Our hard drive test platform has served admirably for more than four and a half years now. We first started using the system in early 2005, a time that seems oh-so long ago. Back then, bundles of garbage sub-prime mortgages were sought-after AAA securities, my inbox was flooded with spam pimping herbal Viagra rather than Acai Berry, and I’d never heard of Heidi Montag or Spencer Pratt. Those were the days.
Among PC enthusiasts, a gigabyte was still considered to be quite a lot of system memory. Windows was arguably in a better place with XP, albeit at only 32 bits. We at least had 64-bit CPUs, but only solitary coresAthlon 64s and, shudder, Prescott Pentium 4s. The GPUs of the day were DirectX 9 parts with discrete pixel and vertex shaders, and not very many of either.
We’ve come a long way in four and a half years. During that time, our venerable test rig has seen no fewer than 70 different storage solutions run its gauntlet. It’s punished everything from 10k-RPM Raptors to exotic solid-state drives to 2.5″ mobile hard drives at three different spindle speeds. And, of course, it’s chewed through a cornucopia of 3.5″ desktop drives.
As Windows 7 approaches and a new SATA 6Gbps standard looms on the horizon, the time has come to put our trusty storage test platform out to pasture. However, before I tear the system down and retire its components to the dusty stacks of old hardware boxes from whence they came, it’s only fitting that we look back on nearly five years of storage performance data.
Much has happened in the hard drive world since 2005. Old-school “parallel” ATA has all but been wiped from our collective consciousness by Serial ATA, banishing bulky ribbon cables from our systems. Mechanical hard drives have switched from longitudinal to perpendicular recording, enabling substantial increases in areal density and overall drive capacity. Hard drives have become smarter, too, adopting command queuing schemes previously reserved for high-end SCSI gear. Oh, and we’ve witnessed the birth of an entirely new class of solid-state storage solutions that represent perhaps the first real paradigm shift in the industry.
To explore how these factors have shaped hard drive performance, we’ve squeezed several years worth of test data into some seriously frightening graphs. Join us as we take a trip down memory lane to see just how far PC storage has come.
Our testing methods
Our test rig’s hardware may look antiquated by today’s standards, but its Intel ICH7R storage controller isn’t all that different from what you get in the latest ICH10R today. Both conform to the second-generation “SATA II” 3Gbps Serial ATA specification, offering support for Native Command Queuing and 300MB/s host-to-disk transfer rates. The next-gen 6Gbps Serial ATA specification wasn’t completed until this year, and it’s not expected to make it to motherboards until the fall.
Even with only a Pentium 4 Extreme Edition onboard, this test system still has plenty of horsepower for the storage-intensive tests we use to probe hard drives for weakness. However, we should note that because Windows XP was optimized for mechanical storage, it’s not an ideal OS for some of the solid-state drives. Our goal here isn’t to look at the fine details of each drive’s performance, but to observe overall trends that should be unaffected by oddball outliers. Besides, we’ve already addressed SSD performance in Windows Vista.
Oh, and rather than testing SSDs in factory-fresh form, we first put them into a simulated used state devoid of empty flash pages. We think this better represents long-term SSD performance.
Speaking of oddballs, a few of the configs we’ve tested over the years are far removed from what we’d consider practical solutions for even high-end desktop use. Neither of the DRAM-based solid-state drives we’ve usedGigabyte’s i-RAM and ACard’s ANS-9010offer enough storage capacity for even just an OS and applications drive. Our four-way Intel X25-E Extreme RAID 0 array, running on an Adaptec RAID 5405 card rather than the ICH7R, also costs about as much as small cara Tata Nano, but still. These outlandish configs may not be reasonable for desktop use, but they do give us an idea of what the exotic reaches of the storage spectrum have to offer.
Getting all this test data consolidated and graphed in a single Excel spreadsheet proved to be a major undertaking. And that was before I tried distill the data down to a reasonable number of readable graphs. With 70 results for nearly every test, readability goes out the window pretty quickly. In an attempt to make things easier to interpret, we’ve segmented drives into 3.5″, 2.5″, 10k-RPM, and SSD categories and color-coded the results accordingly.
Before we get started, I should also note that you won’t find boot time, noise level, or power consumption results in this retrospective. Over the last few years, we’ve switched up our test methods for each slightly, so we can’t compare all our results directly. You’ll find that a couple of drive configurations are missing from the iPEAK results, as well. We didn’t introduce iPEAK to the test suite until August of 2005. Oh, and the i-RAM doesn’t appear in our game level load tests or in WorldBench because its 4GB capacity wasn’t big enough to accommodate the necessary files.
One last time, here are the finer details of our test system.
|Processor||Pentium 4 Extreme Edition 3.4GHz|
|System bus||800MHz (200MHz quad-pumped)|
|Motherboard||Asus P5WD2 Premium|
|North bridge||Intel 955X MCH|
|South bridge||Intel ICH7R|
|Chipset drivers||Chipset 188.8.131.523
|Memory size||1GB (2 DIMMs)|
|Memory type||Micron DDR2 SDRAM at 533MHz|
|CAS latency (CL)||3|
|RAS to CAS delay (tRCD)||3|
|RAS precharge (tRP)||3|
|Cycle time (tRAS)||8|
|Graphics||Radeon X700 Pro 256MB with CATALYST 5.7 drivers|
ACard ANS-9010 RAID 0
Hitachi Deskstar 7K1000
Hitachi Deskstar 7K500
Hitachi Deskstar E7K1000
Hitachi Deskstar T7K250
Hitachi Travelstar 5K100
Hitachi Travelstar 5K500.B
Hitachi Travelstar 7K100
Intel X25-E Extreme
4 x Intel X25-E in RAID 0
Maxtor DiamondMax 10
Maxtor DiamondMax 11
OCZ SATA II
Samsung Spinpoint F1
Samsung Spinpoint M7
Samsung Spinpoint T
Seagate Barracuda 7200.10
Seagate Barracuda 7200.11
Seagate Barracuda 7200.11 1.5TB
Seagate Barracuda 7200.12
Seagate Barracuda 7200.7 NCQ
Seagate Barracuda 7200.8
Seagate Barracuda 7200.9 160GB
Seagate Barracuda 7200.9 500GB
Seagate Barracuda ES
Seagate Barracuda ES.2
Seagate Momentus 5400.2
Seagate Momentus 5400.3
Seagate Momentus 5400.4
Seagate Momentus 5400.6
Seagate Momentus 7200.1 PATA
Seagate Momentus 7200.1 SATA
Seagate Momentus 7200.3
Seagate Momentus 7200.4
Super Talent IDE Flash
Super Talent MasterDrive MX
Super Talent SATA25
Super Talent UltraDrive ME
WD Caviar Black
WD Caviar GP
WD Caviar Green
WD Caviar Green 2TB
WD Caviar SE16 250GB
WD Caviar SE16 500GB
WD Caviar SE16 640GB
WD Caviar SE16 750GB
WD Caviar RE2 400GB
WD Caviar RE2 500GB
WD Raptor WD1500ADFD
WD Raptor WD360GD
WD Raptor WD740GD
WD Raptor X
WD Scorpio Black
WD Scorpio Blue 320GB
WD Scorpio Blue 500GB
WD Scorpio WD1200VE
WD VelociRaptor VR150
|OS||Windows XP Professional|
|OS updates||Service Pack 2|
Thanks to NCIX for getting us the Deskstar 7K1000 and SpinPoint F1.
Our test system was powered by an OCZ PowerStream power supply unit.
We used the following versions of our test applications:
- WorldBench 5.0
- Intel IOMeter v2004.07.30
- Xbit Labs File Copy Test v1.0 beta 13
- HD Tach v3.01
- Far Cry v1.3
- DOOM 3
- Intel iPEAK Storage Performance Toolkit 3.0
The test systems’ Windows desktop was set at 1280×1024 in 32-bit color at an 85Hz screen refresh rate. Vertical refresh sync (vsync) was disabled for all tests.
All the tests and methods we employed are publicly available and reproducible. If you have questions about our methods, hit our forums to talk with us about them.
Performance testing is our bread and butter, but it’s far from the only element to the storage equation. Capacity counts for a lot, and that’s where we’ve seen some of the biggest gains over the years.
The anemic capacity of early SSDs skews the above graph some, but if you just pay attention to the yellow bars, it’s clear that the storage capacity of mechanical hard drives has grown by leaps and bounds. Back when we started using this test system, the biggest hard drive in the market weighed in at 250GB. Today, that’s half the capacity of just one of a 2TB Caviar Green’s four platters.
An eightfold increase in 3.5″ storage capacity in less than five years is quite an achievement. The capacity of 2.5″ notebook drives has grown, too, from a maximum of just 100GB to a half-terabyte today.
I have a soft spot for Western Digital’s 10k-RPM Raptors, and it’s worth noting that these drives have jumped in capacity from just 37GB at launch to 300GB with the latest VelociRaptor. What makes this nearly order-of-magnitude increase so impressive is the fact that the new VelociRaptor is a 2.5″ drive (although not thin enough to fit in most notebooks) while the rest of the Raptor family are of the much larger 3.5″ variety.
SSDs may be the new hotness, but storage capacity remains their greatest challenge. 256GB is a long, long way from 2TB. And that’s before considering the cost per gigabyte. Of course, solid-state drives have other strengths.
WorldBench uses scripting to step through a series of tasks in common Windows applications. Only a handful of these applications benefit from even the fastest storage solutions available today, so we’ve left out the individual results to focus on WorldBench’s overall score.
Although it doesn’t quite make up for their lesser capacities, the best of the latest crop of SSDs dominates the top end of our WorldBench rankings. Coincidentally, some earlier solid-state drives also register the lowest scores of the bunch. Clearly, flash-based storage has come a long way.
Conversely, the WorldBench performance of 3.5″ desktop drives has only improved by about 7%. The 10k-RPM Raptors don’t do much to set themselves apart, either. That’s perhaps to be expected, given that few WorldBench tests truly tax the storage subsystem.
We see a greater performance spread among the 2.5″ mobile drives, but remember we’re dealing with three spindle speeds here. The Fujitsu MHV2040AT is our lone 4,200-RPM model, with the rest split between 5,400 and 7,200 RPM.
Level load times
Our stopwatch testing for game level load times might seem a little crude, but it’s delivered consistent, repeatable results for years.
SSDs have the upper hand in our load time tests, with virtually all of them bunching at the front of the field. No doubt their near-instantaneous access times have much to do with their strong performance here. 2.5″ mechanical drives pull the same trick at the back of the pack, turning in slower load times than just about all of the 3.5″ models we tested. Western Digital’s Scorpio Black is a notable exception, though.
The 10k-RPM Raptors don’t offer much in the way of added performance here. Sure, they’re a little quicker than most of their 3.5″ counterparts, but not by the sort of margins one might expect given the difference in spindle speed. Then again, mechanical drive spindle speeds, which largely determine seek times, have been largely static for years.
File Copy Test
Our FC-Tests results normally span a total of 20 graphs. With 70 results each, that simply wouldn’t do, so we did some pruning. First, we dropped the partition copy results, which rarely deviated from those of the standard copy tests. We also consolidated individual results for each of FC-Test’s five test patterns into a single average file creation, read, and copy speed for each drive.
For the first time, our ridiculous four-way X25-E array makes its presence knownand in dramatic fashion, no less. The top five finishers here are all SSDs, although interestingly, none are based on multi-level cell flash memory. MLC flash can’t write as quickly as single-level cell flash or the DRAM memory used by the i-RAM and ANS-9010, slowing file creation performance dramatically.
As evidenced by the smattering of green across the spectrum, older and even some current SSDs still struggle to keep up with mechanical hard drives when it comes time to write batches of files.
In the land of spinning platters, the middle of the pack is awash in the golden glow of a mob of 3.5″ drives. File creation speeds have just about doubled here over the last four and a half years. They still have a long way to go to catch the X25-E, though.
Performance has more than doubled among the 2.5″ drives in the field, too, although the presence of a lone 4,200-RPM drive might magnify one’s impression of the changes. There’s still quite a gap between the fastest mechanical 2.5″ and 3.5″ drives, though.
The differences in transfer rates draw tighter when we switch to read tests. The X25-E RAID config isn’t as dominant, but SSDs as a whole fare much better here than in the file creation tests. Most are faster than the best mechanical drives around, Raptors included.
Once again the fastest Raptor is no faster than the quickest 3.5″ desktop drive in a sustained transfer rate test. That’s never been the Raptor’s strength, though. The higher areal density offered by 3.5″ desktop drives is usually enough to overcome their spindle speed disadvantage in sequential transfer rate tests.
The copy tests combine reading and writing, and based on the shape and color distribution of the graph, it’s the latter that influences performance more than the former.
The iPEAK Storage Performance Toolkit is itself more than a decade old, but I’ve yet to see its ability to record and play back I/O requests duplicated by a more recent application. Being able to record I/O activity in Windows allowed us to create a suite of disk-intensive multitasking workloads based on real tasks, such as creating compressed files, importing email into Outlook, and loading video into VirtualDub. The results below are an average of the mean service times of each drive across all nine of our custom workloads.
Even if you discount the exotic DRAM-based SSDs and the X25-E RAID array, there’s still quite a range of performance here. SSDs claim the top ten spots, easily edging out the fastest 3.5″ drives and the VelociRaptor. Of course, not all solid-state drives excel in our iPEAK tests. There’s plenty of green sprinkled through the rest of the pack.
The 2.5″ drives fare the worst here, dropping to the back of the pack where, really, you’d expect them to be. For the most part, notebook drives are much slower than desktop models. There are a few exceptions, however. A couple of Seagate’s 7,200-RPM Momentus models are a heck of a lot quicker than a good number of 3.5″ drives and even a few recent SSDs.
As we’ve seen throughout, our collection of desktop drives largely makes up the middle pf the pack. Most of the drives we’ve tested are of the 3.5″ variety, and we have older examples there than we do in the other categories.
IOMeter’s ability to hammer drives with up to 256 concurrent I/O requests nicely simulates demanding multi-user environments. We normally present IOMeter data for multiple load levels between 1 and 256 outstanding I/Os, but we have simply too much data to make such a graph viable here. Instead, we’re going to concentrate on peak IOMeter performance, which is simply the maximum transaction rate recorded during our scaling test.
The near-instantaneous seek times offered by solid-state storage solutions are a perfect match for the random nature of our IOMeter access patterns. Indeed, the SSDs are so overwhelmingly quick that our mechanical drive results are reduced to mere slivers on the graphs above. That says a lot about how dramatically solid-state drives can improve performance when faced with the right workloads. These are easily the most impressive performances we’ve seen to date.
You’ll notice that web server results look a little different from the others. Unlike the other access patterns, which are made up of a mix of read and write requests, the web server pattern is all readsperfect for avoiding the slower write speeds associated with MLC-based flash drives.
Obviously, we can’t make much of our mechanical hard drive scores with the SSDs hogging the limelight. Let’s set them aside for a moment and take a fresh look at how things stack up.
Score one for spindle speed. Wait, make that four. Higher rotational speeds enable lower access times, and that helps a lot in IOMeter, allowing the Raptors to surge to the front of the field. The VelociRaptor really is miles ahead of the next-closest mechanical drive tested. What’s more impressive, however, is how the older Raptors fare against today’s best 7,200-RPM desktop drives. Even the ancient WD360GD, which lacks the command queuing support that could make a big difference under this kind of load, offers higher transaction rates than the vast majority of modern desktop units.
IOMeter serves up all sorts of interesting subplots, such as the tendency of Western Digital drives to offer higher peak transaction rates than their competitors. One particularly shocking result: the brand-spanking-new Seagate Barracuda 7200.12 has a lower max transaction rate than five previous generations of Barracudas with three of four access patterns.
The latest Barracuda even gets beaten by a bunch of notebook drives, whose typically slower access times put them at a distinct disadvantage in IOMeter. Kudos to the Scorpio Black for proving that some notebook drives can take the heat, though. Whatever secret sauce Western Digital has mixed into its 3.5″ formula appears to work for 2.5″ drives, too.
HD Tach falls firmly within the realm of synthetic benchmarks, and its transfer rate tests nicely illustrate the peak sustained throughput of each drive.
SSDs come out on top again, and by substantial margins even if you ignore the outlandish RAID and DRAM-based solutions. Many of today’s solid-state drives can sustain read rates that are double those of the fastest available mechanical drives on the market.
The sequential transfer rates of mechanical hard drives have improved steadily over the years, but the jump from one drive to the next has been incremental. SSD transfer rates have risen at a considerably more rapid pace.
It’s interesting to see the Raptors spread throughout the field here. Traditionally, these 10k-RPM drives have been the fastest of their times. However, in peak transfer rate tests like these, substantially higher areal densities allow 7,200-RPM desktop drives to be more competitive than one might expect given their slower spindle speeds.
Speaking of slower spindle speeds, our collection of 2.5″ drives has a tough time keeping up. Mechanical hard drives have the highest sequential transfer rates on the outer edges of their platters, and 2.5″ discs have a lot less outer edge area than 3.5″ ones. Still, a smattering of blue has crept up into the middle of the pack.
Over the last few years, both 2.5″ and 3.5″ mechanical hard drives have managed to double their sustained transfer rates at the same spindle speed. Rising areal densities have helped on this front, but so has the fact that drive makers are packing more platters into each drive. More platters means more outer edge area to exploit.
Our RAID configs have an unfair advantage here by virtue of their use of multiple Serial ATA connections. Otherwise, performance in HD Tach’s burst speed test appears to be gated by the practical throughput of the 300MB/s SATA interface. 262MB/s is as fast as we’ve seen from a single device, with most managing over 200MB/s.
There’s a second tier of sorts starting at around 137MB/s. Here we find drives designed for the older 150MB/s Serial ATA specification. A bunch of our 2.5″ drives are also IDE models with interfaces capped at 100MB/s.
Perhaps better than any other set of results in this retrospective, our collection of HD Tach random access times clearly illustrates the seek time advantage that solid-state drives have over their mechanical counterparts. Raptors aside, even the quickest mechanical drives are a full two orders of magnitude slower than the SSDs here.
SSDs don’t have to deal with the mechanical or rotational latency inherent to traditional hard drive designs. They don’t have to move a physical drive head, and they don’t have to wait for data points to come spinning ’round on a platter. Increasing spindle speeds is one way to lessen the impact of latency, and as the 10k-RPM Raptors illustrate, doing so can greatly improve random access times.
With only the Raptors exceeding 7,200 RPM, we haven’t seen random access times fall all that much for standard desktop drives. In fact, the old Barracuda 7200.7 and Deskstar 7K500 still have among the quickest access times we’ve measured for mechanical drives. The higher precision required to quickly seek out data on drive platters with ever-increasing areal densities is a challenge for newer models. Some have risen to it, while others have stumbled. The new Barracuda 7200.12’s 17-millisecond access time is particularly disappointing
The access times of mobile hard drives are predictably higher than those of their desktop counterparts. The slower spindle speeds of the 5,400 and 4,200-RPM drives don’t help, but even the 7,200-RPM models lag a few milliseconds behind their 3.5″ brethren.
A great many things have changed in the storage world in the last four and a half years. We’ve mainly focused on performance today, but I suspect the rapid rise of mechanical hard drive capacity has had the greatest impact on users. SSDs remain an expensive proposition even for enthusiasts, while just about anyone can afford a sub-$100 terabyte desktop drive or 500GB mobile unit. The fact that it’s taken less than five years for hard drive capacities to grow by a factor of eight is a testament to just how effective perpendicular recording technology has been at increasing areal densities.
Although the capacities of mechanical hard drives have snowballed, the performance gains have been considerably more modest. Higher areal densities have improved transfer rates, both in real-world file operations and in synthetic tests, but we’re looking at roughly a two-fold increase in performance, at best. Mechanical and rotational latencies are even tougher problems to solve, and we’ve seen little progress on this front. The Raptors have enjoyed success at 10k RPM, but the vast majority of desktop drives have remained at 7,200 RPM. Some, like Western Digital’s Caviar Green series, have even lowered spindle speeds in a bid to cut power consumption and noise levels.
Lower spindle speeds are common among 2.5″ notebook hard drives, but the latest 7,200-RPM models give older desktop units a run for their money. Even if you don’t want to shell out for an SSD, swapping in a fast 7,200-RPM hard drive is usually the best way to improve notebook performance, particularly with budget models that are typically equipped with older 5,400-RPM drivesor, shudder, a 4,200-RPM clunker.
SSDs pose perhaps the biggest threat to notebook hard drives, in part because they share the same 2.5″ form factor. SSDs tend to have higher shock tolerance and lower power consumption, too, and the highest capacity notebook hard drives top out at only half a terabyte.
If you look at our performance results, though, today’s SSDs seem best suited to the sort of random access patterns typical of multitasking and multi-user environments, such as high-end workstations and servers of various types. Solid-state drives may be expensive, but if you consider their performance per dollar, some offer phenomenal value for enterprise applications. Then again, we’ve not tested the 15k-RPM SAS and SCSI drives against which SSDs must to compete in those markets.
SSDs have long had a substantial seek time advantage over their mechanical counterparts. Only recently has silicon-based storage also enjoyed faster sequential transfer rates. Solid-state drives are still relatively new, with lots of room for further performance growth. Mechanical hard drive technology is more mature, and while capacities continue their rapid ascent, the performance improvements seem stunted by comparison.
As solid-state drive prices continue to fall and mechanical hard drive capacities continue to climb, I suspect we’ll see the two technologies coexist in most enthusiast desktop systems. Even today, there’s a good case to be made for running an SSD as an operating system and applications drive backed by a terabyte or two of mechanical storage. That might not be the dominant configuration a few years from now, but based on how rapidly SSDs are evolving, it’ll probably be the best one.