LGTM! Thanks for doing the testing too.
- George
On 5/6/15 3:14 AM, Arne Jansen wrote:
On 05/06/2015 01:41 AM, George Wilson wrote:
line# 1793
- given that you found performance improvements by multiplying the leaf
count by 4 it might be a good idea to make this a tunable.
On 05/05/2015 11:28 PM, Matthew Ahrens wrote:
Theoretically, I think that the ideal number of read i/os depends on whether the
vdevs are mirror (should be based on # leafs, because read hits only one leaf)
vs raidz (should be based on # toplevels, because read hits (nearly) all leafs.
I'm fine with the code as is for now; we can revisit if it becomes a problem.
I added a tunable, mainly to make it easier for me to do a few more test runs
to convince you that no tunable is needed :)
I tested on a zpool with 30 leaf vdevs in 4 raidz2 groups, spinning disks and
about 23000 filesystems on it.
threads time for import
1 11m17s
2 5m55s
4 3m21s
8 1m49s
16 1m7s
32 0m49s
64 0m41s
128 0m38.6s
256 0m35.7s
512 0m31.4s
1024 0m31.9s
2048 ~0m38s
The figures vary a bit from run to run, but starting with 1024 threads they
become quite unstable. In the last few runs the phase with the 2 find_objset
loops takes about 20s.
So I think 4x leaf vdevs (120 in this case) is a good compromise between
performance and risking to have too many threads.
Again, the webrev is updated and attached is the revised patch and the diff to
the previous version.
-Arne
-------------------------------------------
illumos-zfs
Archives: https://www.listbox.com/member/archive/182191/=now
RSS Feed: https://www.listbox.com/member/archive/rss/182191/22008002-303f2ff4
Modify Your Subscription:
https://www.listbox.com/member/?member_id=22008002&id_secret=22008002-f948b352
Powered by Listbox: http://www.listbox.com
_______________________________________________
developer mailing list
[email protected]
http://lists.open-zfs.org/mailman/listinfo/developer
