I'm going to try to reply to all the responses I got together.
> Have you tried backing up other hosts on your network? What are
> the speeds with these hosts? I've noticed that different host
> respond with varying speeds despite being on the same network.
> Wondering if this has to do the client OS doing some throttling
> based on work load.
I am backing up the Bacula server itself, my workstation (with is a FreeBSD box
as well), and our main file server, which is a SunOS server. We aren't doing
any throttling intentionally, but I also see a large variation in throughput
depending on the client in question, but none of them - not even the local
server backing up itself - are all that impressive right now.
> I would start by turning off software compression and do
> performance tests with full backups. A second thing to try is to
> enable attribute spooling so the database does not slow down the
> backup. This can be useful if you have millions of files.
We do not have software compression enabled, as far as I can tell. I've turned
on the "Spool Attributes" option in my job definition, and we'll see if that
helps.
> Compare against a stock, non tuned, Bacula install. Are you
> going between building where you get the slow transfer speed?
> UCSC has 1 Gb links between buildings from my recollection. The
> link to the outside world is not much more than that. Bacula
> also has a batch mode which you can twiddle around with.
For the slowest backup job, the two servers are sitting in the same rack on the
same gigabit switch. The fastest client actually is in a different building.
Yes, we have 1Gb between buildings here, but out Internet connection was
recently upgraded to 10Gb (not that it really applies to this situation anyhow).
I found some Google hits that talked about batch mode, but no documentation
that tells me how to enable it. Can you provide a link?
> Is the MySQL database storage on the same RAID array you are
> writing backups to?
Yes and no. Currently, in our "dev" environment, they are both on the same
physical RAID array, but Bacula operates in a separate jail from mySQL. When
we move to production, the director will probably run on one server and the
storage daemon on another, so maybe that will help?
> It may be useful to run iftop on the network interfaces of the
> Bacula server to see what the network IO is like, and then compare
> that to iotop to see what the disk IO is like.
We actually run Cacti against all our servers. Disk throughput for the Bacula
server can hit as much as 240Mb/s during a backup, whereas the network
throughput at the same time is around 80Mb/s, with a few spikes to 96Mb/s. For
what it's worth, iperf can hit about 780Mb/s between these hosts.
I just twiddled some ZFS parameters last night (turning off the primary and
seconday caches) and reconfigured the zpool to let ZFS handle the striping
(rather than the Adaptec controller handling the RAID array), so we'll see what
numbers we come back with tomorrow. I've also added some other different
hardware/OS combination clients to see if we can work out a pattern.
Tim Gustafson
Baskin School of Engineering
UC Santa Cruz
tjg AT soe.ucsc DOT edu
831-459-5354
------------------------------------------------------------------------------
Beautiful is writing same markup. Internet Explorer 9 supports
standards for HTML5, CSS3, SVG 1.1, ECMAScript5, and DOM L2 & L3.
Spend less time writing and rewriting code and more time creating great
experiences on the web. Be a part of the beta today.
http://p.sf.net/sfu/beautyoftheweb
_______________________________________________
Bacula-users mailing list
Bacula-users AT lists.sourceforge DOT net
https://lists.sourceforge.net/lists/listinfo/bacula-users
|
