Re: [PATCH v2][malloc] Use relaxed atomics for malloc have_fastchunks

[DJ Delorie <dj at redhat dot com>]

Wilco Dijkstra <Wilco.Dijkstra@arm.com> writes:
>> Workload                      Pristine         Patched      
>> 389ds                     9,121,687,695    8,017,021,813   87.89%
>> dj2                       7,901,004,232    8,277,784,940  104.77%
>> ...
>> okular-1                  3,648,656,309    3,220,751,900   88.27%
>> oocalc                    1,053,984,703    1,009,859,213   95.81%
>> qemu-virtio                 781,260,028      766,458,246   98.11%
>> qemu-win7                   655,497,193      626,270,566   95.54%
>> proprietary-2             2,112,159,165    1,977,684,058   93.63%

> Btw have you tried running these traces with say:
> export GLIBC_TUNABLES=glibc.malloc.tcache_count=100

389ds         7,381,521,147 7,318,122,451
dj2           7,385,497,424 7,271,176,052
...
okular-1      3,658,050,304 3,590,747,773 
oocalc        1,151,498,086 1,088,218,893
qemu-virtio     788,062,420   732,039,964
qemu-win7       686,729,339   710,123,160 
proprietary-2 2,192,659,157 2,069,942,722

So, helps some, hurts others.  The default is a compromise, and this is
why it's a tunable :-)

Plus, these results don't show how much memory (RSS) is used by the
cache for each app.  The default is intentionally on the low side to
minimize RSS use.

(the usual caveat about benchmark results apply; I don't have a
dedicated machine to run these on, so "results vary, sometimes a lot" ;)

> It would be interesting to find out whether that (or even larger values)
> helps your traces too like it does the benchmarks I've tried.

If you could capture traces from those benchmarks, I'd appreciate
getting a copy so I can add them to my corpus.  Esp if it's a benchmark
of a real-world app :-)