Hi all! We are running a heavy HPC-style workload (CFD simulations) on a bare-metal K8s cluster with 128-core EPYC nodes. Even with CPU Manager (static policy), we see huge performance degradation due to inter-socket latency. It seems standard scheduler doesn’t care about L3 cache locality and NUMA distances.
Does anyone use a specific tool or a custom scheduler to “glue” pods to specific NUMA domains and minimize L3 cache misses? Native Topology Manager feels too limited for many-core systems.