[Bug 266302] [zfs][iscsi] Periodic drops by ctl with "failed to allocate soft PDU" since 13.1

From: <bugzilla-noreply_at_freebsd.org>
Date: Tue, 20 Dec 2022 00:10:05 UTC
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=266302

--- Comment #2 from eborisch+FreeBSD@gmail.com ---
What's interesting is that, in the latest one I have data for, the free memory
never fell below 10GB. (Granted, that is only being sampled every 30s in my
logging, so it could have much lower briefly.)

I see a spike (monitoring vmstat -z) of a large number of frees just after the
error was detected -- abd_chunk, zio_buf_comb_131072, and dnode_t being the
largest, freeing a combined 2.2GB, and a spike in failures for 'vm pgcache'
immediately before it, and continuing immediately after the failure. I've put a
copy of the logging (1m) of vmstat -z and some analysis here:
https://github.com/eborisch/pdu-error/blob/main/pdu_analysis.ipynb

The log stops _after_ the error was detected, so it has considerable data
leading up to, and then one data point after an instance of this error, if
you're interested in looking at it.

In the interim, I can further restrict vfs.zfs.arc_max; I already have it set
to  75% of the 64GB system.

-- 
You are receiving this mail because:
You are the assignee for the bug.