This is the mail archive of the
libstdc++@gcc.gnu.org
mailing list for the libstdc++ project.
Re: Performance report (350MHz i686 configured as i386-unknown-freebsd4.8)
In article <20030624130604.767129a9.bkoz@nabi.net>,
"B. Kosnik" <bkoz@nabi.net> writes:
>> Run without -pthread:
>>
>> allocator.cc default 1241r 1234u 0s 0pf
>> allocator.cc malloc 2433r 2163u 2s 0pf
>> allocator.cc mt 1389r 1303u 0s 0pf
>> Run with -pthread (by manually hacking libstdc++-v3/scripts/check_performance):
>>
>> allocator.cc default 1190r 1141u 0s 0pf
>> allocator.cc malloc 426r 399u 1s 0pf
>> allocator.cc mt 480r 468u 0s 0pf
> About what I am seeing on linux.
> I don't know how to proceed on sorting out the default allocator issue.
> I like the suggestion here about switching over to mt_alloc if pthreads:
> http://gcc.gnu.org/ml/libstdc++/2003-06/msg00119.html
> That seems easiest to do, considering the alternatives.
Yikes!!!! To Carlo: It is the issue I suspected. However, please
note that we may not compare runs without -pthread against runs with
-pthread since allocator.cc (at least) uses a non-deterministic
pre-scaling routine. Humm... We may only observe that relatively
speaking, the situation flips when mutex overhead is introduced. We
shouldn't even compare that test one day to the next given its design.
>> (BTW, yes, I think we should run the test cases with and without
>> e.g. -pthread esp. if we find it matters even for non-threaded cases.
>> I'll make it possible to run either way without having to change a
>> srcdir file perhaps by respecting the same RUNTESTFLAG format that
>> dejagnu honors.)
> Sounds like a good idea.
OK, I will wack that out when time allows.