Measuring a process's memory on Linux: virtual size, RSS, PSS and what each answers

Goal

Answer "how much memory does this process use" with the number that fits the question being asked: will it fit on this machine, is it leaking, or how much does it add on top of the processes already running.

Prerequisites

Linux with read access to /proc/PID. Vocabulary from the manual pages: VmSize is virtual memory size and VmPeak its peak; VmRSS is the resident set size, the sum of RssAnon (anonymous memory), RssFile (resident file mappings) and RssShmem (shared memory); VmHWM is the peak resident set; VmSwap is swapped-out anonymous memory. The manual page marks the RSS values, VmHWM and VmSwap as inaccurate and refers to /proc/PID/statm. In /proc/PID/smaps each mapping shows Rss and Pss, the process's proportional share of the mapping, which divides shared pages among the processes sharing them; the kernel documentation describes smaps_rollup as the same fields as smaps with values summed over all mappings of the process, which could be derived from smaps but at significantly higher cost.

Steps

1. Read /proc/PID/status and note VmSize, VmRSS, RssAnon, RssFile, RssShmem, VmHWM and VmSwap with a timestamp.
2. Do not use VmSize for fitting decisions. It counts reserved address space (memory-mapped files, thread stacks, allocator arenas, a runtime's reserved heap) that may never be touched; gigabytes of virtual with a modest RSS is normal.
3. For "will it fit": treat RssAnon + VmSwap as the process's own anonymous cost, RssShmem as memory that stays allocated while any sharer holds it, and RssFile as page cache that the kernel can drop and reload.
4. For "what does it add" across many processes sharing libraries or a copy-on-write parent: sum Pss from /proc/PID/smaps_rollup. Summing VmRSS across processes counts each shared page once per process.
5. For "is it leaking": sample RssAnon and VmHWM at fixed intervals under constant load. Growth that never plateaus points to a leak or to an allocator that keeps freed memory; a plateau is the working set.
6. Record PID, command line, kernel version, applied load and the numbers so a second run is comparable.

Expected result

Three numbers with distinct meanings for the same process: address space reserved, memory resident now (shared pages included), and the proportional share. A memory graph labelled with which of the three it plots.

Limits and test basis

Allocators retain freed memory, so RSS overstates live objects; language-level heap statistics (tracemalloc, a runtime's heap profile) measure a different thing and are not directly comparable. Memory-mapped files and huge pages complicate the split. Inside a container the cgroup counts more than the process's RSS; see the open question on container memory metrics. Based on the cited manual pages; no measurements are claimed.