Per-agent memory: report private (anon) and PSS separately; file-backed readers look heavier in PSS because of the page cache

When comparing memory per client for two file readers under concurrency, one number misleads. Report at least two: - anon or private memory per process (from /proc/<pid>/smaps_rollup: Anonymous, or Private_Dirty). This is what the reader itself allocates. - PSS per process, which includes a proportional share of file-backed pages (Pss_File). A reader that maps and touches more of the file (whole index, footers) shows higher PSS even if its private heap is smaller.

In our case, at 90 concurrent clients one reader had 21 MB anon per agent against 29 MB for the other, but 39 MB against 29 MB PSS, because it pulled more of the file into the page cache.

Notes: - Page cache is shared across processes reading the same file, and it is reclaimable. It still counts against the box under pressure. - When every agent reads a different file there is no sharing, so per-agent Pss_File roughly equals the file bytes it touched. - Bytes read from disk per session are a third, separate metric (/proc/<pid>/io read_bytes). Do not conflate them with memory. - Say which metric binds capacity (memory, CPU, IO latency) from the ladder data, rather than assuming.