Yocto Autobuilder: Latency Monitor and AB-INT - Meeting notes: Sept 9, 2021

Randy MacLeod

YP AB Intermittent failures meeting
Sept 9, 2021, 9 AM ET

Attendees: Kiran, Richard, Trevor, Randy!


Ptest results continue to improve but there's still room
for even more improvement.

The make/ninja load average limit is in but it's not clear
if it's effective yet and it breaks dunfell. Trevor investigating
(Aug 26!, Sept 9). Trevor is going to get back to this!

If anyone wants to help, we could use more eyes on the logs,
particularly the summary logs and understanding iostat #
when the dd test times out.

Plans for the week:

Richard: QA results for M3, etc.
Alex: SWAT plans. September email, training. libevent monotonic fix!
Sakib: hook more responsive load average in to latency test. (v3)
Trevor: patch to set PARALLEL_MAKE : -l 50
-> dunfell, gatesgarth, hardknott (Aug 5 - it's a priority)
Investigate dunfell which failed with this change.
- data on WR AB load average.
Saul: SBOM
Randy: Vacation
Kiran: probably work on one of the ptest failures.

Nothing much new below here. Keeping the list since it's still to-do.


Meeting Notes:

1. job server

- ninja could be patched with make's more responsive algorithm
next or is this good enough?

Aug 26:
Randy made some graphs that show that the -l NUM results
in the number of compile jobs oscillates *wildly* between 0 and 200
on a 192 core builder compiling chromium. What I did was:
$ bitbake -c cleansstate chromium-x11
$ bitbake -c configure chromium-x11
$ bitbake -c compile chromium-x11
and while that compile was running:
$ while [ ! -f /tmp/compiling-chromium-is-done ]; do \
cat /proc/loadavg >> procs-load.log ; sleep 0.5 ;
Results so far:
Next step is either:
a. collect data as above for an image build and see if the sub-optimal
ninja behaviour makes a difference
b. patch ninja with make's more responsive load avg

- Richard suggested that we extract make's code for measuring the load
average to a separate binary and run it in the periodic io latency
test. Also can we translate it to python?
- Trevor is working on this and had some problems so next week.
(Aug 19 - Trevor is back from vaction so maybe next week.)

- Trevor to see if the load average change really did reduce load
on WR build systems. (Aug 19)

2. AB status

Trevor is learning about buildbot and working on a scheduling bug
(CentOS worker?)

bitbake layer setup tool should allow multiple backends:
eg: kas, a y-a-helper.

ptest cases are improving, we may be close to done!
Let's wait a week to see how things go.
(July29, Aug 5, Aug 19, we're not done...)

- lttng-tools ptest is failing. RP is working on it with upstream.
The timeout (done on Aug 5) increase hasn't helped.

3. Sakib's improvements to the logging are merged.

Sakib generated a summary of all high latency 'top' logs from
~July 23->July 29 by just running his summary script on the
merged raw top logs.

More analysis required....

Still relevant parts of
Previous Meeting Notes:

4. bitbake server timeout ( no change july 29, Aug 19)

"Timeout while waiting for a reply from the bitbake server (60s)"

5. io stalls (no update: July 29)

Richard said that it would make sense to write an ftrace utility
/ script to monitor io latency and we could install it with sudo
Ch^W mentioned ftrace on IRC.
Sakib and Randy will work on that but not for a week or two
or longer! (Aug 19).

Randy collected iostat data on 3 build server:
We agreed that having -ty-2 be ~ 100 utilization for many hours
in a row is not acceptable and that a threshold of ~ 10 minutes
at 100% utilization may be a reasonable limt. I need to figure out
if I can get data on the fraction of IO done per IO clas since
we do use ionice to do clean-up and other activities.


Join yocto@lists.yoctoproject.org to automatically receive all group messages.