GH repo to measure core-core latency

By: anon2 (anon.delete@this.anon.com), September 19, 2022 8:14 pm
Room: Moderated Discussions
Brendan (btrotter.delete@this.gmail.com) on September 19, 2022 6:14 am wrote:
> anonymous2 (anonymous2.delete@this.example.com) on September 18, 2022 2:40 pm wrote:
> > Do OS schedulers take this into accounts for loads that have a lot of locking?
>
> Ahaha haha ha-ha!
>
>
> Oh wait, you're serious. Let me laugh harder!
>
>
> HAHA haha haa hee-heee. Haha Ha!
>
> Half the people designing schedulers today think we're using time-sharing mainframes from the 1960s and
> that "fair" isn't a mental illness; and the other half still haven't realized that CPUs have caches.
>
> - Brendan
>

Schedulers of today, and in Linux as of a decade or two, have detailed models of cache hierarchy and try to keep related processes close to one another, detect producer-consumer or lock wait-wake relationships between tasks and move them together, etc. So the answer to OP's question is yes.

To a matter of degree of course. There's only so much computation you want to be performing in the scheduler, only so much information you can get from userspace, and any heuristic will of course always have failures. Are schedulers optimal? No. Could they be improved? Undoubtedly. It's just a hard problem though. It's easy to point to your workload and your machine and say the scheduler is braindamaged crap because it probably is for your case, but try to improve it and you will find it's very difficult. And then improving it without causing undue regressions for other cases becomes a huge amount of work. At some point, edge cases and getting performance from 80% to 95% on your complicated parallel workload takes customization in the application and administration layers, and probably always will.
< Previous Post in Thread 
TopicPosted ByDate
GH repo to measure core-core latencyanonymous22022/09/18 02:40 PM
  GH repo to measure core-core latencyFoyle2022/09/19 04:08 AM
    GH repo to measure core-core latencyMichael S2022/09/20 02:56 AM
  GH repo to measure core-core latencyBrendan2022/09/19 06:14 AM
    GH repo to measure core-core latencyanon22022/09/19 08:14 PM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell tangerine? ūüćä