mrb's blog

Disk Vibrations and SSDs

Keywords: hardware performance ssd

Optimizing storage is a big part of my work on rainbow tables (that I have yet to publish). I like to read Robin Harris (from StorageMojo, and Storage Bits) who is a prominent blogger, a true tech geek, and a storage specialist. In Bad, bad, bad vibrations Robin comments on a paper that sheds more data on the problem of vibrations degrading HDD performance. I am myself familiar with this problem. I once helped troubleshoot an otherwise idle Linux server that was rebuilding a 4-drive MD raid5 array at only 5-10 MB/s. "iostat -x" showed the disks were taking 50-100ms to serve ~1024kB sized I/O requests when it should normally take less than 5ms to serve them. After having spent some time on it, unsuccessful, I put my 2 hands on the 1U chassis thinking "what the hell is the problem", and when doing this I felt vibrations that were stronger than normal, which surprised me. I suddenly remembered the shouting in the datacenter video from Sun engineers demonstrating the problems vibrations could cause. And sure enough, I removed the drive trays and noticed that most screws attaching the disks were loose... Tightening them up immediately reduced vibrations and allowed the resync I/O throughput to raise to a respectable 100+ MB/s. Victory!

Going back to Robin's comments about disk vibrations, one of the points he draws in his conclusion is that "disks can improve their [random I/O] performance by 50% through vibration damping, that changes [the] value proposition for SSDs". I disagree with this statement. This does not change anything at all. Modern SSDs provide random IOPS performance that is on the order of 10-100x better than HDDs. 50% is negligible compared to a 10-100x factor. Here is a small table I have compiled representing IOPS numbers for a typical 7200RPM HDD and for an SSD model that seems pretty popular these days (the MLC-based second generation Intel X25-M series):

Device 4kB random read IOPS 4kB random write IOPS
Typical 7200RPM HDD ~100 ~100
Intel G2 MLC 80GB (X25-M) 35000 6600
Intel G2 MLC 160GB (X25-M) 35000 8600

Notice how these SSDs provide 350x the read IOPS performance of a typical 7200RPM HDD. And 66x-86x the write IOPS performance. Even assuming that dampening vibrations can increase random IOPS performance of a regular HDD by 50% or more, in this specific example it would reduce the advantage of SSDs to "only" 175x better read IOPS and 33-43x for write IOPS. As I said, this does not change the value proposition for SSDs. They remain the kings of random IOPS by far.