Sampling peak RSS of a whole process tree on Linux with ps + awk (no /proc walking script)
To measure the memory of a GUI app and all its children over a run (e.g. a launcher that forks the real binary plus helpers), start it in the background, take VP=$!, and sample once a second while kill -0 $VP:
ps -e -o pid=,ppid=,rss= | awk -v r=$VP -v t=$(date +%s%3N) '{p[$1]=$2; m[$1]=$3} END{do{c=0; for(k in p) if(!(k in in_) && (k==r || (p[k] in in_))){in_[k]=1; c=1}}while(c); s=0; n=0; for(k in in_){s+=m[k]; n++} print t, s, n}' >> rss.log
The do/while grows the descendant set until it stops changing, so grandchildren count too. Peak = max of column 2 (KB). Gotchas: if the script itself is fed via ssh host sh -s < script, redirect the backgrounded app with < /dev/null, or it can read the rest of your script from stdin. RSS double-counts shared pages across processes; it is an upper bound, PSS (smem or /proc/PID/smaps_rollup) is fairer when comparing tools. A single sample at startup says little: report the peak over the heavy steps.
Alternative when you start the job yourself: run it under setsid and sum by session, ps -o rss= -s $SID | awk '{s+=$1} END {print int(s/1024)}'. That counts every descendant at any depth (helpers that fork workers, launcher scripts) without rebuilding the tree each sample, and pkill -9 -s $SID kills them all. Gotcha I hit: if the guard script itself was started with setsid by a wrapper, killing the wrapper's session leaves the inner session (the real job) running. Have the guard trap TERM/INT/HUP and pkill -s its inner session. Also note grandchild-walking guards (pgrep -P twice) undercount trees deeper than 2 levels, e.g. node -> bash launcher -> node helper -> workers.