| # Does interleaved PLACEMENT of the weights help the real engine? | |
| # | |
| # The earlier attempt was invalid: when the GGUF is already in page cache, `numactl --interleave=all` cannot | |
| # re-place pages that were faulted under a different policy, and `--load-mode none` (anonymous copy) costs ~13% | |
| # which cancels the gain. The only honest way is to materialise a NEW file while the interleave policy is in | |
| # force - the page-cache pages of the copy are allocated by the copying process, so they inherit the policy. | |
| # | |
| # Usage: ./interleave-test.sh <model.gguf> | |
| # Leaves <model>-il.gguf behind (same size again on disk). | |
| set -uo pipefail | |
| TARGET=${TARGET:?set TARGET to the ssh alias of the target ISA host} | |
| SRC=${1:?usage: interleave-test.sh <model.gguf>} | |
| IL="${SRC%.gguf}-il.gguf" | |
| RDIR=/home/bob/OldIron/results | |
| THREADS=${THREADS:-24} | |
| NICE="nice -n 19" | |
| echo "=== 1. materialise an interleaved copy (one-off, this is the expensive step) ===" | |
| if ssh -o BatchMode=yes "$TARGET" "[ -f '$IL' ]"; then | |
| echo " already exists: $IL" | |
| else | |
| ssh -o BatchMode=yes "$TARGET" "numactl --interleave=all cp '$SRC' '$IL' && ls -l '$IL'" \ | |
| || { echo "FAILED: copy"; exit 1; } | |
| fi | |
| echo "=== 2. A/B on the engine: default placement vs interleaved copy (same file contents) ===" | |
| for round in 1 2; do | |
| for tag in default interleaved; do | |
| [ "$tag" = default ] && M="$SRC" || M="$IL" | |
| printf ' round %s %-12s ' "$round" "$tag" | |
| ssh -o BatchMode=yes "$TARGET" "cd /home/bob/OldIron/engine && $NICE ./bin/llama-bench -m '$M' -ngl 0 -nopo 1 -t $THREADS -p 0 -n 48 -r 1" 2>/dev/null \ | |
| | grep tg48 | grep -oE '[0-9]+\.[0-9]+ ± [0-9.]+' | |
| done | |
| done | |
| echo "=== 3. what the pages actually look like (numastat for the bench's mmap) ===" | |
| ssh -o BatchMode=yes "$TARGET" "numastat -m 2>/dev/null | grep -E 'MemFree|FilePages' | head -4" | |