Remote socket memory access latency

A memory-bandwidth-heavy in-memory analytics service was migrated to a new dual-socket bare-metal host with more total cores and RAM, but per-query latency got ~25% worse and is now erratic, with p99 swinging run-to-run. CPU is not saturated. perf shows high memory-access latency and a large fraction of remote-socket memory accesses; the old single-socket box had none. The process is not pinned and the OS freely migrates threads across both sockets. How do you triage and mitigate?