[OpenIndiana-discuss] AMCC/3Ware 9550SX-8LP

[Sašo Kiselkov]

On 01/25/2013 03:11 PM, Edward Ned Harvey (openindiana) wrote:
>> From: Christopher Chan [mailto:christopher.chan at bradbury.edu.hk]
>>
>> :-D I'm here to entertain since I have not been able to spring for a ssd
>> for use as a slog. :-D
> 
> LOL, you mean you have a HDD slog device?   :-D
> It's actually very surprising how well that works, especially if you have a high rpm drive.  Because most of the wasted access time in a HDD is waiting for head seek time.  Usually, the rotational latency is like 1ms or so, so it's irrelevant compared to the head seek, but if you do a good job of eliminating the head seek time, then the rotational latency becomes completely relevant, or even dominant.  If your drive is used only for slog, guess what ...  ZFS does a pretty good job of keeping the ZIL clustered together in tightly grouped tracks, so you've done a pretty good job of eliminating the seek.    ;-)
> 
> If you let the ZIL sit on main pool, it will both be adversely affected by other reads and writes ...  And it will also adversely affect other reads and writes, mutually.  And it's more multiplicative rather than additive.  Because you now have large random seek times, and cache flushes, so all your IO optimization techniques get messed up.

Agreed, short-stroking a 15k HDD to something like 1/100th of its
capacity actually gives you really good transactional throughput.
Considering a 15k 300GB SAS drive is comparatively cheap nowadays, this
gives you 3GB of slog (more than enough) at fairly low latency (1000
iops is achievable).

--
Saso