Re: [PATCH] RE: gcc parallel make check

[Jakub Jelinek <jakub at redhat dot com>]

On Tue, Sep 09, 2014 at 03:14:08PM +0000, VandeVondele  Joost wrote:
> Attached is a further revision of the patch, now dealing with check-c++. Roughly 50% speedup here at '-j32' (18m vs 12m). For my setup (--enable-languages=c,c++,fortran) I have now improved all targets called in 'make -j32 -k check'. The latter is now 30% faster (15m vs 20m). Note that there are +- 1m fluctuations in these numbers, easily.
> 
> I currently have no plans to work on other check targets before this patch is committed.
> 
> OK for trunk ?

So, I've tested your patch on a few years old 16way AMD,
--enable-checking=yes and with make check -j1 and make check -j2
(the former doesn't parallelize, the latter does).

First of all, the -j2 testing shows more tests tested in gcc and libstdc++:

                === gcc Summary ===
 
-# of expected passes           108100
+# of expected passes           108130
 # of unexpected failures       175
 # of unexpected successes      33
 # of expected failures         262
...
                === libstdc++ Summary ===
 
-# of expected passes           10133
+# of expected passes           10152

In gcc testsuite the difference seems to be (according to sorted PASS
lines from *.sum files):
+PASS: gcc.target/i386/vect-args.c -m3dnow -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -m3dnow -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -m3dnow -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -m3dnow  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -m3dnow  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -m3dnow  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -mmmx -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -mmmx -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -mmmx -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -mmmx  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -mmmx  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -mmmx  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse2 -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse2 -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse2 -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse2  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse2  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse2  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c -msse  (test for excess errors)
+PASS: gcc.target/i386/vect-args.c  -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c  -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c  -O (test for excess errors)
+PASS: gcc.target/i386/vect-args.c   (test for excess errors)
+PASS: gcc.target/i386/vect-args.c   (test for excess errors)
+PASS: gcc.target/i386/vect-args.c   (test for excess errors)
and in libstdc++:
+PASS: 23_containers/map/modifiers/erase/abi_tag.cc scan-assembler _ZNSt3mapIiiSt4lessIiESaISt4pairIKiiEEE5eraseB5cxx11ESt17_Rb_tree_iteratorIS4_E
+PASS: 23_containers/map/modifiers/erase/abi_tag.cc (test for excess errors)
+PASS: 23_containers/multimap/modifiers/erase/abi_tag.cc scan-assembler _ZNSt8multimapIiiSt4lessIiESaISt4pairIKiiEEE5eraseB5cxx11ESt17_Rb_tree_iteratorIS4_E
+PASS: 23_containers/multimap/modifiers/erase/abi_tag.cc (test for excess errors)
+PASS: 23_containers/multiset/modifiers/erase/abi_tag.cc scan-assembler _ZNSt8multisetIiSt4lessIiESaIiEE5eraseB5cxx11ESt23_Rb_tree_const_iteratorIiE
+PASS: 23_containers/multiset/modifiers/erase/abi_tag.cc scan-assembler _ZNSt8multisetIiSt4lessIiESaIiEE5eraseB5cxx11ESt23_Rb_tree_const_iteratorIiES5_
+PASS: 23_containers/multiset/modifiers/erase/abi_tag.cc (test for excess errors)
+PASS: 23_containers/set/modifiers/erase/abi_tag.cc scan-assembler _ZNSt3setIiSt4lessIiESaIiEE5eraseB5cxx11ESt23_Rb_tree_const_iteratorIiE
+PASS: 23_containers/set/modifiers/erase/abi_tag.cc scan-assembler _ZNSt3setIiSt4lessIiESaIiEE5eraseB5cxx11ESt23_Rb_tree_const_iteratorIiES5_
+PASS: 23_containers/set/modifiers/erase/abi_tag.cc (test for excess errors)
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexIdE4imagB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexIdE4realB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexI[eg]E4imagB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexI[eg]E4realB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexIfE4imagB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexIfE4realB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexIiE4imagB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc scan-assembler _ZNKSt7complexIiE4realB5cxx11Ev
+PASS: 26_numerics/complex/abi_tag.cc (test for excess errors)

Not sure where the bug is, could be e.g. in i386.exp for gcc, but for
libstdc++ less likely to be there rather than in the split.

Attached are seconds result of the various *.log.sep times, in the first
attachment from my last night's stock make -j48 -k check regtest (so, note
the times might be higher because most of the CPUs are kept much more busy)
and for make -j2 -k check regtest with your patch in the second attachment.
The normal{1,2,3} libstdc++ times are still way too high (bet one of those 3
is regex, which really has to be split), while e.g. the various 0/3 or even
< 30 seconds jobs are IMHO undesirable.

	Jakub

Attachment: 1
Description: Text document

Attachment: 2
Description: Text document