On Wed, Nov 24, 2021 at 2:18 AM Richard Purdie
On Wed, 2021-11-24 at 09:00 +0100, Alexander Kanavin wrote:
But the AB has not been exhibiting any problems with zstd, and this willI'm not sure I agree with that.
degrade performance. Let's only fix what is broken.
We have 60+ "intermittent" bugs and some of us are in weekly meetings trying to
do something about working out why these are failing. It feels like we're not
really getting too far with some subset of them and it is using up a lot of the
SWAT and bug triage time.
We've made a few changes to try and reduce the load spikes on the systems and
this fits with the other changes we've made.
From a different data point, we have clipped the parallelism for XZ
and ZSTD internally to very low ( 2 and 4 )
and it has in fact reduced unexpected failures and seen no impact on
build performance. Since parallelism settings
are myopic for these tools, they greedily take every CPU which is not
best for overall build thats doing many
things in parallel, so its best to curtail them to a conservative value.