TR DriveBench 2.0 — More disk-intensive multitasking
As much as we like DriveBench 1.0's individual workloads, the traces cover only slices of disk activity. Because we fire the recorded I/Os at the disks as fast as possible, the drives also have no downtime during which to engage background garbage collection or other optimization algorithms. DriveBench 2.0 addresses both of those issues with a much larger trace and a slightly different testing methodology.
For our new trace, I recorded about two weeks of disk activity on a test system pressed into service as my primary desktop. The system was left on at all times, and it was used mostly for web surfing, email, photo editing, gaming, and working with the HTML, Excel, and image files that become TR content. To make things more interesting, I fired up disk-intensive multitasking workloads alongside those more mundane desktop tasks. The multitasking workloads were similar to what's included in DriveBench 1.0: compiling code, copying files, downloading torrents, transcoding video, and scanning for viruses.
Although my bursts of disk-intensive multitasking were contrived in nature, the goal was to come up with a demanding test that would probe drives for weakness in a sea of everyday I/O. The system was limited to a single partition, which housed not only the OS and applications but also all of the associated data and downloaded, ahem, Linux ISOs. We plan to add at least one more DriveBench workload that more strictly models the life of an OS and applications drive, but that'll have to wait for a future article.
As it stands, our multitasking-infused trace is loaded with more than 25 million read operations totaling over 1.1TB of data. The workload has plenty of writes, too: 14 million that add up to nearly 525GB. That's a busy couple of weeks.
DriveBench 1.0 all but eliminates disk idling, but this second revision gives the SSDs plenty of idle time for background processing. The test begins with drives fresh from a secure erase, but it writes across their full capacity to ensure that all flash pages are filled before performance is measured. Things are a little different on that front, too. Instead of looking at a raw IOps rate, we're going to explore service times—the amount of time that it takes drives to complete an I/O request. We'll start with an overall mean service time before slicing and dicing the results.
The SandForce drives reign supreme, with the synchronous Force GT, Vertex 3, and HyperX configs topping their asynchronous counterparts. Intel lurks just shy of that second pack of SandForce offerings, while the Crucial m4 trails even the 320 Series. Behind it, the Performance 3 shows serious weakness and is nearly as slow as our mechanical hard drive.
Let's try to make some sense of these numbers with a breakdown of reads and writes.
The standings shift slightly from the overall results when we only consider reads. While the synchronous SandForce drives remain in the lead, they're now followed by the 510 Series and the m4. Corsair's Performance 3 Series fares much better here, trailing the 320 Series by a relatively small margin.
Switch to writes, however, and the middle of the pack shuffles completely. The asynchronous SandForce configs tuck in behind their synchronous relatives, while the 510 Series slips behind not only the 320, but also the Caviar Black. Write service times are even slower on the Crucial m4, and the Performance 3 is a complete disaster.
There are millions of I/O requests in this trace, so we can't easily graph service times to look at the variance. However, our analysis tools do report the standard deviation, which can give us a sense of how much service times vary from the mean.
With reads, the standard deviation results stack up much like the mean service times. One notable exception is the Crucial m4, which falls three places.
The SandForce configs have a low standard deviation with both reads and writes, indicating more consistent performance than the competition. Coupled with lower mean service times, that's a pretty appealing combo. Once again, the Marvell-based SSDs falter with writes and fall behind our lone mechanical hard drive.
If I haven't already scared you off with too many graphs and statistics, this next pair will do it. We're going to close out our DriveBench analysis with a look at the distribution of service times. I've split the tally between I/O requests that complete in 0-1 milliseconds, 1-100 ms, and those that take longer than 100 ms to complete.
The top contenders are pretty closely matched with reads. Only the Agility 3, Force 3, and 320 Series drop off the lead group. Our write results are more interesting, and they hint at why the Performance 3 fares so poorly overall: 3.6% of all write request take over 100 milliseconds to complete. That's a long time within the context of a modern PC. We're currently looking at other ways to crunch these data to determine whether those slower service times occur in bunches that might cause perceptible stuttering or lag.
Even if they don't, the percentage of 100+ ms service times on the Performance 3 is several orders of magnitude higher than on most of the other drives—including the Caviar Black. The m4 and 510 Series also have a higher number of 100+ ms requests than the mechanical drive, highlighting a potential weakness of the Marvell controller.
|The SSD Endurance Experiment: Only two remain after 1.5PB||32|
|Friday night topic: Conspiracy theories||31|
|GeForce 344.11 WHQL drivers support new cards, new games, G-Sync||3|
|Deal of the week: A 23'' IPS monitor for $150, a 200-mm fan for free, and more||22|
|GeForce GTX 970, 980 cards already widely available||14|
|Curved VA panel powers 27'' Samsung monitor||15|
|Android L to encrypt devices by default||6|
|Nvidia's GeForce GTX 980 and 970 graphics cards reviewed||283|
|Thursday Night Shortbread||17|