6

I have a 10-disk RAIDZ in Solaris 11.1. Its connected by 8G fibre. I have created a ZVOL, made it available with COMSTAR and mounted it on another machine connected by fibre. When I benchmark it with large files, I get ~700MB/s read and ~150MB/s write. That's averaged. Upon closer inspection the read speed is totally steady but the write speed is constantly fluctuating from all the way down from 50MB/s to 400MB/s.

I know ZVOLs use sync writes but should I get this slow write speed and be seeing this fluctuation? I feel like I should get better (or at least consistent) write without having to have a ZIL.

BruceLeroy
  • 176
  • 4
  • 2
    Whether or not the ZVOL is actually doing sync writes depends on a number of settings. First, show us the 'sync' setting on the dataset, the writeback cache ('wce' or sometimes 'wcd') setting on the COMSTAR LU, and what if any sync setting is being set by the client? Usually when I see spiky write performance, it is because you're sending too much. ZFS write throttling is a hard nut to crack, lots of moving pieces and tuning to touch. To be frank, it is fairly stupid at times - one of those times is if you slam the box with more than it can handle with no end in sight. – Nex7 Jul 24 '13 at 07:20
  • 1
    +1 to Nex7's comment, and 150MB/s average sounds pretty good for RAIDZ. I know it's not your ideal answer, but a ZIL on SSD would probably smooth things out and speed things up, depending upon how you implement it. – s.co.tt Jul 30 '13 at 18:42

1 Answers1

2

RaidZ has write performance of a single disk. It is not directly equivalent to raid5. If you want performance you need to stripe it or use mirrors.

JamesRyan
  • 8,138
  • 2
  • 24
  • 36