Gerhard den Hollander <[EMAIL PROTECTED]> writes:
> * Johannes Niess <[EMAIL PROTECTED]> (Thu, Feb 01, 2001 at 04:01:03PM +0100)
>
> > My first thought was to disable compression, but that's already
> > done. What's your network technology? 2 MBytes/sec is not the optimum
>
> They're all local disks
> (Wide scsi 2, so that's 80 Mbaud IIRC)
>
> > On the other hand disk speed might be the bottle neck. I'd run bonnie
> > on the clients and the server holding disk. I assume you've checked
> > SCSI settings/errors/DMA for IDE controllers etc. of the holding disk.
Let's do the math:
We need 10 M/s sustained transfer rate from holding disk to the tape
and 10 M/s from data disk to holding disk. A 80 M/s SCSI bus should be
able to do that.
20 M/s with a lot of seeks are quite a load for a single holding
disk. Depending on you holding disk hardware I'd guestimate 3 M/s a
reasonable throughput under these circumstances. Two simultaneous runs
of bonnie (read/write) on the same disk could give reasonably good
values. How does amanda determine the block size of writes/reads? I
hope that there are large buffers on all sides between client, server
and tape writing parts. RAID in stripping (or concatenation) mode
might help on the holding disks. Different physical holding disks
might do the same trick. What exactly is meant by "round robing" of
holding disks? It looks like you need two holding disks for maximum
performance: 1 for writing to , one for simultaneous reads. In your
case I'd set maxdumps=1, too. The extended formula is N = D + 1, where
N is the optimum number of independend holding disks and D is the
number of simultaneous dumpers, 1 is for the taper reading data.
Nobody asked about seek problems on your data disks. It definitely
kills performance to dump two partitions of the same disk at the same
time. Amanda's default value of the spindle parameter is not at an
intuitive value:
(man amanda, disk list parameters)
spindle
Default: -1. A number used to balance backup load
on a host. Amanda will not run multiple backups at
the same time on the same spindle, unless the spin
dle number is -1, which means there is no spindle
restriction.
This assumption might date from the times when disks where too small
to be split into pieces. IMHO 1 is a more reasonable default as it
does only one dump at a time on a host.
HTH,
Johannes Niess
P.S: Feel free to make a faq-o-matic section on performace out of
this. What about a weekly post of the FAQ to this list?
