On 11/10/26 12:28, Lloyd Brown via Bacula-users wrote:
On 11/10/26 11:48, Arno Lehmann via Bacula-users wrote:
This really gets to the crux of my original question. Why is the
cycle bad? Suboptimal performance, perhaps, though I'd argue that
for disks the difference is negligible.
We stumbled over that because performance was so bad that users
actually couldn't complete their copy and virtual backups in decent
time, couldn't saturate their tape drives, and found that SD threads,
doing such jobs, wasted more than 90% (iirc) of a CPUs time going
through bootstrap records. Negligible only if you accepted a whole
university's backups to simply not take place any more 🙂
Interesting. This was occurring during the run of the VF?
We certainly didn't experience that. I wonder if the fact that the
old 9.6.6 build had many more jobs per day, and had 100G disk volumes
vs 500G currently on bacula 15, meant we avoided this effect without
realizing it.
Lloyd
This is additional speculation on my part, but I wonder also if it
depends on how interleaved the jobs' data are on the disk volumes. A
long time ago we had performance problems during a restore when it had
to do a lot of seeks within the disk volumes. We switched to having many
`Device` entries in the SD's config, each with `Maximum Concurrent Jobs
= 1`. The result is that interleaving within a specific job run almost
entirely disappeared. And that would mean many fewer transitions
between volumes when running the VF, and therefore fewer and likely
shorter cycles, if they occurred at all.
Lloyd
_______________________________________________
Bacula-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/bacula-users