Re: SSD performance

On 1/23/09 3:35 AM, "david(at)lang(dot)hm" <david(at)lang(dot)hm> wrote:
http://techreport.com/articles.x/15931/1

one thing that both of these reviews show is that if you are doing a
significant amount of writing the X-25M is no better than a normal hard
drive (and much of the time in the middle to bottom of the pack compared
to normal hard drives)

David Lang

The X-25-M may not have write STR rates that high compared to normal disks, but for write latency, it is FAR superior to a normal disk, and for random writes will demolish most small and medium sized raid arrays by itself. It will push 30MB to 60MB /sec of random 8k writes, or ~2000 to 12000 8k fsyncs/sec. The -E is definitely a lot better, but the -M can get you pretty far.

For any postgres installation where you don't expect to write to a WAL log at more than 30MB/sec (the vast majority), it is good enough to use (mirrored) as a WAL device, without a battery back up, with very good performance. A normal disk cannot do that.

Also, it can be used very well for the OS swap, and some other temp space to prevent swap storms from severely impacting the system.

For anyone worried about the X 25-M's ability to withstand lots of write cycles ... Calculate how long it would take you to write 800TB to the drive at a typical rate. For most use cases that's going to be > 5 years. For the 160GB version, it will take 2x as much data and time to wear it down.

Samsung, SanDisk, Toshiba, Micron, and several others are expected to have low random write latency, next gen SSD's this year. A few of these are claiming > 150MB/sec for the writes, even for MLC based drives.

A RAM based device is intriguing, but an ordinary SSD will be enough to make most Postgres databases CPU bound, and with those there is no concern about data loss on power failure. The Intel X 25 series does not even use the RAM on it for write cache! (it uses some SRAM on the controller chip for that, and its fsync safe) The RAM is working memory for the controller chip to cache the LBA to Physical flash block mappings and other data needed for the wear leveling, contrary to what many reviews may claim.