Extra swapping when using gfortran 4.3-20061224 vs gfortran 4.2-20061211

[Toon Moene]

I wrote, in December 2006:

> A run of HIRLAM for which the grid was carefully tuned (half a year ago) 
> to the RAM size of my laptop suddenly started to swap two weeks ago when 
> using gfortran 4.3 as the compiler.
> 
> However, as the HIRLAM suite I'm testing with also constantly changes, I 
> first suspected our own source code.  Today I finished the back-to-back 
> comparison of gfortran 4.3-20061224 vs. gfortran 4.2-20061211. Compiling 
> all the source with the first (experimental) compiler makes HIRLAM swap 
> during execution, whereas with the second (prerelease) compiler, it 
> doesn't.

While I still believe this experiment was valid and produced the results 
I describe here, I now have proof that this extra swapping is not caused 
by the MEM-SSA changes.

At a request by Diego Novillo, I ran two versions of HIRLAM (7.0, which 
dates from May last year and 7.1beta2, which dates from two weeks ago) 
using both the Fortran/C compiler from trunk revision 119759 (just 
before the MEM-SSA changes went in) and revision 119760 (with those 
changes).

The result is: No swapping for the 7.0 version of HIRLAM *with both 
compilers*:

119759's top log, about half way through the forecast:

30070 hirlam    25   0  449m 333m 2144 R 99.9 66.8  16:00.26 hlprog.x 

30070 hirlam    25   0  448m 333m 2144 R 99.9 66.7  16:03.27 hlprog.x 

30070 hirlam    25   0  486m 359m 2144 R 99.9 71.9  16:06.27 hlprog.x 

30070 hirlam    25   0  448m 296m 2144 R 99.9 59.3  16:09.27 hlprog.x 

30070 hirlam    25   0  441m 322m 2144 R 99.9 64.6  16:12.29 hlprog.x 

30070 hirlam    25   0  448m 333m 2144 R 99.9 66.7  16:15.29 hlprog.x 

30070 hirlam    25   0  449m 333m 2144 R 99.9 66.8  16:18.29 hlprog.x 

30070 hirlam    25   0  448m 332m 2144 R 99.9 66.6  16:21.29 hlprog.x 

30070 hirlam    25   0  486m 376m 2144 R 99.9 75.5  16:24.30 hlprog.x 


119760's top log:

30347 hirlam    25   0  448m 329m 1844 R 99.9 66.0  16:01.05 hlprog.x 

30347 hirlam    25   0  486m 381m 1844 R 99.9 76.3  16:04.05 hlprog.x 

30347 hirlam    25   0  456m 304m 1844 R 99.9 61.0  16:07.06 hlprog.x 

30347 hirlam    25   0  486m 344m 1844 R 99.9 69.0  16:10.06 hlprog.x 

30347 hirlam    25   0  449m 330m 1844 R 99.9 66.1  16:13.06 hlprog.x 

30347 hirlam    25   0  448m 329m 1844 R 99.9 66.0  16:16.07 hlprog.x 

30347 hirlam    25   0  448m 330m 1844 R 99.9 66.1  16:19.07 hlprog.x 

30347 hirlam    25   0  486m 374m 1844 R 99.9 74.9  16:22.07 hlprog.x 

30347 hirlam    25   0  448m 302m 1844 R 99.9 60.6  16:25.07 hlprog.x 

30347 hirlam    25   0  486m 348m 1844 R 99.9 69.7  16:28.08 hlprog.x 


Note that for both compilers, the maximum Resident Size is ~ 375 Mbyte, 
and both executables ran 99.9 % of CPU usage.

The picture is completely different for the HIRLAM version 7.1beta2:

119759's top log:

25770 hirlam    18   0  543m 424m 1800 D 19.6 85.0  16:00.28 hlprog.x 

25770 hirlam    18   0  543m 424m 1800 D 48.3 85.0  16:01.73 hlprog.x 

25770 hirlam    18   0  544m 425m 1800 R 65.2 85.2  16:03.69 hlprog.x 

25770 hirlam    18   0  544m 426m 1800 R 71.2 85.3  16:05.83 hlprog.x 

25770 hirlam    18   0  550m 447m 1820 D 72.6 89.7  16:08.01 hlprog.x 

25770 hirlam    18   0  550m 454m 1824 D  6.7 91.0  16:08.21 hlprog.x 

25770 hirlam    18   0  547m 448m 1824 D 13.6 89.9  16:08.62 hlprog.x 

25770 hirlam    18   0  537m 444m 1824 D 10.0 89.0  16:08.92 hlprog.x 

25770 hirlam    18   0  477m 386m 1824 D  6.7 77.4  16:09.12 hlprog.x 


119760's top log:

23023 hirlam    18   0  550m 434m 1520 D  4.3 86.9  16:00.02 hlprog.x 

23023 hirlam    18   0  546m 437m 1520 R 17.0 87.5  16:00.53 hlprog.x 

23023 hirlam    18   0  537m 436m 1520 R  9.3 87.4  16:00.81 hlprog.x 

23023 hirlam    18   0  477m 379m 1524 R  6.7 76.0  16:01.01 hlprog.x 

23023 hirlam    18   0  477m 382m 1656 R  2.0 76.6  16:01.07 hlprog.x 

23023 hirlam    18   0  537m 393m 1672 R 28.3 78.7  16:01.92 hlprog.x 

23023 hirlam    19   0  543m 409m 1672 D 60.6 81.9  16:03.74 hlprog.x 

23023 hirlam    19   0  584m 449m 1672 R 78.6 89.9  16:06.10 hlprog.x 

23023 hirlam    18   0  584m 453m 1672 D 33.3 90.8  16:07.10 hlprog.x 


Note that the Resident Size maxes out at slightly over 450 Mbyte,  which 
is consistent with a RAM size of 511 Mbyte - you simply can't get much 
more for a single executable on this system with X and scores of daemons 
running.  The swapping is evident from the 'D' process state (waiting 
for disk) and the less than 99.9 % CPU utilization.

So the conclusion is: MEM-SSA doesn't lead to memory bloat.  The test 
results from 26 December last year must point to *another* difference 
between the tested compilers (gfortran 4.3-20061224 vs. gfortran 
4.2-20061211).

Cheers,

-- 
Toon Moene - e-mail: toon@moene.indiv.nluug.nl - phone: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
A maintainer of GNU Fortran: http://gcc.gnu.org/fortran/
Who's working on GNU Fortran: 
http://gcc.gnu.org/ml/gcc/2007-01/msg00059.html