Hacker News new | ask | show | jobs
by iigs 5728 days ago
Where I work we presently have between 5-10 TB (actual, not raw, not counting replicants) across several MySQL shards. $/iop they just slay magnetic solutions.

The most interesting insight I can offer is that because there is no seek time penalty and SSD throughput increases as the I/O scatters, our tests with RAID 5 are essentially as fast as RAID 10, which makes $/gb much better than it looks at first blush.

TRIM through hardware RAID is still an unsolved problem for us, as is getting wear life SMART data (Intel X25M G2) through the RAID controller. My gut feeling is that the final solution is probably a JBOD controller and Linux software RAID.
1 comments
macros 5728 days ago
We ended up with the JBOD solutions on a standard ahci controller with soft raid. We've tried a few different controllers and so far the performance with them has been been kinda sad, they just aren't made for the amount of iops the ssds provide.

I saved the results from tiobench from some of my initial testing with 6 intel x-25m here: http://gist.github.com/348417

Not bad for $2400 worth of ssds.

link
Chris_Morrell 5728 days ago
You may want to consider testing the LSI 9211-8i for a very fast JBOD controller or the 9260-8i for an equally fast RAID5/6 card. I've been very happy with these controllers and they are typically much cheaper than the equivalents from ARECA. Just make sure to use the latest drivers and firmware as LSI has been improving SSD performance by leaps and bounds in the last year.

On thing I noticed with the X-25M G2s is that if you are in a write intensive environment they will degrade in performance dramatically. The SandForce SF1200/1500 drives seems to handle garbage collection much better and run with lower write amplification levels. I've found that the SandForce drives do a much better job at maintaining acceptable random write performance after months of being beat on. That all being said a beat to hell X25-M is still 10 times better than a spinning disk.

If you want I can run an iometer batch or figure out tiobench if you'd like some runs on the 9211 or 9260 to compare with your ICH and Adaptec runs. Ping me at gomeler at gmail if you are interested.

link