This benchmark regularly varies by +/- 15% max-rss on any given perf run. This should be indicated somehow, so that changes aren't regarded as regressions/improvements when it's just variance. If we don't have per-statistic variance indications, perhaps it would be good to add them.