On 2018-07-27 14:15, Stefan G. Weichinger wrote:
Am 27.07.2018 um 19:37 schrieb Austin S. Hemmelgarn:
Perhaps I can help with that.
Great stuff, thanks for your informative reply, that's exactly the
information I would like to have in the docs etc
Will consult that in detail asap.
A quick note on what I try to solve here:
I have servers with only one big RAID-array consisting of maybe 4 or 6
physical disks, and based on that (software-)RAID there is one LVM
volume group. So the logical volumes containing the data to be backed up
(DLEs) are on the same array as the other LV providing the amanda
holding disk.
Yes, I know, that's not optimal, though I can't easily change that (I
would have to add separate disks for holding disk purpose ... cost and
space/controller issues)
Don't worry, I've got to deal with similarly sub-optimal stuff where I
work (our backup server has to multiplex all the dumps _and_ taping over
a single GbE connection, so our backups are _always_ network-bound, even
when we do really aggressive compression), so I entirely understand.
So I want to avoid too much parallel activity of dumper and taper
processes because that lets the throughput drop down massively (not to
mention the additional stress on the hardware).
So it would be great to be able to tell amanda "the DLEs coming from the
amanda client which is the amanda server (~localhost) should be dumped
to holdingdisk while no taper processes run"
Or something in that direction.
I will consider reducing maxdumps to 4 as well and test "TTTT" for
tonight's run.
And yes, I also test "holdingdisk no" for some DLEs already: I have big
chunks of VM backups where it doesn't make sense to copy them within the
RAID array ... I tape them directly.
If you're taping to vtapes, you might actually be able to set things up
to not need a holding disk at all. I'm a bit fuzzy on how to configure
it, but I know it's possible to set up vtapes to tape things in
parallel. If you do that, you could (probably, again not 100% certain)
get rid of the holding disk, dump direct to the vtapes, and still have
the dumps run in parallel. That would avoid having to worry about the
taper processes competing with the dumper processes. The only caveat is
that failure to tape would mean failure to dump too, but the number of
situations where you would fail to tape but still be able to dump to the
same array as a holding disk is near zero, and the only one I can think
of off the top of my head is completely avoided by not having a holding
disk.
