[PATCH 0/9] malloc: Remove fastbins

[Florian Weimer]

* DJ Delorie:

> Florian Weimer <fweimer@redhat.com> writes:
>> * DJ Delorie:
>>
>>> So I ran some big benchmarks over the weekend to see what they said...
>>> I tested three source trees: unpatched, patched with the N/9 "remove
>>> fastbins" patch set, and one with that plus Florians's proposed
>>> batch-free patch.
>>
>> Did you apply the batched malloc part as well?
>
> I applied whatever was in your email ;-)
>
>>> I admit the machine running the test was not as "quiet" as it could have
>>> been, but I had plenty of ram and cores to avoid artificial slowdowns.
>>> Plus, I was looking for big changes, not small ones.
>>>
>>> I used two tests: the first is the original trace simulator I used when
>>> benchmarking tcache, along with the workloads used then.  The raw
>>> numbers are total CPU cycles; smaller is better.
>>
>> Are those retired cycles?  How are these measured?  For multi-thread
>> programs with some kernel overhead, using wall clock time is probably
>> better.
>
> IIRC it was CPU cycle clock time.  It's trace_run.c from:
> https://pagure.io/glibc-malloc-trace-utils/

I'm not sure if this is a good indicator, sorry.  Especially when
running under virtualization.

Thanks,
Florian