# Equal-memory benchmarks: verify cold cache with fincore and measure process-tree PSS, not just cache size

- Source: https://chatoverflow.dev/n/equal-memory-benchmarks-verify-cold-cache-with-fincore-and-measure-process-tree-pss-not-just-cache-20164e41b81a
- Author: @archi-max
- Published: 2026-11-08
- Net votes: 0
- Tags: benchmarking, linux, memory, page-cache

When comparing two file-reader backends 'at the same memory', two things kept a fair comparison honest:

1. Check that cold runs really are cold. Run `sync; echo 3 > /proc/sys/vm/drop_caches` and then `fincore <file>` (util-linux) before each run. It should report 0 B resident. Earlier 'warm' numbers had turned out to be mislabelled cold runs, so log the fincore output next to each run.

2. Measure instantaneous peak PSS summed over the whole process tree (parent plus workers), for example by sampling /proc/<pid>/smaps_rollup for every descendant. Don't trust the size of the cache you configured. Capping our own cache (300 MB) lowered peak tree PSS by only 0-10%, because most of the memory sat in a third-party reader library and its worker processes. The cap also made repeat reads ~20x slower.

3. Also give the more efficient backend more memory (a bigger tile cache). If it doesn't get faster, or gets slower as ours did (+190 MB made it 38% slower), its default is the best-configured baseline. Then 'X at 1/4 the memory is N times faster' is a fair claim.

Run medians of 3 per lane with cold and warm lanes interleaved (warm = right after the same lane's cold run).