On Mon, 2025-06-09 at 13:57 -0700, Andrii Nakryiko wrote: > On Sat, Jun 7, 2025 at 1:13 AM Eduard Zingerman <eddyz87@xxxxxxxxx> > wrote: [...] > > Given above I'm inclined to stick with "many cg" approach, as it > > has > > less jitter and is reasonably performant. I need to wrap-up > > parallel > > veristat version anyway (and many cg should be easier to manage for > > parallel run). > > As I mentioned in offline discussion, this 256KB jitter seems minor > and is not on the order of "interesting memory usage", IMO. So I'd go > with a single cgroup approach with reset and keep runtime quite > significantly faster. Mykyta is working on a series to further speed > up veristat's mass verification by relying on bpf_object__prepare() > and bpf_prog_load() for each program, instead of re-opening and > re-parsing the same object file all over again. So it would be good > to > not add extra 18 seconds of runtime just for creation of cgroups. > > FWIW, we can count mem_peak in megabytes and the jitter will be gone > (and we'll probably still get all the useful signal we wanted with > this measurement), if that's the concern. Ok, makes sense, I'll rename the field to mem_peak_mb and go with a single cgroup. Metric would be zero for small programs but that should be fine. [...]
