Hi Dinko, On Mon, Oct 13, 2025 at 11:04:04AM +0200, Dinko Korunic wrote: > > > > Maybe we should start a discussion on the github pages about allocators, > > how to test them and share feedback. That could possibly encourage some > > users to give them a try on a less busy day. > > > [...] > > Hi, > > I've been testing mimalloc for a while now especially due to the fact that > jemalloc repository has been archived since June 2025. I have found the > performance in real life and in specific scenarios common to installations I > have lying around is not statistically much different from jemalloc but only > when additional heap-oriented security features (guard pages, randomized > allocations, encrypted free lists) have been disabled. > Also I have found maximum RSS in containers with mimalloc to be slightly > lower than maximum RSS with jemalloc, but again I suspect that it is very > very dependent on a specific traffic load and active configuration.
Interesting, thanks for sharing your observations! While jemalloc has been working extremely well for us for a long time, recently it started to batch too many objects upon some free() operations, considerably slowing down certain multi-threaded workloads to the point of almost deadlocking threads and triggering the watchdog. We've tried to play with the tuning a bit but it's a pain to deal with and we were not completely convinced about the results. Interestingly a few times it turned out that glibc's simpler ptmalloc regained interest by being more predicable, and I'm aware that a number of users turned back to plain glibc instead of jemalloc. Intuitively I'd think that an allocator that manages to reduce the RSS a bit would more aggressively release objects and would probably not cause large batches, which could be encouraging. Another point is, is it sufficient to just pass ADDLIB=-lmimalloc or is anything else required ? If the former, maybe we could at least mention it in the build instructions to encourage users to try it so that we can progressively get good feedback about it in various workloads. That's already how jemalloc became popular in our context. Cheers, willy
