This is the mail archive of the libc-help@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: About JSON output generated by "bench-malloc-thread"


On 12 February 2018 at 01:25, Francesco <francesco.montorsi@gmail.com> wrote:
> Hi Siddesh,
> You're right, I have to sign another copyright assignment. I will eventually
> do that in the future but perhaps it does not make sense to do it now for
> such a small contribution.

No worries, I'll fix it up when I have some time so that it's usable
for your tests.

> However that didn't stop me from approaching in a slightly different manner
> my benchmarking:
> I wrote a small Makefile and a couple of python scripts and uploaded them
> and the benchmark I collected on 3 different machines here:
>    https://github.com/f18m/malloc-benchmarks
> The test compares GNU libc implementation, tcmalloc and jemalloc.
> Apparently tcmalloc is the winner on the first 2 systems, having 4 and 8
> cores.
> GNU libc 2.26 in those 2 systems was much faster than earlier GNU libc
> implementations and, although slower than tcmalloc, did good.
>
> This 3rd machine is a dual-CPU server having a total of 40 cores and
> unfortunately it runs Centos7 and I could not build GNU libc 2.26 there. The
> system GNU libc there (2.17) was pretty bad w.r.t. performances if compared
> against jemalloc!
>
>
> Hope these results may turn useful to others...

This is very good, thank you for reporting the results.  glibc 2.26
got a thread cache for malloc, which is why you see those
improvements.

It would be really nice to have the ability to hook alternative
implementations like tcmalloc and jemalloc using tunables and not have
to rely on the function interposition semantics.  That will make
implementations like tcmalloc and jemalloc more accessible to
applications.

Siddhesh


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]