New Scheduler Optimization Can Help Out PostgreSQL & More On Sapphire Rapids

[phoronix]

Phoronix: New Scheduler Optimization Can Help Out PostgreSQL & More On Sapphire Rapids

Stemming from Intel engineers finding significant overhead in some Linux scheduler functions when running PostgreSQL within a Docker instance, a new scheduler patch is on the way for Linux 6.7 that will help out at least Ice Lake and Sapphire Rapids with some migration-heavy workloads. With the change being in the common scheduler code, it's also likely to help out other hardware platforms too...

https://www.phoronix.com/news/Linux-Sched-Postgres-SPR

[Kjell]

Podman/Docker is amazing. Hope this patch is compatible with EEVDF!

[milkylainen]

Would be more interesting to know why this is an issue on select CPUs?
Symptom fix by decreasing the update rate, sounds like a bad idea.

[kgardas]

Would be more interesting to know why this is an issue on select CPUs?
Symptom fix by decreasing the update rate, sounds like a bad idea.

Well, those selected CPUs are Intel last two server generations. Do you expect Intel own engineers to test/debug/work on competitive CPUs? I would not. And btw, communication between CPUs is quite expensive...

[sdack]

Would be more interesting to know why this is an issue on select CPUs?
Symptom fix by decreasing the update rate, sounds like a bad idea.

It is not a symptom fix when they have identified the source being too many updates to a single value. The load average is an approximation and a low priority value (i.e. the number of active processes is more important). Limiting updates to the average to a millisecond is still generous. An application, which needs a real-time value of this number should use the count of currently active processes instead of an approximated average.

The reason why it is a problem is because of cache updates. When you have i.e. 128 cores, each with its own L2/L1 cache, then all the caches holding a copy of this value need to be updated. When this happens for a low priority value like the load average then it is considered a waste of bandwidth between the caches when the bandwidth could be used for more important cache updates.