Linux 4.20 Scheduler To Better Deal With "Misfit" Tasks On ARM big.LITTLE Systems

[yurikoles]

more much more

Oh, Michael

[polarathene]

You can already do that, you can disable 1 of the 2 CCXs on a die and overclock the remain 4 cores. So on a 2 die Threadripper you can have 1 die with 4 cores active (1 CCX) at 4Ghz and the 2nd die with 8 cores active (2 CCX) at 3Ghz.

Which ThreadRipper were you referring to? The 1950x is 3.4GHz and I recall running it at around 3.8-4GHz on OC with AIO cooling I think(all cores). Why would one die be underclocked to 3GHz if you were disabling a CCX to OC the other in the 2nd die?

I was thinking of a better disparity, EPYC 7601P for example is 2.2 base and 3.2 boost. It'd be nice to get a CCX there to be more like a TR 1950X or Ryzen 2700 CCX clockspeed and boost, even without OC I think a 2700X can boost to 4.3GHz from 3.7GHz?(with no CCX disabled either). I've seen the 2700X get OC to 4.2-4.4GHz, even almost 4.5GHz: https://www.reddit.com/r/Amd/comment...hreaded_loads/

If the TR and EPYC parts are meant to be using higher quality bins, shouldn't their dies be able to achieve something similar to that? Getting one of the EPYC parts to get over 4GHz for single threaded workloads would be great, with a TR I'd hope you'd be able to get a bit further than 4GHz with your OC approach(does it boost beyond that much?)

[dh04000]

I wish either intel or and would release a big.little for the benefit of low thread use cases. 1 or 2 big cores that have the transistor of two traditional cores in one, then a number of small cores with the remaining space to handle background processes, OS processes, hardware processes, network processes, ect. No idea if this is possible though......

[Spazturtle]

Makes me wonder if with AMD's architecture there might be a EPYC cpu at some point where you get many cores but one die provides fewer core count at higher clockspeed. I've had some software with that processing behaviour where it'd be nice to have more cores/threads for multi-threaded but the single-threaded workload suffers as the clockspeed takes a dive

You can already do that, you can disable 1 of the 2 CCXs on a die and overclock the remain 4 cores. So on a 2 die Threadripper you can have 1 die with 4 cores active (1 CCX) at 4Ghz and the 2nd die with 8 cores active (2 CCX) at 3Ghz.

[polarathene]

So if you have a workload that is multi-threaded and pushes all cores, then it switches to single-threaded, that thread might be on a less performant core but with these changes will now be migrated to a more suitable one? That's great

Makes me wonder if with AMD's architecture there might be a EPYC cpu at some point where you get many cores but one die provides fewer core count at higher clockspeed. I've had some software with that processing behaviour where it'd be nice to have more cores/threads for multi-threaded but the single-threaded workload suffers as the clockspeed takes a dive

[phoronix]

Phoronix: Linux 4.20 Scheduler To Better Deal With "Misfit" Tasks On ARM big.LITTLE Systems

The next Linux kernel has scheduler improvements that will benefit some tasks when running on ARM big.LITTLE type systems where select CPU cores are more much more powerful than the other cores...

http://www.phoronix.com/scan.php?page=news_item&px=Linux-4.20-Scheduler-Misfit