Technology card
ShippingHard Disk Drive
HDD · hard drive
Read capacity, recording density and transfer-rate specifications separately: more tracks can hold more data without making a single track pass its head any faster.
Glossary entry on hesela.dev · JSON record · Areal density · Linear recording density · Track density · Zoned bit recording · Track skew · Read-on-arrival
Related
On this page7 sections
In one minute
A hard disk drive is a non-volatile storage device that records data magnetically on rotating platters and accesses it with movable read/write heads.
Read-on-arrival changes the order in which requested sectors are collected, then buffers them for logical-order delivery. The historical name zero-latency read is misleading: a full physical track can be gathered in one revolution after positioning, but seeking, transfer and host delivery are not instantaneous. Track alignment matters; a similarly sized logical range can cross a track boundary. [5] [7]
In a constructed eight-sector track, arriving at sector 3 and requesting sectors 0 through 7 takes 13 sector ticks if the reader waits for sector 0 first. Reading 3,4,5,6,7,0,1,2 instead takes eight ticks, with the same logical-order result. The companion code enumerates all eight boundary phases and includes a one-sector no-benefit control. Its units are sector-transfer ticks, not measured milliseconds; seeks, queues and host transfer are excluded. [3]
A device-level mechanism is not an application speedup guarantee. FAST 2002 tested the feature through timing-accurate emulation, and its PostMark workload barely changed. That result belongs to the reported historical workload and model, not a physical firmware modification or a claim about current HDDs. [2]
Spindle speed supplies a period, not an application IOPS ceiling. At 7,200 RPM one revolution takes about 8.33 ms. A uniform target-relative phase after positioning gives a mean remaining wait of about 4.17 ms; the platter keeps rotating during a seek. One revolution bounds that single uninterrupted wait for the next sector passage, not complete request latency. [4]
Under these assumptions, the reciprocal unrounded mean gives a model rate of 240 operations/s for continuously busy serial service with rotation as the sole cost. Locality, sequential access, read-ahead and rotation-aware scheduling change the service process, so this is not a workload-independent limit. The 1994 disk-modeling paper explains these mechanisms, not current-drive performance. The companion six-row calculation preserves its original numbers while making those assumptions explicit. [5] [6]
A rate needs the right denominator. In a constructed serial sequence with no idle gaps, 1 ms and 9 ms service times give two completions in 10 ms: 200 operations/s. Averaging the individual reciprocal times gives about 555.6 operations/s instead, which is not that sequence's throughput. This is arithmetic, not measured HDD performance. Idle gaps must count toward elapsed time, and overlapping request latencies are not additive elapsed time. The companion calculation publishes the example and its generator. [6]
The actuator positions a head over a track; rotation brings successive sectors past it. A drive specification therefore needs several separate quantities: accessible bytes, seek and rotational behavior, sustained media transfer and host-interface rate. A cached response is a different workload from reading the magnetic surface. [26]
Areal density combines two dimensions. Linear recording density measures bits along a track, commonly in bits per inch (BPI). Track density measures tracks across the surface, in tracks per inch (TPI). Their local product is bits per square inch when both describe the same recording conditions and compatible bit-count conventions. Seagate's 2019 presentation reports 2,194 kBPI x 900 kTPI, about 1.97 Tbit/in2, as a laboratory demonstration, not a shipping-drive capacity. [24]
The calculation below isolates those dimensions. At the same radius and rotation speed, increasing only TPI raises areal density but leaves the number of bits passing one head per revolution unchanged. Increasing BPI raises that count. The table assumes the same bit-count convention throughout and is a geometric comparison, not a benchmark or a claim that every such design is physically achievable.
Zoned bit recording (ZBR) groups tracks into radial recording zones, with more sectors per track in outer zones. Zone-dependent recording frequency makes use of the longer outer circumference. At a fixed RPM, the outer track has higher tangential speed, not higher angular speed. ZBR describes a layout; linear recording density measures bits per unit length. [10]
Within a recording zone containing S payload sectors per track, with B payload bytes per sector and fixed RPM, S x B x RPM / 60 gives an idealized payload rate in bytes/s for one head. More payload bytes per revolution can raise this rate without changing RPM. Track/head switches, seeks, retries, cache behavior and workload still affect sustained host throughput; the formula does not establish a universal outer-to-inner speed ratio.
A geometry field is not necessarily a local measurement. The ST3655 manual reports average sectors per track, not a zone-local S for the formula above. Dempsey (FAST 2003) found that the IBM Microdrive's translated cylinder/head/sector address was not its true physical location, and instead inferred layout from timed operations. Its validation covered two mobile disk models, not today's entire HDD population. [9] [8]
Track skew is a sector-position offset across successive tracks that accommodates rotation during a track switch. Atropos (FAST 2004) used track size and skew to arrange two-dimensional access, evaluating a prototype with four Maxtor Atlas 10K III disks. This explains why boundary crossings matter; it is not a universal modern-drive tuning rule or a reason to assume that adjacent logical blocks have identical access cost. [7]
A skew count is not a duration. In Atropos's zoned-geometry discussion, track-switch time stays fixed while the number of sectors passing during that interval changes between zones. Comparing sector offsets without the local sectors-per-track and rotation period can therefore misrepresent switching cost. This is a distinction in the paper's model, not a measurement of a current drive. [7]
A recording zone is not automatically a host-managed write zone. ZBC/ZAC and ZNS define zone access and management behavior, including write pointers and resets for sequential-write-required zones. ZBR alone does not impose those rules. Nor must its geometry be invisible: the 1994 ST3655 SCSI manual documents a Notch mode page reporting recording-zone boundaries. [11] [9]
The word zone has a third meaning here. In the ST3655 Format Device page (03H), Tracks per zone describes defect-management groups for spare allocation. Those groups are not the recording notches described by the Notch page (0CH). [9]
The same Notch page reports LPN = 0: boundaries are cylinder/head coordinates, not logical block addresses. Active notch = 0 selects parameters for all notches, not a first individual notch. Read the coordinate type and selector before interpreting the boundary fields. [9]
Check which bits a density figure counts. Seagate's 2018 demonstration reports user areal density after allowing for error-correction overhead; its raw recorded density is 13% higher. For that example, raw density = 1.13 x user density, so the reverse conversion divides by 1.13 instead of subtracting 13% of the raw value. This factor belongs to that demonstration, not all magnetic recording. [25]
Product tables also mix scopes. Seagate's May 2014 manual labels BPI as peak and TPI as average, and lists internal, sustained and interface rates separately. Preserve those labels: multiplying unrelated maxima does not establish a whole-disk average, and an interface ceiling is not a media-transfer measurement. [26]
User capacity also depends on active recording surfaces, usable track area, formatting and reserved space. A platter's full geometric area and a peak density figure are insufficient to calculate the bytes exposed to the host. For host-visible logical capacity, multiply the logical-block count by bytes per logical block; filesystem-usable space is a separate quantity. [26]
Density creates physical trade-offs as well as arithmetic ones. In a simplified grain model, the anisotropy energy barrier is Ku times grain volume V, compared with thermal energy kB times temperature T. Reducing V at fixed Ku and T reduces that barrier ratio; this thermal stability issue is not merely magnetic interference from neighboring grains. No universal retention lifetime follows from an areal-density number. [27]
Keep the recording layers distinct when choosing software. CMR and SMR describe track layout; PMR describes magnetization orientation; HAMR adds heat-assisted writing. Drive-managed SMR exposes ordinary block writes, while host-managed SMR exposes sequential-write constraints. Capacity or a HAMR label alone does not identify the host contract. [28]
An SSHD combines magnetic storage with flash inside one drive. [13] Seagate's 2011 Momentus XT description uses a learned selection of read blocks in NAND, with magnetic writes preceding selected flash copies. Consequently, repeated cached reads and uncached reads exercise different paths. A useful comparison records warm-up, working-set size and read/write mix instead of assuming that every access has SSD-like latency. [14]
Read the memory fields separately. The FireCuda ST2000LX001 family manual lists 8 GB of NAND and a 128 MB buffer; neither is extra user capacity to add to its formatted-capacity table. It also warns that DRAM-cached writes can be lost when power is interrupted. Nonvolatile flash somewhere in a drive does not establish persistence for every acknowledged write. [15]
Do not generalize one generation's cache policy. Seagate's Enterprise Turbo SSHD separately specifies a 32 GB NAND cache and an 8 MB nonvolatile cache for faster write responses. Compare the documented read and write paths for the exact model, including its power-loss behavior. A cache within the same device is not an independent backup. [16]
Hybrid storage is broader than an SSHD. Hystor combined separate SSD and HDD devices in a Linux prototype. Its placement metric combined access frequency with request size to favor costly disk accesses, not simply the most popular blocks. The study supports evaluating workload and policy together; it does not establish the behavior or speed of a commercial SSHD. [12]
Copy placement changes rotational wait
Rotational replication adds angular choices for a read. Copy count alone does not specify the benefit. This exact illustration uses a 12-coordinate circle at 7,200 RPM with continuous uniform arrival phase after positioning, not measured drive requests.
Mean time to encounter the first or last copy position, in milliseconds. Last-copy encounter excludes payload transfer and is not durable write completion.
| Layout | First copy (ms) | Last copy (ms) | Payload fraction |
|---|---|---|---|
| One copy | 4.1667 | 4.1667 | 1 |
| Two evenly spaced | 2.0833 | 6.2500 | 1/2 |
| Two clustered | 3.5301 | 4.8032 | 1/2 |
| Three evenly spaced | 1.3889 | 6.9444 | 1/3 |
| Four evenly spaced | 1.0417 | 7.2917 | 1/4 |
For gap fractions g_i summing to one, the mean next-copy wait is R * sum(g_i squared) / 2. Equal gaps minimize it at R/(2D); two clustered copies do not halve the single-copy mean. The largest gap governs the worst-phase wait. Payload fraction 1/D counts only redundant payload, not formatting, spares or a separate reliability mirror.
Under the same idealization, a foreground sweep reaches the last copy after R minus the first-copy mean. An equal read/write mix therefore stays at 4.1667 ms for every row. This excludes seeking, transfer, track switches, queueing and consistency protocols; it is not an application speedup forecast. Copies on one disk do not protect against losing that disk.
Historical context: Yu et al., OSDI 2000, section 2.2. Hesela's derivation and reproduction instructions, JSON rows, generator and exhaustive tests distinguish chosen geometry from observed performance.
Density and transfer
The examples hold track radius, RPM and bit-count convention fixed. Actual host throughput also depends on formatting, error correction, servo overhead, cache behavior and the workload. Tape uses longitudinal motion along a strip; optical discs are rotating media, not tape-like linear-access devices.
Illustrative density changes. Relative bits per revolution use the same radius and bit-count convention; RPM is also held fixed. M means one million.
- Baseline: 2 x 0.5
- Density (Tbit/in2)1.00
- Bits per revolution1.00 x baseline
- 25% more tracks: 2 x 0.625
- Density (Tbit/in2)1.25
- Bits per revolution1.00 x baseline
- 25% more bits along a track: 2.5 x 0.5
- Density (Tbit/in2)1.25
- Bits per revolution1.25 x baseline
Uncertainty notes
- Reviewed 24 October 2026. The table is generated arithmetic with explicit fixed conditions; it does not predict a specific drive's throughput.
- The 2014 manual and 2019 demonstration are dated examples of specification scope, not current product limits.
- The 1956 milestone refers to the commercial IBM 350 RAMAC disk unit, not the invention of magnetic storage.
- Backblaze failure rates describe its observed fleet and methodology, not a universal HDD reliability guarantee.
Sources
- 01Yu et al.: Trading Capacity for Performance in a Disk Array, OSDI 2000, pp. 243-258, sections 2.2 and 3.2-3.4 · accessed 2026-11-10
- 02Griffin, Schindler, Schlosser, Bucy and Ganger: Timing-accurate Storage Emulation, FAST 2002, section 6.1 and Table 5 · accessed 2026-11-09
- 03Hesela: exact eight-sector read-on-arrival illustration, code, assumptions and tests · accessed 2026-11-09
- 04Arpaci-Dusseau and Arpaci-Dusseau: OSTEP, chapter 37, sections 37.3-37.5 · accessed 2026-11-06
- 05Ruemmler and Wilkes: An Introduction to Disk Drive Modeling, IEEE Computer 27(3):17-29 (March 1994), controller and rotational-position modeling sections · accessed 2026-11-06
- 06Hesela: HDD rotational latency by spindle speed and conditional serial-model rates · accessed 2026-11-06
- 07Schindler, Schlosser, Shao, Ailamaki and Ganger: Atropos, USENIX FAST 2004, sections 3.2, 3.3.1 and 5.1-5.3 · accessed 2026-10-24
- 08Zedlewski, Sobti, Garg, Zheng, Krishnamurthy and Wang: Modeling Hard-Disk Power Consumption, USENIX FAST 2003, sections 3.2 and 4 · accessed 2026-10-24
- 09Seagate: ST3655 Family SCSI Product Manual, Rev. A, 18 January 1994, sections C.3 and C.7, pp. 96-98 and 108-109 · accessed 2026-10-23
- 10Leow et al.: Adaptive zone frequency compensation for disc drives, US6757119B2 (2004), Background of the Invention · accessed 2026-10-23
- 11Western Digital, Zoned Storage: device models, zone types and management commands · accessed 2026-10-23
- 12Chen, Koufaty and Zhang: Hystor, ICS 2011, sections 3 and 7; DOI 10.1145/1995896.1995902 · accessed 2026-10-20
- 13Hesela: SSHD definition and cache-policy boundaries · accessed 2026-10-20
- 14Seagate: SSHD Goes the Distance, MB618.2-1107US, July 2011, pp. 2-3 (manufacturer marketing bulletin) · accessed 2026-10-20
- 15Seagate: FireCuda Product Manual, 100802299 Rev. H, February 2019, sections 1, 2.1 and 2.11.1 · accessed 2026-10-20
- 16Seagate: Enterprise Turbo SSHD datasheet, DS1802.1-1309US, September 2013, pp. 1-2 · accessed 2026-10-20
- 17Computer History Museum — 1956: First commercial hard disk drive shipped (The Storage Engine, IBM Model 350 RAMAC) · accessed 2026-10-16
- 18Computer History Museum, The Storage Engine — 1973: “Winchester” pioneers key HDD technology · accessed 2026-10-11
- 19Computer History Museum — IBM 3340 Direct Access Storage Facility, 1973 (catalogue document) · accessed 2026-10-11
- 20Backblaze — Drive Stats for 2025 (fleet methodology and AFR) · accessed 2026-10-06
- 21Backblaze — Open hard drive stats data set description · accessed 2026-10-06
- 22Seagate — CMR/SMR product matrix (recording types in market) · accessed 2026-10-06
- 23Western Digital — SMR HDD technology white paper (HDD recording context) · accessed 2026-10-06
- 24Seagate: HDD Technology Update, Library of Congress DSA 2019, slide 14 · accessed 2026-10-19
- 25Seagate: Disk Drive Technology Trends, Library of Congress DSA 2018, slide 10 · accessed 2026-10-19
- 26Seagate: Enterprise Capacity 3.5 HDD v4 SAS Product Manual, 100726011 Rev. A, May 2014, sections 3.5 and 4.1 · accessed 2026-10-19
- 27Seagate: HDD Storage Trends and Emerging Technology, Library of Congress DSA 2016, slide 5 · accessed 2026-10-19
- 28Zoned Storage: Shingled Magnetic Recording Hard Disks, overview and interface implementations · accessed 2026-10-19