This is the mail archive of the
libstdc++@gcc.gnu.org
mailing list for the libstdc++ project.
Re: Diff mt_allocator
- From: Loren James Rittle <rittle at latour dot rsch dot comm dot mot dot com>
- To: libstdc++ at gcc dot gnu dot org
- Cc: stefan at xapa dot se
- Date: Thu, 22 Jan 2004 21:12:17 -0600 (CST)
- Subject: Re: Diff mt_allocator
- Organization: Networks and Infrastructure Lab (IL02/2240), Motorola Labs
> here's the diff that allocates chunks of memory directly to threads
> according to rfc.
> I have tested it quite a bit (including recompiling a 10k line MT
> application and running that) but I appriciate all reviewing and testing
> - it's been a long day =)
Stefan,
Interesting patches (please consider using '-c' in the future; and
please avoid WS changes). Please consider writing the changelog entry
(since we'd rather use your words than mine). Logically, I see no
downside to these patches but haven't tested them beyond...
> Here's the before numbers (on the dual 2.4 Ghz machine):
> real 0m18.026s
> user 0m12.620s
> sys 0m13.480s
Any idea why you saw so much sys time here? Is that the global
contention factor that you have removed with the second patch?
> And this is after this diff:
> real 0m4.123s
> user 0m13.740s
> sys 0m1.460s
Using the included test case, I see the following on a single PII/350
running i386-*-freebsd4.9:
-r1.9 (-O2 -g -pthread):
30r 27.0u 1.8s ./a.out
w/ first patch only (to address single-threaded speedup):
30r 27.2u 1.7s ./a.out
w/ both patches:
15r 14.4u 1.2s ./a.out
Nice to see my ~5 year old HW can keep up... (ignore wall time please ;-)
Regards,
Loren