Yocto Autobuilder: Latency Monitor and AB-INT - Meeting notes: July 8, 2021


Trevor Gamblin
 

On 2021-07-08 9:53 a.m., Randy MacLeod wrote:

YP AB Intermittent failures meeting
===================================
July 1, 2021, 9 AM ET
https://windriver.zoom.us/j/3696693975

Attendees:  Tony, Richard, Trevor, Randy


Summary:

========



The autobuilder RCU hang is still fixed ;-),
   master branch builds greener.

ptest failures are the top problem now.



Add Michael Halstead, see questions below in section 4.



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: glibc upgrade, etc.



  Alex: ?



  Sakib: pub/non-release link upgrade, script clean-up.



  Trevor: make job server test. Try it on YP AB!!! What type of build?



  Tony: fix/work-around valgrind ptest bug:
     none/tests/amd64/fb_test_amd64
  Saul: nothing this week for YP.



  Randy: vacation, then email catch-up!




Meeting Notes:

==============


1, runqemu

Tony having trouble with runqemu on some Wind River machine.

Richard has a fix for a race in runqemu in master-next.

These might be related but if not Tony should debug the
issue/ collect logs.

2. job server

- Trevor has conclusive evidence that the 'make' job server is useful.
  email summary to come. Need to fix some assumptoins in code that
  parses PARALLEL_MAKE then send patch to yocto-autobuilder-helper.

- ninja will have to be done next.
I've submitted an initial patch to yocto@.... My short notes to go along with it:

MAKE JOBSERVERS, SYSTEM LOADS, AND THE YOCTO AUTOBUILDER

INTRO

- Yocto Autobuilder has been experiencing several issues with nightly builds, some intermittent
- Previous discussions looking for ways to limit CPU/RAM loading (see " [yocto] bitbake controlling memory use" on yocto@... mailing list)
- Source packages built as part of YP builds rely on various build tools, including Make, Ninja. Can we use them to limit system resource loading, and therefore reduce build failures that occur near/beyond these limits?

MAKE JOBSERVER

- When the "-j" option is provided (e.g. by PARALLEL_MAKE and/or EXTRA_OEMAKE in local.conf), this tells Make to use that many job slots, e.g. PARALLEL_MAKE = "-j 10" tells make to use at most 10 job slots during compilation
- When the "-l" or "--max-load" option is provided (also via PARALLEL_MAKE or EXTRA_OEMAKE in local.conf), this tells Make not to start *more than one job* if the system load average (as perceived by Make) exceeds that value. If the system load exceeds the average specified with the "-l" argument and there is already a Make job running, no new ones are started until either the existing jobs finish, or the system load falls below the threshold.
- No limit by default

Sources:
https://www.gnu.org/software/make/manual/make.html#Parallel
http://make.mad-scientist.net/papers/jobserver-implementation/

PROPOSED SOLUTION

- "Just do it" - put a patch in for use with the Autobuilder and see what shakes out
- add 'PARALLEL_MAKE = "-j N -l M"' to the yocto-autobuilder-helper config.json default
- Complicated slightly by Ninja not supporting the same "--debug" functionality as Make - this is what has been used to track the system load in the initial testing. To get around this, we can temporarily set EXTRA_OEMAKE. Without "--debug", the "-l" option still works, but there is no output in the do_compile logs to tell us how the actual system load compares to the percentage targeted by the "-l" option

I did not see any issues in my manual testing with PARALLEL_MAKE and EXTRA_OEMAKE set in the same manner. It appears from a recursive grep through the poky repository that PARALLEL_MAKE gets referenced in various places, but it is usually either passed directly to the build tool of choice, or the argument to the "-j" flag is explicitly stripped out for use. An example of the former is meta/classes/meson.bbclass:

...

meson_do_compile() {
    ninja -v ${PARALLEL_MAKE}
}

...

It's possible that we will run into some edge cases of course, but I haven't seen them in my manual testing. Builds are running now at https://autobuilder.yoctoproject.org/typhoon/#/builders/85/builds/1478 , so we should have some initial results soon.

- Trevor



3. AB status

 generally better but...

 ptests are having some recurring problems.
 - parted - only on arm?
 - valgrind - none/tests/amd64/fb_test_amd64
 - gdb test failing again. - Randy!

4. Richard reported
   - something really flaky going on with serial ports.
   - particularly bad on qemuppc but also x86.
   - related to Saul's QMP data dump?

5. Sakib needs to send patch to make testimage failures
   generate summary logs.

6. Richard says that we may need to redesign the data collection system
   that Sakib's AB INT tests are based on.





Still relevant parts of
Previous Meeting Notes:
=======================

1. The qemu RCU hang has been fixed to not deadlock anymore!

It still  hangs at times but this dramatically reduces the
AB failures.



4. bitbake server timeout.

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

   Randy mentioned that the bitbake server timeouts seen in the
   Wind River build cluster have gone away after upgrading to
   a newer version of docker.

   Old: Docker Version: Docker version 18.09.4, build d14af54266
   New: Docker Version: Docker version 20.10.7, build f0df350


   Clearly the YP ABs aren't running in docker but what
   about firmware and kernel tunings.

   Michael,

   Is the BIOS/firmware kept up to date on most nodes?

   It seems that we are running stock kernels which makes sense but
   given that we don't have concerns about privacy since system access
   is controlled and the nodes are being used to test open
   source software, we might consider optimizing for performance
   rather than security.

   Alex pointed at: https://make-linux-fast-again.com/
   Which just lists a set of kernel boot options:
      noibrs noibpb nopti nospectre_v2 nospectre_v1  \
      l1tf=off nospec_store_bypass_disable no_stf_barrier \
      mds=off tsx=on tsx_async_abort=off mitigations=off

   Can we enable some or all of these on a node to see what the
   performance difference is?


5. io stalls

   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.

6. Switch the pub/non-release links from full log to summary.
   The host data links on:
     https://autobuilder.yocto.io/pub/non-release/
   should include links to the summary data. I think we have room to
   include both links like this:
   0 1 2 3  (Full: 0 1 2 3 )





../Randy


Randy MacLeod
 

YP AB Intermittent failures meeting
===================================
July 1, 2021, 9 AM ET
https://windriver.zoom.us/j/3696693975

Attendees: Tony, Richard, Trevor, Randy


Summary:

========



The autobuilder RCU hang is still fixed ;-),
master branch builds greener.

ptest failures are the top problem now.



Add Michael Halstead, see questions below in section 4.



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: glibc upgrade, etc.



Alex: ?



Sakib: pub/non-release link upgrade, script clean-up.



Trevor: make job server test. Try it on YP AB!!! What type of build?



Tony: fix/work-around valgrind ptest bug:
none/tests/amd64/fb_test_amd64
Saul: nothing this week for YP.



Randy: vacation, then email catch-up!




Meeting Notes:

==============


1, runqemu

Tony having trouble with runqemu on some Wind River machine.

Richard has a fix for a race in runqemu in master-next.

These might be related but if not Tony should debug the
issue/ collect logs.

2. job server

- Trevor has conclusive evidence that the 'make' job server is useful.
email summary to come. Need to fix some assumptoins in code that
parses PARALLEL_MAKE then send patch to yocto-autobuilder-helper.

- ninja will have to be done next.


3. AB status

generally better but...

ptests are having some recurring problems.
- parted - only on arm?
- valgrind - none/tests/amd64/fb_test_amd64
- gdb test failing again. - Randy!

4. Richard reported
- something really flaky going on with serial ports.
- particularly bad on qemuppc but also x86.
- related to Saul's QMP data dump?

5. Sakib needs to send patch to make testimage failures
generate summary logs.

6. Richard says that we may need to redesign the data collection system
that Sakib's AB INT tests are based on.





Still relevant parts of
Previous Meeting Notes:
=======================

1. The qemu RCU hang has been fixed to not deadlock anymore!

It still hangs at times but this dramatically reduces the
AB failures.



4. bitbake server timeout.

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

Randy mentioned that the bitbake server timeouts seen in the
Wind River build cluster have gone away after upgrading to
a newer version of docker.

Old: Docker Version: Docker version 18.09.4, build d14af54266
New: Docker Version: Docker version 20.10.7, build f0df350


Clearly the YP ABs aren't running in docker but what
about firmware and kernel tunings.

Michael,

Is the BIOS/firmware kept up to date on most nodes?

It seems that we are running stock kernels which makes sense but
given that we don't have concerns about privacy since system access
is controlled and the nodes are being used to test open
source software, we might consider optimizing for performance
rather than security.

Alex pointed at: https://make-linux-fast-again.com/
Which just lists a set of kernel boot options:
noibrs noibpb nopti nospectre_v2 nospectre_v1 \
l1tf=off nospec_store_bypass_disable no_stf_barrier \
mds=off tsx=on tsx_async_abort=off mitigations=off

Can we enable some or all of these on a node to see what the
performance difference is?


5. io stalls

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.

6. Switch the pub/non-release links from full log to summary.
The host data links on:
https://autobuilder.yocto.io/pub/non-release/
should include links to the summary data. I think we have room to
include both links like this:
0 1 2 3 (Full: 0 1 2 3 )





../Randy