On Sun, Dec 22, 2024 at 03:21:32PM +0800, Huang, Ying wrote: > Hyeonggon Yoo <hyeonggon.yoo@xxxxxx> writes: > > > On this server, ideally weighted interleaving should be configured > > within a socket (e.g. local NUMA node + local CXL node) because > > weighted interleaving does not consider the bandwidth when accessed > > from a remote socket. > > If multiple sockets are considered, what is the best behavior? > > The process may be cross-socket too. So, we will need to use > set_mempolicy() to bind tasks to sockets firstly. Then, it may be > better to use per-task weights. > If we want to revisit this, we might be able to make task-local weights work without a new syscall, but the use case was not clear enough which is why it was soft-nak'd originally. vma-local weights are arguably more usable, but require the task to be numa-aware and probably require a new mempolicy syscall because mbind has no remaining arguments. recall my original testing results from stream: https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@xxxxxxxxxxxx/ Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device) Default interleave : -78% (slower than DRAM) Global weighting : -6% to +4% (workload dependant) Targeted weights : +2.5% to +4% (consistently better than DRAM) Just some context ~Gregory
