Foo wrote:
> Thanks, that helped, although there is still some packetloss (about a
> quarter of the previous value).
Try setting the high priority job to -c1 (realtime) and the bacula fd to
-c3 (idle) priority.
If that's still not sufficient, you may need to tune your disk subsystem
for low latency instead of the usual high-throughput-and-high-latency
configuration. Things that may help include:
- Reducing disk level, block device level, and filesystem-level
readaheads (but not TOO much otherwise you'll waste so much time seeking
that you'll be unable to service all the required I/O);
- Trying different I/O schedulers. cfq is the best for many, but not
all, workloads.
- If you have a hardware RAID device:
- If and only if it has a battery backed cache, enable write back
caching mode
- Tune (probably turn down) the hardware and/or driver I/O queue
depth, so fewer requests are waiting outside the control of
the Linux I/O priority queue system.
I landed up having to hack the sources of the driver for my 3Ware
Escalade 8500-8 to get decent I/O latency, since the driver queued up to
256 (!!) requests internally. This prevented Linux from pre-empting
already issued I/O with higher priority new I/O requests.
--
Craig Ringer
------------------------------------------------------------------------------
Open Source Business Conference (OSBC), March 24-25, 2009, San Francisco, CA
-OSBC tackles the biggest issue in open source: Open Sourcing the Enterprise
-Strategies to boost innovation and cut costs with open source participation
-Receive a $600 discount off the registration fee with the source code: SFAD
http://p.sf.net/sfu/XcvMzF8H
_______________________________________________
Bacula-users mailing list
Bacula-users AT lists.sourceforge DOT net
https://lists.sourceforge.net/lists/listinfo/bacula-users
|