There is a new article at PC Perspective that talks about the really hot Intel SSDs. But it goes beyond a normal review and gets into some of the guts of SSDs and talks about an important topic: file system performance as it ages. So, I want to write about the review and point out the file system aging problems (which I write about later).
I think most everyone has seen that the new Intel X25-M is the new high-performing SSD in the world. While it is the "desktop" version of the X25 (the X25-E is the enterprise version) the review discovers some very interesting things about the drive. In particular, it talks about the cool things Intel has done to crank up the performance of SSDs.The article explains that Intel has used a custom-built 10-channel flash memory controller that can take advantage of Native Command Queuing (NCQ). NCQ allows the controller to reduce seek time, and Intel has used it to prefetch data, which reduces latency in the drives. One problem that plagues SSDs is the problem of write performance. In a previous blog I wrote about poor write performance of SSDs. This problem is particularly noticeable for small writes. The Intel controller has the ability to combine writes and put them into a single flash block operation. I think this ability is one of the reasons why the benchmarks on the drive are so good.The article is particularly good in reviewing the drive in depth (in my opinion many of the reviews you see floating around the Web are really superficial, but that's another blog). One extremely interesting thing that the reviewer points out is that he found as the drive "aged"— that is, the drive was used for some period of time with files of various sizes—the performance tanked. Figure 1 is from the article (Allyn Malventano at www.pcper.com and used with permission).
The article goes on to talk about reasons why this happened. One in particular discusses the effect of combining writes to fill a single flash block. They argue that this has a definite impact on performance because it basically leads to fragmentation of the drive (write combining may even force more fragmentation than "normal," but that point has yet to be tested). In fact, their testing showed that the drive performance can drop below the manufacturer's specification. Intel gave them some options for "fixing" this problem including using IOMeter to write to the entire drive (so you can't use the drive as your OS drive). The authors tried using a similar approach by writing a very large file to the drive to force it to adapt the fragmentation, but they did not have as much success with this approach.The second option from Intel was to use a tool that performed a SECURE ERASE command on the drive. But you have to use a specific application that can issue this command to the drive.
One aspect of the article that many people will miss is that as the drive was used it became fragmented and could not recover performance without some substantial work on the drive (i.e., writing over the data). This point reminded me of an aspect of HPCC storage that people don't really think about—the effect of aging on performance. I'll be writing a blog on this in the near future so stay tuned. :)Jeff