ZFS performance as the FS fills up?

[Patrick M. Hausen]

Hello,


Am 09.03.2011 um 14:28 schrieb Tom Evans:
> I don't think this is accurate. Prefetch was certainly disabled by
> default on 8.0 if you had 4GB of RAM or less, requiring the sysctl
> vfs.zfs.prefetch_disable=0 to be set if you wanted prefetch and had
> 4GB of RAM or less.

My configuration is this:

JBOD with 24 disks before the performance increase, now 30.
5 vdevs of 6 disks each, raidz2. 16 GB RAM, 8.2 amd64.

vm.kmem_size="12288M"
vfs.zfs.arc_max="8192M"
vfs.zfs.prefetch_disable="1"
vfs.zfs.txg.timeout="5"

-------

net.inet.tcp.sendbuf_max=16777216
net.inet.tcp.recvbuf_max=16777216

net.inet.tcp.sendspace=65536
net.inet.tcp.recvspace=131072

kern.maxvnodes=250000
vfs.zfs.txg.write_limit_override=1073741824


After I added the last 6 disks net write performance (that means
how quick Amanda fills the current vtape) at least doubled.
Without any further Amanda tuning.

I am aware that there are many components involved that
affect performance, so what would be the best way to measure the
ZFS part? Working crystal balls have become rare these days ;-)


Kind regards,
Patrick
-- 
punkt.de GmbH * Kaiserallee 13a * 76133 Karlsruhe
Tel. 0721 9109 0 * Fax 0721 9109 100
info at punkt.de       http://www.punkt.de
Gf: Jürgen Egeling      AG Mannheim 108285