PCMark 7: Storage Suite
PCMark 7 is the latest synthetic from Futuremark. The older Vantage test gives a moderate approximation of storage performance, but there are situations where the scores vary too much from one run to another to make it a great tool for evaluating SSDs. Both metrics use the same underlying trace-based technology from Intel's IPEAK. However, Futuremark's programmers made a few changes in the latest version that it claims improve its accuracy, though it's still hard to make sense of this benchmark's results. Fortunately, at least the variability is gone.
The difference between the latest SSDs from Intel, Crucial, OCZ, Mushkin, Corsair, Patriot, and Adata all fall within a tight spectrum in PCMark 7, but the overall storage score is based on a geometric mean of the subtests. That's why it's important to take a deeper look into the individual tests themselves.
The Windows Defender test is based on a trace of Windows' Defender utility performing a Quick Scan of the system. This is very read-heavy scenario comprised of lots of random access. In fact, read operations make up 97.9% of the trace. And of the read operations, 93% are random accesses. Naturally, anti-virus scanning and file searches are the two more applicable real-world scenarios for a test like this.
Overall, there's really no difference between the SSDs. The 120 GB Vertex 3 and S511 perform marginally better than the other new SandForce-based drives. However, compared to a desktop hard drive, the crowd of SSDs offers about 2x the performance.
The second test in PCMark 7's storage suite is a trace based on importing 68 images (434 MB total) from a USB thumb drive into Windows Live Photo Gallery. This doesn't actually include copying the image files. The trace only includes the I/O activity pertaining to indexing. This type of scenario involves writing more data than you read, and most of the writes are random.
This is another scenario where we see a difference between the SF-2200-based drives employing synchronous and asynchronous NAND. The gap varies from 15% to 30%, depending on the drive.
The Video Editing test is based on the I/O activity of publishing a 1080p video in Windows Live Movie Maker. We're dealing specifically with a scenario where multiple high-def sources are combined and written to a single output file. Overall, the split is 30/70 between random and sequential reads, and reads are more prevalent than writes. In fact, the ratio is about 1:9 in favor of reads.
The problem here is the writing portion of the workload, which consists almost completely of incompressible video data. This puts SandForce-based SSDs at a disadvantage.
The Windows Media Center test is a based on a trace of an HTPC recording two simultaneous TV shows in Windows Media Center, while playing a separate prerecorded show. We're basically reading one file and writing two others.
This type of scenario involves a lot of random writes (94%), because Windows Media Center incrementally adds data to the video file as the TV show progresses. Reads are another story, as they're almost all sequential (84%). Playing a video file is different than recording one; when you play a video file, you're loading it up and playing it back as a continuous stream.
In this scenario, there's practically very little difference between SSDs.
The Adding Music test in PCMark 7 is a not exactly what it sounds like. Futuremark hooked a drive filled with 68 GiB of music files (lossless WMA) to a computer and recorded the I/O activity while Windows Media Player added the audio tracks to the music library. The important point is that this doesn't actually involve copying files to the disk. Rather, it's all about scanning and indexing music files. You'd think that means more random reads and almost no writes, but indexing involves adding to a database of information. That's why we're dealing with more sequential writes (75% of all writes) and a situation where reads are outnumbered by writes 2:1.
In this benchmark, we're presumably restricted by the low transfer rate of the external disk with all of the music files.
The Starting Application trace is extremely brief in that it's only made up of loading the PCMark 7 Whitepaper v1.0 PDF and opening Internet Explorer from the taskbar (19.236 seconds). That adds up to reading a 717 KB PDF file and loading executables, along with related file dependencies from the system drive. The amount of data read outnumbers the amount written 63:1, and most of the read accesses are random in nature (86%).
There's a clear delineation here between the drives complemented by synchronous NAND and the ones that use asynchronous memory to help cut costs.
The Gaming test involves starting and loading World of Warcaft, which is why we're dealing almost exclusively with reads. Most of the read operations are random in nature, but in terms of the total amount of data read, there's a fairly even split between sequential and random accesses. Even though there are 3002 random reads and 575 sequential reads with block sizes up to 4 KB, this cumulatively only accounts for less than 14 MB of the total 123 MB read. At block sizes between 1 and 2 MB, there are more sequential reads than there are random reads.
The margin of difference is much smaller than the Starting Applications test, so it's difficult to make any comment on performance. Of course, there's still a clear line in the sand between SSD and HDD performance.
