The nightmare comes slowly... ZFS prefetch algorithm can cause performance drawbacks

| | TrackBacks (0)
Just a side note:

We are using ZFS across two SAN mirror pairs for our production mailserver. It is running Cyrus IMAP with 14000 users and 50000 mailboxes. The zpool has 8 Terabytes and the system experienced some latencies when accessing data on the zpools.

The first problem was easy to resolve
but still we had IMAP SELECT times greater 0,5 seconds on some accesses.

The machine was doing approx 2000 random read accesses per second (we have approx 1000 concurrent IMAP users including our webmail application which is heavily using IMAP commands).

The first step was to put more RAM for the ARC cache in the machine, which resulted in 100-200 random read accesses per second. Cyrus metadata files of logged in users are now mainly in RAM.

The second step was to disable the prefetching algorithm. I put that in /etc/system:

set zfs:zfs_prefetch_disable=1

And - hooray - SELECT times are under 0,1 seconds. The prefetch method seems not to be optimized for a highly random i/o environment.

0 TrackBacks

Listed below are links to blogs that reference this entry: The nightmare comes slowly... ZFS prefetch algorithm can cause performance drawbacks.

TrackBack URL for this entry: http://southbrain.com/mt/mt-tb.cgi/21

December 2015

Sun Mon Tue Wed Thu Fri Sat
    1 2 3 4 5
6 7 8 9 10 11 12
13 14 15 16 17 18 19
20 21 22 23 24 25 26
27 28 29 30 31    

About

This blog is owned by:

Pascal Gienger
J├Ągerstrasse 77
8406 Winterthur
Switzerland


Google+: Profile
YouTube Channel: pascalgienger