[OmniOS-discuss] Slow NFS writes in 151026

Lee Damon nomad at ee.washington.edu
Thu Aug 23 15:37:01 UTC 2018


I recently installed a new host. So new I couldn't install LTS on it so
I've installed 151026.

This host is strictly for serving ZFS-based NFS & CIFS. Everything else is
just default.

Over time it has become fairly obvious to me that NFS writes are ... well,
abysmal.

This example is copying a 36GB directory of mixed size/type files. The
first copy is strictly on a filesystem on the new server. The second is
reading from the new server to an existing one. The third is doing the same
read/write activity as test one but on an existing server running 151022.

on new fileserver:

: || nomad at omics1 fs2test ; time cp -rp 004test omics1/004test-1

real 22m27.225s
user 0m0.188s
sys 0m29.880s

reading from new fileserver, writing to existing fileserver:

: || nomad at omics1 hvfs2test ; time cp -rp /misc/fs2test/004test .

real 2m9.770s
user 0m0.180s
sys 0m28.694s

existing fileserver:

: || nomad at omics1 hvfs2test ; time cp -rp 004test omics1/004test-1

real 2m14.158s
user 0m0.242s
sys 0m30.313s

While the user and system times are consistent across all tests the wall
clock time of the first test is 10x that of the others. I've seen wall
clock time on these tests take as long as 50 minutes. All tests were done
on the same CentOS 7 host.

Watching snoop collect packets I see multiple-minutes-long pauses while
writing to the new server. If I'm reading the heat maps right -
https://drive.google.com/open?id=1zcX9ryXjrPMH0_uUbfywiTTnJDau4WW0 - it
seems to be spending about 81% of its time in _t_cancel, waiting on a
thread to cancel. I'm not a dev, haven't looked at the code, so it's quite
possible I'm misunderstanding what the map is saying.

The client spends a lot of time so stuck in diskwait that it can take
several minutes to respond after a SIGINT, SIGHUP, or SIGKILL to the cp
process.

Is anyone else seeing similar problems?
nomad
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://omniosce.org/ml-archive/attachments/20180823/3753dfff/attachment.html>


More information about the OmniOS-discuss mailing list