 billion by 2025 [171]), regulatory changes in retention periods for data such as health records, and the growth in the size of training datasets for AI models and the need to preserve the raw data for provenance purposes. Emerging AI technologies also hold the promise of providing new means for extracting value from data, providing an incentive to preserve data for potential future use. A white paper from Horizon Storage Strategies reported that over 60% of all data is archival and projected that this could reach 80% or more by 2024 [187]. Considering all these factors, we conclude that the demand for archival storage capacity will grow exponentially in the coming years.

Magnetic tape storage provides a cost-effective way to retain the exponentially increasing volumes of data currently being created. Tape’s low cost per terabyte and enhanced cyber-resiliency combined with its low energy consumption and CO₂ footprint make it an appealing option for storing infrequently accessed data. These factors combined with tape’s potential to continue scaling capacity and data rate have resulted in a resurgence in use of the technology and will likely lead to an expanded role to meet the growing need for cost effective, green storage solutions.

The first practical demonstration of magnetic recording was performed by Valdemar Poulsen in 1898 who recorded analog audio on a steel piano wire [46]. Recording on something resembling modern tape was first demonstrated by Fritz Pfleumer in 1927 who recorded analog audio on thin strips of paper coated with iron oxide powder [47]. Another key step was the 1934 development by Eduard Schüller of the ring recording head which provided a more focused magnetic field and enabled more precise control of the magnetization of small regions of the magnetic media [47]. Over the next two decades, a foundation for the development of tape data storage were laid in the continued improvement of analog audio recording on magnetic tape, led primarily by AEG in Germany and Ampex in the U.S.A. [79]. The first magnetic tape system for data storage, the UNISERVO I, was commercialized by the Remington Rand corporation in 1951 and was the primary I/O device for the UNIVAC I computer. The UNISERVO I used a metal tape made from a very abrasive nickel-plated phosphor bronze alloy. It recorded data using a staggered head in an 8-track format using six tracks for data, one for parity and one for timing. A full reel of tape weighted about 25 pounds requires the use of large motors and a complex pulley and lever mechanism to buffer the tape and enable the rapid start and stop motion required to input small blocks of data to the computer for processing. The block size was limited by the small amount of memory available in early compute systems. Data was recorded at 128 bits/inch at a tape speed of 100 inches/sec. The large inertia of the reels necessitated the use of a relatively large 2.5 inch inter block gap between data blocks of 60 words of 12 characters, resulting in a transfer rate of about 7,200 characters per second. The 1,200 foot long tape of the UNISERVO I had a capacity of about 1.5 MB [201]. Note that early tape systems recorded data as characters rather than bytes, however, to simplify comparisons to later products the capacities of these systems have been converted to an effective capacity in bytes. In 1952, IBM announced the IBM 726 magnetic tape unit for use with the IBM 701 computer. The 726, shown in Figure 33, implemented an innovative vacuum column technology to buffer the tape during start and stop, which enabled the use of a much lighter but more fragile polymer-based tape and a smaller 0.75 inch inter block gap. The 726 had a 7-track format with an in-line head, using 6 tracks for data and one for parity. The 726 recorded at a linear density of 100 bits/inch in an NRZI encoding format, at tape speeds of 75–100 inches per second. The tape had a length of 1,200 feet and a capacity of about 2 MB. The use of vacuum columns, polymer tape and the 7-track format were widely adopted and became a de facto industry standard. Signal amplification and logic were implemented using vacuum tubes. An important focus in early tape development was on increasing data rates to match the increasing speed of compute systems. Initially this was achieved through a combination of linear density and tape speed scaling. For example, the IBM 727, released in 1955, operated at 15 kB/s and the 729-3, released in 1958 at 63 kB/s. Starting with the 729-2, the vacuum tubes used in earlier drives were replaced with transistors. Another important innovation of the 729 was the introduction of a dual gap head that enabled on the fly read while write verification of the data [83].

The first automated tape library, the IBM 7955 (Tractor), was developed in 1962 and delivered to the NSA as part of the Harvest computer system. The design was very innovative for its time, using 1.75-inch wide tape, housed in dual reel cartridges and the library was equipped with an automated mechanism to fetch, load and unload cartridges from the six tape drives that made up the system [165]. However, only one system was ever made, and it would be another 30 years before automated tape libraries began to see widespread adoption. The next decade saw a continued focus on tape I/O performance as well as innovation in tape handling and system integration. In 1964, IBM launched the IBM 2104 (Models 1-3) for use with the IBM System/360. The 2104 was the first 9 track tape format (8 for data and 1 for parity) and introduced CRC. It had a capacity of about 5 MB per reel and a data rate of up to 90 kB/s. The 2104 Models 4-6 followed in 1965 that had a capacity of about 10 MB and a data rate up to 180 kB/s. Two key innovations that enabled this increased performance were the implementation of electronic skew buffers and self-clocking tracks [83]. The IBM 2420, announced in 1968, operated at a data rate of 320 kB/s and introduced a self-threading mechanism for the tape that provided significant operator time savings. The use of DC motors enabled a tape speed of 200 inch/sec and a remarkable start/stop time of less than 2 ms [83]. Three models of the IBM 3420 were introduced in 1971 for use with the IBM System/370 computer and operated at a linear density of 1,600 bpi and data rates up to 300 kB/s. In 1973, three additional models were introduced operating at a linear density of 6,250 bpi and data rates up to 1,250 kB/s. The 3420 embedded a set of around one thousand hardware-assist instructions that eliminated the need for a separate switching unit to provide more than one CPU access to the tape drive and can be viewed as the beginning of tapemicrocode [110].

At the end of 1964 Digital Equipment Corporation (DEC) introduced the relatively low cost DECtape system (also called Microtape) with its PDP-7 minicomputer. DECtape was based on the LINCtape system designed at MIT-Lincoln Labs. Compared to the other tape systems discussed above (and those that follow), DECtape had many unique aspects. For example, DECtape was a random access block addressable storage device that essentially behaved like a hard disk with very high latency and could be used as the main storage for the operating system. It used 250 feet of 3/4${}^{\prime \prime }$ tape wound on an $\sim 4^{\prime \prime }$ reel with 10 tracks, (6 data, 2 mark tracks and 2 clock tracks). Mark and clock tracks were pre-formatted during tape manufacturing. To improve reliability, each track was paired with a non-adjacent track that contained the same data. To make the tape durable enough to support random IO, the surface of the magnetic recording layer was coated with mylar to protect the recording layer from wear. In 1978, DEC introduced DECtape II which was a similar block addressable random-access device but used 0.15${}^{\prime \prime }$ tape housed in a miniature cartridge [16, 40, 41, 42].

In 1974, IBM released the first commercially available automated tape library, the IBM 3850 Mass Storage System (MSS). The 3850, shown in Figure 34, was very innovative for its time, utilizing a variant of helical scan [46, 157] recording technology and cylindrical plastic cartridges with a 1.86-inch diameter and 3.49-inch length. The cartridges held 770 inches (20 m) of tape with a 50 MB capacity and were housed in honeycomb cells along 2 walls of the library that were accessed by two robots (accessors). Data was staged in and out of the system via DASD (HDD) and the host and application treated all data as if it were stored on DASD making the 3850 the first instance of a virtualized storage system. Despite the innovation, the concept was ahead of its time and the 3850 had limited commercial success with no follow-on products [46, 84, 121].

In 1984, IBM released the 3480 magnetic tape subsystem that became a de-facto industry standard; companies including Fujitsu, M4 Data, Overland Data, StorageTek, and Victor Data made tape drives compatible with the 3480 standard. The 3480, shown in Figure 35, was a significant departure from previous open reel tape drive designs and introduced multiple new innovations including a chromium dioxide particle based media housed in a 4” x 5” x 1” cartridge that replaced the previously used 10.5” reels and set the stage for automated tape libraries. In addition, the 3480 introduced an 18-track MR head manufactured using thin film technology. This was the first use of thin film technology in magnetic recording and was a key technology to enable future areal density scaling via the continuous miniaturization of the write/read transducers through improvements in lithography technology. In addition, the 3480 eliminated the vacuum columns used in previous open reel drives enabling a more compact drive that required half the floor space of the 3420. The 3480 also achieved significant improvements in error detection and correction using adaptive cross parity (AXP) coding. The 3480 had a capacity of up to 400 MB and operated at a data rate of 3 MB/s. In 1986, IBM released the IBM3480 IDRC (improved data recording capability) which added hardware-based data compression and enabled a 2x increase in capacity and data rate [46, 95]. In the same year autoloaders were added to the 3480. The autoloader held up to seven cartridges and automatically exchanged the current tape for a queued cartridge from the bottom of the loader. The 3490E drive released in 1989 used a 36-track format and provided an increase in native capacity to 800 GB [23, 84].

In the same year the 3480 was released, DEC released the TK50 Compac Tape drive which was intended for use with minicomputers rather than mainframes. The TK50 used a 22-track tape format, recording with a single channel head in a linear serpentine fashion. It had a data rate of 45 KB/s and a cartridge capacity of 94.5 MB [43]. The second generation, the TK70, was released in 1987 and provided 294 MB capacity at the same data rate as TK50 [44]. The third generation, the THZ01, which was later rebranded as the DLT260, was released in 1991 with a capacity of 2.6 GB and a data rate of 800 kB/s. The THZ01/DLT260 introduced the use of cylindrical guide rollers to guide the tape through the tape path. Over the next 8 years there were multiple follow-on versions of DLT (digital linear tape) each providing increases in capacity and data rate. The number of channels was first increased to two in the DLT2000 drive and then 4 channels in the DLT 7000. The final version, the DLT8000, was released in 1999 and had a capacity of 40 GB and a data rate of 6 MB/s. In 2000, Quantum introduced the SuperDLT (SDLT format) that used an optically read servo pattern on the backside of the tape for track following [176]. The first generation SDLT 220 had a capacity of 110 GB and a data rate of 16 MB/s. The final generation was released in 2007 with a capacity of 800 GB and a data rate of 60 MB/s.

In 1987, Storage Tech (STK) launched the Cimarron 4400 ACS (Automated Cartridge System) which became the first tape library to achieve major commercial success and introduced the concept of “Nearline” storage. The 4400 used IBM 3480 compliant tape drives and was based on a modular cylindrical library called a silo. A silo supported up to 16 tape drives, held up to 5,500 200 MB cartridges for a total capacity of over 1TB. The robot accessor had integrated cameras to read barcoded labels on the cartridges. In 1992 STK released the follow-on PowderHorn library that held up to 6,000 cartridges and provided up to 350 cartridges/hour accessor performance as well as smaller libraries called TimberWolf and WolfCreek that held up to 500 and 1,000 cartridges, respectively [2]. In 2005, Storage Tech launched the SL8500 library with a capacity of up to 10,088 cartridges and up to 64 drives but has since discontinued development.

In the late 1980’s in Japan, a variety of automated tape libraries were also developed. Examples include the NEC N7645 library announced in 1988 with a capacity of 6,250 cartridges, the Fujitsu F6455 with up to 5152 cartridges and the Hitachi H-6951-1 library with up to 6,560 cartridges [155].

Exabyte Corp. was founded in the mid 80’s with the goal of using consumer videotape technology [46, 157] for data storage. In 1987 Exabyte released the EXB-8200 drive, the first commercial data tape drive to use video helical scan technology. The EXB-8200 operated at a data rate of 246 kB/s, had a native capacity of up to 3.5 GB and used 8 mm particulate based consumer video tape media. In 1990, Exabyte released that EXB-8500 that provided an increased data rate of 500 kB/s and in 1992 they released the EXB-8505 that provided an increased native capacity of up to 5 GB. The EXB-8900 Mammoth drive was released in 1996 and used advanced metal evaporated (AME) media to achieve a 3 MB/s data rate and 20 GB capacity. The final version, the Mammoth-2 was released in 1999 with a data rate of 12 MB/s and capacity up to 60 GB.

In 1989, Sony released the first generation of Digital Data Storage (DDS) which was based on digital audio tape (DAT). It used helical scan recording [46, 157] on a 3.81 mm tape and achieved a capacity of 2 GB and a data rate of 183 kB/s. Over the next decade, 5 more generations of the technology were released, scaling capacity and data rate with each generation. The last two generations used 8 mm wide tape and operated at capacities 80 GB and 160 GB and data rates of 6.9 MB/s and 12 MB/s, respectively.

In 1993, IBM launched the 3495 robotic tape library that used a large (c.a. 400 kg), bright yellow, 6-axis industrial robot to serve cartridges to tape drives that were housed in a linear string of frames, as illustrated in Figure 36. The base model was 13.4 m long and held 5,660 cartridges. Three larger configurations were also available, the largest of which was 28 m long and held up to 64 tape drives equipped with autoloaders and up to 18,920 cartridges. The robot could mount 120 cartridges per hour but was not fast enough to keep 64 tape drives continuously busy [76, 84]. The internal IBM code name for the product was Caballero, but it was often referred to as Conan, as in Conan the Librarian. In the same year, IBM also released the 3494 library for mid-range/open systems. Cartridges were stored horizontally rather than vertically as in the 3495, which simplified the accessor design considerably. The 3494 was a more compact design with a much lighter custom designed robot that could perform up to 250 mounts per hour. The first version consisted of 2 frames, with two 3480 tape drives without autoloaders. This was extended to 8 frames in 1994 and then 16 frames in 1996 providing a capacity of up to 6420 cartridges. In 1997, dual accessors were added, improving the mount rate to 610 mounts per hour. A further enhancement to the 3494 was the introduction of the Virtual Tape Server (VTS). The VTS was a large disk cache and server responsible for managing the data and tape media and was placed between the host and tape library. The VTS enabled a much more efficient use of tape capacity and simplified migration to newer tape technology [84, 131]. In 2000, IBM released the 3584 library that could handle both LTO (see below) and DLT technology. The 3584 used a similar linear architecture as the 3494 and was configurable with 1 to 16 frames and up to 192 drives. In 2006 it was renamed the TS3500. In 2008 the capacity was enhanced through the introduction of HD (high density) slots in which multiple cartridges are stored one behind the other, providing a capacity of up to 1,320 LTO cartridges in a single frame [76, 84]. Starting around 2004 IBM also began offering a range of mini and mid-range libraries. Examples of other companies producing tape libraries in this period include Qualstar, ADIC (Advanced Digital Information Corporation), Hewlett Packard, and later additional companies including Overland and Spectralogic. IBM’s latest large scale tape library offerings include the TS4500 introduced in 2014 which has a similar architecture to the TS3500 and the single frame TS6000 Diamondback introduced in 2022.

In 1995, IBM released the 3590 Magstar MP (multi purpose) tape drive which was the first tape drive to implement track follow servo control. See Figure 37. Track follow servo was a key innovation that enabled a new era of much faster track density scaling that continues today. The 3590 used an amplitude based servo pattern that is described in reference [159]. The first version of the 3590 (Model B) had 128 tracks and provided a capacity of 20 GB and data rate of 9 MB/s. It was followed by the Models E and H in 1999 and 2002 which had capacities of 40 GB and 60 GB respectively; both operated at 14 MB/s.

In 1996, IBM introduced the 3570 tape drive targeting mid-range computer systems. See Figure 38. It used a dual reel cartridge and implemented a novel midpoint load design to minimize data access time and boasted an impressive load ready time of 6.7 s and average seek time of only 8 s. The 3570 was the first tape drive to use the TBS technique that has become standard in all recent linear tape drives. The 3570 operated at a data rate of 7 MB/s with an initial capacity of 5 GB that was extended to 7.5 GB in 1999 and 10 GB in 2002 [76].

In 1996, Sony introduced Advanced Intelligent Tape (AIT) which was based on helical scan recording [46, 157] and used AME media. The first generation had a capacity of up to 35 GB and a data rate of up to 4 MB/s. A total of five generations were released, scaling capacity and data rate to 400 GB and 24 MB/s, respectively, in the 5th generation that was released in 2006. Gen5 of AIT was the first tape drive to use giant magneto resistive (GMR) reader technology.

In 1999, Ecrix, which later merged with Exabyte corporation, released VXA tape based on an 8mm helical scan technology [46, 157]. The first generation, VXA-1 had a capacity of 33 GB and a data rate of 3 MB/s and was followed by two more generations in 2002 and 2005 that had capacities of 80 GB and 160 GB and data rates of 6 MB/s and 12 MB/s, respectively.

In the 1970’s, 80’s, and 90’s there was a proliferation of tape products and formats. In addition to those discussed above, other examples include the quarter inch cartridge (QIC) format launched in 1972, QIC-Wide launched in 1994, Travan launched in 1995, QIC-EXtra (QIC-EX) launched in 1996, and SLR (scalable linear recording) launched in 1997. Most of the formats of this era were proprietary. Incompatibility between the many formats made it difficult for customers to change technology and vendor and hence tended to lock customers in. In response to this situation, in the late 90’s IBM, HP and Seagate formed the LTO consortium with the goal of developing a new more open format that provided interchangeability between drives and media from different manufactures. In 1998, the consortium announced the LTO Roadmap. Initially two formats were planned, a single reel format called Ultrium and a dual reel, mid-point load format based on the IBM 3570 format that was called Accellis. The Accellis Gen 1 format was never commercialized and no follow on formats were developed. As a result the term LTO is currently used to refer exclusively to the Ultrium format. In 2000, IBM, HP, and Seagate each released LTO Ultrium Gen 1 drives (see Figure 39). LTO-1 had a capacity of 100 GB and a data rate of 20 MB/s. It used a linear serpentine recording format with an 8-track head that spanned only a quarter of the width of tape reducing the sensitivity to TDS effects by about a factor of 4. To read and write across the full width of tape, the IBM LTO-1 drive adopted a novel architecture that combined a coarse actuator with a fine actuator for track follow servo and replaced the air bearing tape guides used in early IBM drives with tape guide rollers. Another major innovation was the introduction of flat lapped heads built using HDD head fabrication technology on AlTiC wafers to replace the much larger contoured and slotted nickel-zinc ferrite heads used in earlier tape drives [20]. In addition to other advantages, the light weight flat lapped heads enabled higher bandwidth servo control and hence better track following performance. However, the most revolutionary innovation of LTO was interchangeability, i.e., any LTO1 cartridge provided by the multiple LTO licensed media vendors could be written/read in drives manufactured by any of the LTO drive vendors and then later written/read in a drive from any of the other vendors. LTO-1 drives and follow on generations were made in a higher performance FH form factor and a lower performance/lower cost HH form factor intended for lighter workloads. LTO tape is often referred to as a “mid range” tape solution in terms of performance and cost relative to “Enterprise” solutions often used with mainframes and low-cost solutions used with PCs. Competition between the multiple LTO drive and media providers resulted in more competitive pricing compared to proprietary formats and contributed to LTO becoming the dominant tape format. To date, nine generations of LTO have been brought to market. LTO-2 was introduced in 2002 with a 200 GB capacity and 40 MB/s data rate and introduced a PRML (partial response maximum likelihood) data channel and a more efficient 16/17 modulation code. LTO-2 drives were also backwards compatible in that they were able to read and write LTO-1 cartridges. Seagate’s tape business was renamed Certance in 2003 and then acquired by Quantum in 2004. LTO-3 was also released in 2004 with a new 16 track head format and again doubled capacity and data rate to 400 GB and 80 MB/s, respectively. The potential in linear tape recording to scale the number of parallel channels in a straightforward manner enabled faster data rate scaling compared to helical scan based technologies and was also a contributing factor to LTO becoming the dominant format. LTO-4 was released in 2007 with an 800 GB capacity and 120 MB/s data rate. Tandberg Storage also participated in the LTO consortium for three generations, releasing HH versions of LTO-2 (TS400, 2005), LTO-3 (TS800, 2007), and LTO-4 (TS1600, 2008) drives. LTO-5 was released in 2010 with a 1.5 TB capacity and 140 MB/s data rate followed by LTO6 in 2012 with a 2.5 TB capacity and 160 MB/s data rate. LTO-5 introduced the capability to partition the tape into 2 tape partitions and LTO-6 extended this capability to 4 partitions. LTO-7 was released in 2015 with a 6 TB capacity and used a 32-channel head to enable a data rate of 300 MB/s. The two most recent generations, LTO-8 (2017) and LTO-9 (2021) use the same 32 channel head format and operate at capacities of 12 TB and 18 TB and data rates of 360 MB/s and 400 MB/s, respectively. The latest LTO roadmap describes 5 future generations of the technology, each of which is expected to double the capacity of the previous generation.

SUN Microsystems acquired STK in 2005 and subsequently launched a new family of Enterprise class drives in 2006 that used a linear serpentine format and a single reel cartridge. The first generation, the T10000 had a capacity of 500 GB and a data rate of 120 MB/s. The second generation, the T10000B was released in 2008 with a capacity of 1 TB and a data rate of 120 MB/s. In 2009, Oracle acquired SUN and subsequently released the T10000C in 2011 with a 5 TB capacity and 240 MB/s data rate. In 2013, Oracle launched the T10000D drive with a capacity of 8.5 TB and a data rate of 252 MB/s. A T10000E drive was planned but eventually cancelled when Oracle stopped tape development in 2016.

In 2003, IBM also launched a new family of Enterprise class drives. The first generation was branded 3592 and operated at a capacity of 300 GB and a data rate of 40 MB/s. See Figure 39. The second generation, the TS1120, operated with a capacity of 700 GB and a data rate of 100 MB/s. The TS1120 was the first storage device to provide built-in hardware-based encryption which later also became standard in LTO drives. The TS1130 was released in 2008 and used GMR reader technology to provide a capacity of 1 TB and data rate of 160 MB/s. Compared to LTO drives of the same time frame, Enterprise class drives typically provided higher performance in terms of capacity, data rate, access time and error rate and in addition, enabled “up formatting” of the previous generation of media to a higher capacity using the latest generation of tape drive. The TS1140 was released in 2011 with a 4 TB capacity and a data rate of 250 MB/s. The TS1140 introduced flangeless tape guides that reduced media edge wear as well as active skew control and technology to improve rewrite performance. In addition, the TS1140 introduced a new data dependent noise predictive maximum likelihood (DD-NPML) data channel, a more powerful ECC scheme that doubled the length of the C2 code, a novel 3 module head design with 32 active channels and a new servo pattern. In 2014, the TS1150 was released with a 10 TB capacity and 360 MB/s data rate. The TS1150 introduced a new read head based on TMR (tunneling magneto resistive) readers, the first use of TMR technology in tape. The TS1155 was released in 2017 with a 360 MB/s data rate and a 15 TB capacity that was partially enabled by the introduction of a new writer with notched poles and a new high moment liner technology. In 2018 the TS1160 introduced a new tension based active TDS compensation scheme that helped to enable a 20 TB capacity and 400 MB/s data rate. The TS1160 introduced new ECC technology that included a new C2 code that again doubled the code length as well as a novel iterative decoding scheme. The latest generation, the TS1170 was released in 2023 with a 50 TB capacity and 400 MB/s data rate. Many of the technologies introduced in the TS11xx family were also implemented in subsequent generations of LTO drives and were critical to enabling the scaling of IBM’s LTO drives. Currently, LTO and IBM TS11xx Enterprise are the only tape formats still under active development. However, despite consolidation in the industry and convergence to two formats, tape has a healthy ecosystem with companies including IBM, HPE, Dell, Quantum, Spectralogic, Overland and Oracle selling tape drives and libraries, two media manufactures and multiple brands of media. Moreover, tape has significant potential for continued scaling as discussed in Section 12.

Apart from the UNISERVO 1 tape system that used a metal tape made from nickel plated phosphor bronze alloy, all other commercial data tape media have used a polymer substrate coated with a thin layer of magnetic material. Companies who have manufactured magnetic data tape media include: BASF, Datatape/Graham Magnetics, Fujifilm, Imation, Maxell, Memorex, SONY, TDK, and 3M, with both Fujifilm and Sony still actively developing new tape media products. Two types of magnetic coating have been used: particulate and metal evaporated (ME) coatings. In particulate based tape, the recording layer (mag-layer) is made up of small magnetic particles fixed to the polymer substrate with a binder (i.e., “glued” to the substrate). The tape is manufactured by coating the particles onto the substrate in a liquid slurry that also contains solvent and binder, followed by evaporation of the solvent to “bind” the particles to the substrate. In ME tape the mag-layer consists of a continuous, granular magnetic metal film deposited on the polymer substrate under vacuum by evaporation. Exabyte’s Mammoth 1 and 2 as well as Sony’s AIT and SAIT tape, used a metal evaporated mag-layer technology that is described by Kawana et al. [130]. All other polymer tape media to date, including state-of-the-art products, are particulate based and, therefore, we will focus on the evolution of particulate tape technology here.

Early generations of particulate tape, such as that used for the IBM 726, had a simple bilayer structure consisting of an approximately 10 $\mu m$ thick layer of magnetic particles and binder coated directly onto an acetate substrate. The magnetic particles were an acicular form of gamma ferric oxide with particle lengths on the order of 0.2 to 0.8 microns. To achieve the required surface quality, i.e., low defect density, “IBM designed and built the world’s most advanced tape coater in Poughkeepsie, and the first clean room used in manufacturing.” [23]. In later generations of tape, a thin, carbon loaded back-coat layer was added to control tribo-charging effects and improve the winding/unwinding properties of the tape. Over time, the thickness of the magnetic recording layer and the size of the magnetic particles have been continuously reduced to enable scaling of the areal recording density, as discussed in Section 3. In the 1980’s, FeCo particles with a length on the order of 250 nm were used in a mag-layer with a thickness on the order of 3–5 $\mu m$. In the 1990’s, CrO2 particles were used in the IBM 3480 and then 3490 tape media. In the late 1990’s Fujifilm developed a dual coating technique in which a non-magnetic undercoat and a much thinner mag-layer are deposited simultaneously. This technology enabled a significant reduction in the mag-layer thickness to around 300 nm initially with further decreases in subsequent generations to around 100 nm over the next decade. The introduction of a more advanced dual layer coating technology in 2011 enabled a further reduction to around 60 nm [154].

Early generations of LTO and Enterprise tape media (IBM TS11xx and STK T10000) used MP (metal particle) technology which was based on acellular particles with a CoFe core and a (non-magnetic) Yttrium based shell a few nm thick which prevented oxidation of the magnetic core. The magnetic axis of the particles was aligned in the longitudinal direction of the tape, i.e., parallel to the tape transport direction through the application of a magnetic field during the coating process. The first generation of LTO media used particles with a volume on the order of 10,000 $n{m}^3$. LTO-1 media had a mag-layer thickness of 220 nm which was reduced to 100 nm for Gen2 [20]. With each generation of media, the volume of the particles was continually reduced. Reductions were also made in the width of the distribution of particle sizes to improve SNR. LTO Gen4 media, released in 2007 used a particle volume of about 4500 $n{m}^3$ [82] and LTO Gen 5 (2010) used particles with a volume of about 2800 $n{m}^3$ [28]. Below a volume of around 2800 $n{m}^3$ it is difficult to maintain a sufficient particle coercivity to ensure the thermal stability of recorded data [154], as discussed in Section 2.

In 2011, Barium Ferrite (BaFe) particle technology was introduced in Enterprise tape media and then later also adopted in LTO media. LTO Gen6 supported both MP and BaFe media, whereas generations 7 to 9 are based exclusively on BaFe. BaFe particles have a hexagonal platelet shape and a magnetization that results from crystalline order rather than shape anisotropy. The coercivity of the particles can be tuned by doping the particles with elements such as Co, Zi, or Ti. Moreover, BaFe (BaFe₁₂O₁₉) is an oxide and, therefore, does not require a non-magnetic shell to protect against oxidation. As a result, BaFe particles can be scaled to much smaller sizes than MP technology. Initial generations of BaFe media (e.g., IBM TS1140 JC tape) utilized a particle volume of 2,100 $n{m}^3$ and a mag layer thickness of about 70 nm [71]. In the TS1150 JD media, the particle volume was reduced to around 1950 $n{m}^3$ [137]. In both JC and JD media, the particles had an essentially random orientation. The particle volume of the TS1160 JE media was further reduced to 1,700 $n{m}^3$ and the particles were partially oriented in the perpendicular direction through the application of a magnetic field during the coating process [69].

The effective distance between the read and write transducers and the magnetic particles, referred to as magnetic spacing, is determined predominantly by the roughness of the surface of the mag-layer. To enable areal density scaling, tape roughness has been continuously reduced as new generations of media were introduced. For example, in the early 90’s, the roughness of the mag-layer as measured by optical interferometry, was on the order of Ra 7 nm. The introduction of dual coat technology in the late 90’s enabled a reduction to around Ra 5 nm and a continual decrease in subsequent generations to around 1.5 nm by 2012 [154].

Early open reel tape products used an Acetate substrate which was later replaced by PET to improve tape robustness against breakage during the continuous start/stop usage pattern of early tape drives. Later, PEN was also used as a substrate (for example in the IBM 3590 Extended cartridge). Recent generations of LTO tape have used PET, PEN, and Spaltan, which is a blend of PET and Aramid, whereas the most recent generations of enterprise tape media have used Aramid as a substrate. Of these four materials, Aramid is the most robust and has the lowest dimensional change versus environmental conditions, but is the most expensive.

Over the history of tape, the thickness of the tape media has been gradually reduced to enable longer tape lengths and hence higher reel/cartridge capacities. For example, the 8-inch diameter (203 mm) reel used in the IBM 726 held 1,200 ft (365 m) of tape that was about 58 $\mu m$ thick. The tape cartridge introduced with the IBM 3480 had a diameter of about 96 mm and held 541 ft (165 m) of tape that was about 30 $\mu m$ thick. LTO cartridges have about the same reel diameter but in Gen 1 they held 609 m of 8.6 $\mu m$ thick tape and by Gen 9 the length was scaled to 1,035 m of 5.2 $\mu m$ thick.

The most common tape width format is ½ inch (12.65 mm), introduced with the Uniservo 1 and IBM 726 and still used in LTO and recent Enterprise tape media. However, over tape’s 70+ year history a variety of other formats have been used, including 4 inch, 1 inch, ¾ inch, 8 mm, ¼ inch, 4 mm and 1/16 inch.

The Uniservo 1 and IBM 726 used open reels with an 8-inch diameter. In 1953, IBM introduced the 727 magnetic tape unit with 10.5-inch reels that became a de facto industry standard for around 25 years [46]. In 1984, IBM introduced the 3480 that replaced the 10.5${}^{\prime \prime }$ reels with a $4^{\prime \prime }$ $\times$ $5^{\prime \prime }$ $\times$ $1^{\prime \prime }$ ($101.6$ mm $\times$ $127$ mm $\times$ $25.4$ mm) cartridge. More recent generations of IBM and STK/Oracle Enterprise tape have used a similar size cartridge. The LTO tape cartridge format is slightly smaller (102 mm x 105.4 mm x 21.5 mm) and is similar to the dimensions of the DLT cartridge. A variety of dual reel cartridge formats have also been used, such as in the DDS, IBM 3570, and AIT. Dual reel cartridges can provide an advantage in terms of access time, but have a lower volumetric storage efficiency.

In the first generation of AIT, released in 1996, Sony introduced the concept of a CM with their MIC (memory in cassette) technology that was used to store meta data and improve data access performance. In the second generation of AIT, Sony introduced a contactless CM technology called R-MIC and used it to enable the first WORM tape cartridge. LTO and recent Enterprise tape products also use a contactless CM and have also offered WORM capability in all recent generations, starting from STK T9940A, IBM 3592, and LTO-3.

State-of-the-art media technology is described in Section 3 and potential future media technologies are discussed in Section 12.

Over the more than 70 year history of magnetic tape storage there have been profound changes in the way tape is utilized in the computing environment, and corresponding changes in the types of software available for making use of tape storage.

Magnetic tape storage and the wide availability of general-purpose computers both appeared in the early years of the 1950s. However, neither of these technologies entered a vacuum; in fact, both replaced or augmented equipment and processes that had been developed over many decades.

In the 1890s, Herman Hollerith’s use of punched-card equipment to process the 1890 U.S. Census ushered in an era of card-based record-keeping and accounting procedures that would form the basis for most early automated business processing. Hollerith’s Tabulating Machine Company eventually became a part of the Computing-Tabulating-Recording (CTR) Company, which was renamed in 1924 to International Business Machines (IBM) [12].

By the 1930s, IBM and other manufacturers had a wide array of card-based equipment available, including card readers, sorters, printers, collators, and even plugboard-programmable “tabulators” capable of counting, addition and subtraction, and providing multiple levels of accumulators. This equipment was often referred to collectively as Electric Accounting Machines (EAM) [12].