| # Does NUMA interleave help the real engine? Only if the weights are in ANONYMOUS memory: | |
| # when the file is served from page cache the pages keep the placement they were faulted with, | |
| # so numactl's policy has no effect. -mm 0 forces a fresh anonymous read under the policy. | |
| cd /home/bob/OldIron/engine || exit 1 | |
| M=/mnt/flare/OldIron/weights/Qwen3.6-35B-A3B-Q4_0-otq4.gguf | |
| R=/home/bob/OldIron/results | |
| echo "=== interleave, fresh anon pages (-mm 0) ===" | |
| numactl --interleave=all ./bin/llama-bench -m "$M" -ngl 0 -nopo 1 --load-mode none -t 24 -p 128 -n 48 -r 2 > "$R/numa-interleave-nommap.log" 2>&1 | |
| echo "=== default placement, fresh anon pages (-mm 0) ===" | |
| ./bin/llama-bench -m "$M" -ngl 0 -nopo 1 --load-mode none -t 24 -p 128 -n 48 -r 2 > "$R/numa-default-nommap.log" 2>&1 | |
| echo "=== done ===" | |
| grep -hE "tg48|pp128" "$R/numa-interleave-nommap.log" "$R/numa-default-nommap.log" 2>/dev/null | sed 's/^ *|//' | |