ZFS scrub CPU bound?

[Ivan Voras]

On 09/08/2012 17:39, Daniel Kalchev wrote:

> CPU:  0.4% user,  0.0% nice, 51.4% system,  0.8% interrupt, 47.3% idle
> Mem: 2171M Active, 1541M Inact, 5407M Wired, 25M Cache, 416K Buf, 22G Free
> Swap: 8192M Total, 8192M Free
> 
>   PID USERNAME  THR PRI NICE   SIZE    RES STATE   C   TIME   WCPU COMMAND
>    11 root       32 155 ki31     0K   512K CPU31  31 190.2H 1818.46% idle
>     0 root      268  -8    0     0K  4288K -       0 141:25 1261.33% kernel
>     4 root        4  -8    -     0K    80K CPU30  30   9:13 95.17% zfskern
>    13 root        3  -8    -     0K    48K -       4   5:37 42.97% geom
>    12 root       66 -84    -     0K  1056K WAIT    0   6:19 30.86% intr
> [...]

> It seems that zfskern will top to 100%. This is an 32 core system, and
> as you see scrub, at 600MB/sec is able to eat 16 cores (from 2x 2.2 GHz

If you hit "H", top will show threads separately, so you will see which
threads exactly do the work (you might need to reduce your font so that
all of them fit on the screen :) ).

> Opteron 6274). There is high load on geom as well... but geom does go
> over 100% CPU, so I suppose it scales.

Actually, GEOM itself is a bottleneck with SSDs and not really scalable
in its current version. If you hit H, you could see that e.g. the g_down
thread is sometimes pinned to 100% CPU or something similar. If all GEOM
threads in your case are individually below 100%, then you didn't hit
the GEOM bottleneck yet.



-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 260 bytes
Desc: OpenPGP digital signature
Url : http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20120820/9df183fb/signature.pgp