Bob Hetzel <beh AT case DOT edu> writes:
>> What kind of backup throughput are people seeing?
>>
>> I'm backing up Linux clients to a Linux server across a Gb network to
>> local disk (hardware RAID 5) on the backup server and rarely see
>> speeds above 8MB/s. What kinds of performance tweaks are available?
>> Is it all about tuning IO on the client or can Bacula be tweaked?
>
> 1) Writes to a RAID-5 array slow down the more drives you have. For
> most things like use as a file share performance is still great so
> long as you've got the writes cached, but for sustained writing the
> cache will fill up before the writes complete and you wind up being
> at the mercy of the platter speeds. For a write to a RAID-5 set of
> 5 drives, it has to do reads as well as writes on all the drives in
> the set.
this is too simplified, IMHO. for streaming writes, RAID-5 will write
complete stripes (typically 64 KiB * number of disks), and you get
performance linear with the number of spindles. however, if you use
RAID-5 for spool data, Bacula will be writing to many different files
simultaneously (one for each concurrent client), and you will be more
likely to see performance detoriate due to seeking and smaller I/O
sizes. still, it's rather unlikely to be a problem when you have
hardware RAID 5 with cache in battery backed RAM, IMO. software RAID
5 is a completely different story, though, since the kernel can't
delay writes for very long without risking dataloss.
> That's why some people do RAID-1 and RAID-10 for projects like high
> performance backup to disk, this but of course that's more
> expensive.
it may be worthwhile to look into using an SSD disk for the spool
directory.
> 2) The more simultaneous backups you have going at the same time,
> the slower it will be too.
>
> Depending on your config... that might be all you could expect. The
> rotational speed and interface of the drives matters as well as the
> speed of the computers being backed up--desktops with a single sata
> 7,200 rpm sata drive won't be very quick. That's where concurrency
> helps, but as I mentioned above, if you get too many writers at the
> same time you'll reduce performance of all of them.
if you do spooling, this will be less of an issue. it is of course
essential that the spool and data areas don't share spindles.
--
regards,
Kjetil T. Homme
Linpro AS
-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/
_______________________________________________
Bacula-users mailing list
Bacula-users AT lists.sourceforge DOT net
https://lists.sourceforge.net/lists/listinfo/bacula-users
|
