5 ms·

https://github.com/microsoft/mimalloc?tab=readme-ov-file#performance https://github.com/microsoft/mimalloc?tab=readme-ov-file#per...

by scns 2y ago

Y_Y 2y ago
Thanks for that. I have an innate skepticism of benchmarks from a company who wants to show their solution is best, and it seems their latest results are from 2021, but I couldn't find a better comparison myself either. I do note that rpmalloc (old), Hermes (not public? doesn't compare with mimalloc) and also snmalloc (also Microsoft) have benchmarks of their own showing themselves to be best in some circumstances. https://github.com/mjansson/rpmalloc-benchmark https://github.com/mjansson/rpmalloc-benchmark https://arxiv.org/abs/2109.02922 https://arxiv.org/abs/2109.02922 https://github.com/SchrodingerZhu/bench_suite/blob/master/output/index.md https://github.com/SchrodingerZhu/bench_suite/blob/master/ou...
- Tuna-Fish 2y ago
  More than that, based on the basic architecture of JEmalloc, mimalloc and tcmalloc, you should always expect their fragmentation behavior to be better for long-running software than the glibc malloc. (At the expense of consuming substantially more memory for very small programs with only a few allocations of a given size). The glibc malloc has nearly pessimal fragmentation behavior, you are very confused here.
  - Y_Y 2y ago
    > you are very confused here Is this addressed to me? If so would you do me the kindness of attempting to disabuse me of my confusion?
    - Tuna-Fish 2y ago
      Wait, no, sorry. That reply was supposed to go to a different post.
jeffbee 2y ago
Oddly enough I've written about those results before: https://www.dropbox.com/scl/fi/evnn6yoornh9p6l7nq1t9/Irreproducible-malloc-benchmarks.paper?rlkey=73mi4op4hfn7euqmry9bmnxtr&dl=0 https://www.dropbox.com/scl/fi/evnn6yoornh9p6l7nq1t9/Irrepro...
- vlovich123 2y ago
  but we all can agree that glibc is trash right? :) I’d be careful extrapolating from just one benchmark but generally if I had to choose I’d pick the new tcmalloc if I could. It seems to be a higher quality codebase.
  - jeffbee 2y ago
    The problem with the current tcmalloc, which I agree is very good, is the difficulty of integrating it with random builds. It works well for large systems and especially those that are already bazelized, but I couldn't whip up a quick way to link it to `jq` in this gist.

https://github.com/microsoft/mimalloc?tab=readme-ov-file#performance https://github.com/microsoft/mimalloc?tab=readme-ov-file#per...

by scns 2y ago

https://github.com/microsoft/mimalloc?tab=readme-ov-file#performance https://github.com/microsoft/mimalloc?tab=readme-ov-file#per...

Y_Y 2y ago

Thanks for that. I have an innate skepticism of benchmarks from a company who wants to show their solution is best, and it seems their latest results are from 2021, but I couldn't find a better comparison myself either. I do note that rpmalloc (old), Hermes (not public? doesn't compare with mimalloc) and also snmalloc (also Microsoft) have benchmarks of their own showing themselves to be best in some circumstances. https://github.com/mjansson/rpmalloc-benchmark https://github.com/mjansson/rpmalloc-benchmark https://arxiv.org/abs/2109.02922 https://arxiv.org/abs/2109.02922 https://github.com/SchrodingerZhu/bench_suite/blob/master/output/index.md https://github.com/SchrodingerZhu/bench_suite/blob/master/ou...

Tuna-Fish 2y ago
More than that, based on the basic architecture of JEmalloc, mimalloc and tcmalloc, you should always expect their fragmentation behavior to be better for long-running software than the glibc malloc. (At the expense of consuming substantially more memory for very small programs with only a few allocations of a given size). The glibc malloc has nearly pessimal fragmentation behavior, you are very confused here.
- Y_Y 2y ago
  > you are very confused here Is this addressed to me? If so would you do me the kindness of attempting to disabuse me of my confusion?
  - Tuna-Fish 2y ago
    Wait, no, sorry. That reply was supposed to go to a different post.

jeffbee 2y ago

Oddly enough I've written about those results before: https://www.dropbox.com/scl/fi/evnn6yoornh9p6l7nq1t9/Irreproducible-malloc-benchmarks.paper?rlkey=73mi4op4hfn7euqmry9bmnxtr&dl=0 https://www.dropbox.com/scl/fi/evnn6yoornh9p6l7nq1t9/Irrepro...

vlovich123 2y ago
but we all can agree that glibc is trash right? :) I’d be careful extrapolating from just one benchmark but generally if I had to choose I’d pick the new tcmalloc if I could. It seems to be a higher quality codebase.
- jeffbee 2y ago
  The problem with the current tcmalloc, which I agree is very good, is the difficulty of integrating it with random builds. It works well for large systems and especially those that are already bazelized, but I couldn't whip up a quick way to link it to `jq` in this gist.