| Message ID | yddei8aoxn9.fsf@manam.CeBiTec.Uni-Bielefeld.DE |
|---|---|
| State | New |
| Headers | show |
Rainer Orth wrote: > I see two fortran testcases that regularly time out on my platforms: > > * gfortran.dg/array_constructor_33.f90 always times out on > Solaris/SPARC: even on an unloaded machine, it takes over 3 minutes: > real 3:05.04 SPARC Enterprise T5220, 1.2 GHz UltraSPARC-T2 > I'd like to set a timeout factor of 4 to allow for that. > > * gfortran.dg/vect/fast-math-pr38968.f90 times out in a very strange > way. I only observe the timeouts on Solaris 10/x86, but as I've > found, it has nothing to do with Solaris 10: identical binaries and > libraries take 1:30 min on the test machine (the first one), which is > a factor of 3 slower than two other machines (either Solaris 10 or > 11), but with different CPUs: > > real 1:31.91 Sun Fire X4450, 2.93 GHz Intel Xeon X7350 > real 32.11 Sun Fire X4200, 2.6 Ghz AMD Opteron 285 > real 33.46 Sun Fire V60x, 2.8 GHz Intel Xeon > > Two other machines running Solaris 11 resp. 9 are in between: > > real 47.43 Sun Fire X4600, 2.6 GHz AMD Opteron 8218 > real 50.08 Sun Fire V20z, 1.8 GHz AMD Opteron 244 > > So this seems to be a problem with this particular CPU; still I'd like > to accomodate this range by setting a timeout factor of 2. > Only mainline is affected. That's really odd. > Ok for mainline (and 4.5 branch for the first)? OK. Thanks for the effort to reduce the failures. Tobias
Tobias Burnus <burnus@net-b.de> writes: > Rainer Orth wrote: >> I see two fortran testcases that regularly time out on my platforms: >> >> * gfortran.dg/array_constructor_33.f90 always times out on >> Solaris/SPARC: even on an unloaded machine, it takes over 3 minutes: >> real 3:05.04 SPARC Enterprise T5220, 1.2 GHz UltraSPARC-T2 >> I'd like to set a timeout factor of 4 to allow for that. >> >> * gfortran.dg/vect/fast-math-pr38968.f90 times out in a very strange >> way. I only observe the timeouts on Solaris 10/x86, but as I've >> found, it has nothing to do with Solaris 10: identical binaries and >> libraries take 1:30 min on the test machine (the first one), which is >> a factor of 3 slower than two other machines (either Solaris 10 or >> 11), but with different CPUs: >> >> real 1:31.91 Sun Fire X4450, 2.93 GHz Intel Xeon X7350 >> real 32.11 Sun Fire X4200, 2.6 Ghz AMD Opteron 285 >> real 33.46 Sun Fire V60x, 2.8 GHz Intel Xeon >> >> Two other machines running Solaris 11 resp. 9 are in between: >> >> real 47.43 Sun Fire X4600, 2.6 GHz AMD Opteron 8218 >> real 50.08 Sun Fire V20z, 1.8 GHz AMD Opteron 244 >> >> So this seems to be a problem with this particular CPU; still I'd like >> to accomodate this range by setting a timeout factor of 2. >> Only mainline is affected. > > That's really odd. indeed. Perhaps Uros (or another of the x86 maintainers) can shed some light? Rainer
Index: gcc/testsuite/gfortran.dg/array_constructor_33.f90 =================================================================== --- gcc/testsuite/gfortran.dg/array_constructor_33.f90 (revision 168958) +++ gcc/testsuite/gfortran.dg/array_constructor_33.f90 (working copy) @@ -1,4 +1,5 @@ ! { dg-do compile } +! { dg-timeout-factor 4 } ! PR20923 gfortran slow for large array constructors. ! Test case prepared from PR by Jerry DeLisle <jvdelisle@gcc.gnu.org> program sel Index: gcc/testsuite/gfortran.dg/vect/fast-math-pr38968.f90 =================================================================== --- gcc/testsuite/gfortran.dg/vect/fast-math-pr38968.f90 (revision 168958) +++ gcc/testsuite/gfortran.dg/vect/fast-math-pr38968.f90 (working copy) @@ -2,6 +2,8 @@ ! to handle complex types. This test is very slow on these platforms so ! skipping is better then running it unvectorized. ! { dg-skip-if "" { ia64-*-* sparc*-*-* } { "*" } { "" } } +! It can be slow on some x86 CPUs. +! { dg-timeout-factor 2 } program mymatmul implicit none integer, parameter :: kp = 4
I see two fortran testcases that regularly time out on my platforms: * gfortran.dg/array_constructor_33.f90 always times out on Solaris/SPARC: even on an unloaded machine, it takes over 3 minutes: real 3:05.04 SPARC Enterprise T5220, 1.2 GHz UltraSPARC-T2 I'd like to set a timeout factor of 4 to allow for that. I'd like to apply this patch to both mainline and the 4.5 branch which are both affected. * gfortran.dg/vect/fast-math-pr38968.f90 times out in a very strange way. I only observe the timeouts on Solaris 10/x86, but as I've found, it has nothing to do with Solaris 10: identical binaries and libraries take 1:30 min on the test machine (the first one), which is a factor of 3 slower than two other machines (either Solaris 10 or 11), but with different CPUs: real 1:31.91 Sun Fire X4450, 2.93 GHz Intel Xeon X7350 real 32.11 Sun Fire X4200, 2.6 Ghz AMD Opteron 285 real 33.46 Sun Fire V60x, 2.8 GHz Intel Xeon Two other machines running Solaris 11 resp. 9 are in between: real 47.43 Sun Fire X4600, 2.6 GHz AMD Opteron 8218 real 50.08 Sun Fire V20z, 1.8 GHz AMD Opteron 244 So this seems to be a problem with this particular CPU; still I'd like to accomodate this range by setting a timeout factor of 2. Only mainline is affected. Ok for mainline (and 4.5 branch for the first)? Thanks. Rainer 2011-01-18 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE> * gfortran.dg/array_constructor_33.f90: Use dg-timeout-factor 4. PR fortran/45988 * gfortran.dg/vect/fast-math-pr38968.f90: Use dg-timeout-factor 2.