Now the question is, how terrible of an idea is it to try playing VR games on my Threadripper in a VM while it's compiling Gentoo packages? I mean, emerge is niced, I have an NVMe drive, and RAM to spare. It *should* work... right? The scheduler *should* handle it, right? Right?
-
Show this thread
-
Report: that worked until GCC started compiling. Turns out no, the scheduler cannot handle it. VM was already isolated to one NUMA node, so now I'm sending emerge off to the other one (CPUs only, let's let it use all free RAM).
2 replies 0 retweets 46 likesShow this thread -
Replying to @marcan42
Are you specifically pinning the vCPUs to particular cores? Or just restricting them to one NUMA node by some other means
1 reply 0 retweets 0 likes -
Replying to @JohnHedge
The 8 vCPUs are pinned to 8 physical cores on one NUMA node (8 SMT siblings left unused), and the vRAM is pinned to the adjacent memory controller, and the GPU is attached to PCIe lanes on that side. emerge is allowed to use all RAM and the 16 threads on the other NUMA node.
2 replies 0 retweets 3 likes -
Replying to @marcan42 @JohnHedge
on my setup (1920X), I have cgroups set up to reserve cores and RAM from one NUMA node for the VM also, based on advice I saw online, I expose each CCX within that node as a NUMA node within the VM not sure how much that helps performance, but it definitely didn't make it worse
1 reply 0 retweets 1 like -
also, I pin all 12 vCPUs to a specific hardware thread (so the VM does get SMT)
1 reply 0 retweets 1 like
SMT requires special scheduler consideration to not end up pathological, and I don't trust Windows enough not to screw that up, and it can't use all 8 cores for games anyway, hence avoiding SMT is the safe bet for me.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.