Following this article about the X25-E SSD and performance when the SSD's write-cache turned off, I made my own observations.
Setup:
- Adaptec 5805 + 512MB battery-backed cache
- 4x X25-E 64GB (SSDSA2SH064G1GC) in RAID10, 256KB stripe)
With an active MySQL server with InnoDB tables, I selectively set to write-back/write-through the cache of the the controller (arcconf setcache logicaldrive wt/wbb) and drives ((arcconf setcache device wt/wb) and observed iostat over a 15-second interval.
| Controller | Disks | rrqm/s | wrqm/s | r/s | w/s | rKB/s | wkB/s | avgrq-sz | avgqu-sz | await | svctm | %util |
| WB | WT | 0.00 | 70.67 | 2.53 | 1158.73 | 40.80 | 5434.13 | 9.43 | 0.08 | 0.07 | 0.07 | 7.71 |
| WB | WB | 0.00 | 62.93 | 9.40 | 1107.67 | 172.80 | 4976.27 | 9.22 | 0.08 | 0.07 | 0.07 | 7.64 |
| WT | WB | 0.00 | 62.53 | 4.53 | 1126.33 | 72.53 | 6157.07 | 11.02 | 0.29 | 0.26 | 0.24 | 27.46 |
| WT | WT | 0.00 | 63.53 | 2.60 | 1169.00 | 42.13 | 5757.07 | 9.90 | 0.83 | 0.71 | 0.68 | 80.01 |
Fig. 1 - iostat -k 15 -x across four different cache settings
Performance clearly suffers with write cache turned off, with a 3x increase in (write) latency.
What's interesting though is that the controller's cache resolves this problem and improves latency by another 3x over the SSD's write-back performance.
So a RAID controller is a good complement to SSDs when data integrity is important, improving performance by 10x in our case. With the associated risk of the controller fubar'ing and wiping the cache of course :)
Post new comment