[Sending to dev after accidentally replying just to Liviu]

On 09/30/2012 12:09 PM, Martin Sebor wrote:
On 09/27/2012 06:36 PM, Liviu Nicoara wrote:
On 9/27/12 8:27 PM, Martin Sebor wrote:
On 09/27/2012 06:41 AM, Liviu Nicoara wrote:
On 09/26/12 20:12, Liviu Nicoara wrote:
I have created STDCXX-1071 and linked to STDCXX-1056. [...]

I am open to all questions, the more the better. Most of my opinions
have been expressed earlier, but please ask if you want to know more.

I am attaching here the proposed (4.3.x) patch and the timings results
(after re-verifying the correctness of the timing program and the
results). The 4.2.x patch, the 4.3.x patch, the test program and the
results file are also attached to the incident.

The patch isn't binary compatible. We can't remove data members
in a minor release. We could only do it in a major release.

There are two of them, one for 4.2.x, one for 4.3.x.

I'm still curious about the performance, though. It doesn't make
sense to me that a call to a virtual function is faster than one
to an almost trivial inline function.

I scratched my head long on that. I wager that it depends on the
machine, too.

Here are my timings for library-reduction.cpp when compiled
GCC 4.5.3 on Solaris 10 (4 SPARCV9 CPUs). I had to make a small
number of trivial changes to get it to compile:

With cache No cache
real 1m38.332s 8m58.568s
user 6m30.244s 34m25.942s
sys 0m0.060s 0m3.922s

I also experimented with the program on Linux (CEL 4 with 16
CPUs). Initially, I saw no differences between the two versions.
So I modified it a bit to make it closer to the library (the
modified program is attached). With those changes the timings
are below:

With cache No cache
real 0m 1.107s 0m 5.669s
user 0m17.204s 0m 5.669s
sys 0m 0.000s 0m22.347s

I also recompiled and re-ran the test on Solaris. To speed
things along, I set the number threads and loops to 8 and
1000000. The numbers are as follows:

With cache No cache
real 0m3.341s 0m26.333s
user 0m13.052s 1m37.470s
sys 0m0.009s 0m0.132s

The numbers match my expectation. The overhead without the
"numpunct cache" is considerable.

Somewhat unexpectedly, the test with the cache didn't crash.

I also tried timing the numpunct patch attached to the issue
with the test program. The test program runs to completion
without the patch but it crashes with a SIGSEGV after the
patch is applied. That suggests there's a bug somewhere in
do_thousands_sep() (in addition to the caching).


Let me know if there is anything I could help with.


Reply via email to