I've recently started doing ON nightly builds on zfs filesystems on the
internal ATA disk of a Blade 1500 running snv_42. Unfortunately, the
builds are extremely slow compared to building on an external IEEE 1394
disk attached to the same machine:
ATA disk:
==== Elapsed build time (DEBUG) ====
real 21:40:57.7
user 4:32:15.6
sys 8:22:24.1
IEEE 1394 disk:
==== Elapsed build time (DEBUG) ====
real 6:14:11.4
user 4:28:54.1
sys 36:04.1
Running kernel profile with lockstat (lockstat -kIW -D 20 sleep 300), I
find in the ATA case:
Profiling interrupt: 29117 events in 300.142 seconds (97 events/sec)
Count indv cuml rcnt nsec Hottest CPU+PIL Caller
-------------------------------------------------------------------------------
15082 52% 52% 0.00 1492 cpu[0] (usermode)
9565 33% 85% 0.00 318 cpu[0] usec_delay
compared to IEEE 1394:
Profiling interrupt: 29195 events in 300.969 seconds (97 events/sec)
Count indv cuml rcnt nsec Hottest CPU+PIL Caller
-------------------------------------------------------------------------------
20042 69% 69% 0.00 2000 cpu[0] (usermode)
5414 19% 87% 0.00 317 cpu[0] usec_delay
At other times, the kernel time can be even as high as 80%. Unfortunately,
I've not been able to investigate how usec_delay is called since there's no
fbt provider for that function (nor for the alternative entry point
drv_usecwait found in uts/sun4[uv]/cpu/common_asm.s), so I'm a bit stuck
how to further investigate this. I suspect that the dad(7D) driver is the
culprit, but it is only included in the closed tarball. In the EDU S9
sources, I find that dcd_flush_cache() calls drv_usecwait(1000000), which
might be the cause of this.
How should I proceed to further investigate this, and can this be fixed
somehow? This way, the machine is almost unusable as a build machine.
Rainer
-----------------------------------------------------------------------------
Rainer Orth, Faculty of Technology, Bielefeld University
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
