[Bug 1613996] Re: 30% slowdown in numerical solver exection on 16.04.1 vs. 14.04 with same solver binary

[Launchpad Bug Tracker]

Status changed to 'Confirmed' because the bug affects multiple users.

** Changed in: glibc (Ubuntu)
       Status: New => Confirmed

-- 
You received this bug notification because you are a member of Ubuntu
Foundations Bugs, which is subscribed to glibc in Ubuntu.
https://bugs.launchpad.net/bugs/1613996

Title:
  30% slowdown in numerical solver exection on 16.04.1 vs. 14.04 with
  same solver binary

Status in gcc:
  New
Status in glibc package in Ubuntu:
  Confirmed

Bug description:
  I noticed that a numerical solver I develop runs much slower on
  16.04.1 than on 14.04. See for example this output:

  The counters (top part of each result section) show that the solver
  does the same on both variants. The timings (lower part, beginning
  with FrameworkTimeWriteOutputs) are execution time in seconds. Overall
  time is in the last row (WallClockTime). First column shows results on
  Ubuntu 14.04, second column shows time on 16.04.1.

  ../../data/tests/CCMTest/Kirchhoff.d6p
                                       Reference             New
    IntegratorErrorTestFails                1026 ==         1026
    IntegratorFunctionEvals                32474 ==        32474
    IntegratorLESSetup                      3114 ==         3114
    IntegratorLESSolve                     32473 ==        32473
    IntegratorSteps                        25809 ==        25809
    LESJacEvals                              463 ==          463
    LESRHSEvals                             3241 ==         3241
    LESSetups                               3114 ==         3114
    --
    FrameworkTimeWriteOutputs               0.00 ~~         0.00
    IntegratorTimeFunctionEvals             4.96 <>         9.46
    IntegratorTimeLESSetup                  0.38 ~~         0.58
    IntegratorTimeLESSolve                  0.36 ~~         0.35
    LESTimeJacEvals                         0.08 ~~         0.08
    LESTimeRHSEvals                         0.27 ~~         0.46
    WallClockTime                           6.13 <>        10.79

  MoistField.d6o
  RHField.d6o
  ../../data/tests/EN15026/Kirchhoff.d6p
                                       Reference             New
    IntegratorErrorTestFails                   2 ==            2
    IntegratorFunctionEvals                17685 ==        17685
    IntegratorLESSetup                       903 ==          903
    IntegratorLESSolve                     17684 ==        17684
    IntegratorSteps                        17635 ==        17635
    LESJacEvals                              295 ==          295
    LESRHSEvals                             2065 ==         2065
    LESSetups                                903 ==          903
    --
    FrameworkTimeWriteOutputs               0.03 ~~         0.03
    IntegratorTimeFunctionEvals            31.04 <>        58.89
    IntegratorTimeLESSetup                  2.47 ~~         3.76
    IntegratorTimeLESSolve                  3.05 ~~         2.98
    LESTimeJacEvals                         0.28 ~~         0.28
    LESTimeRHSEvals                         2.02 ~~         3.30
    WallClockTime                          40.39 <>        69.39

  Particularly affected is the physics part of the code
  (IntegratorTimeFunctionEvals), which does by far the most memory
  access and uses pow(), sqrt(), exp() functions.

  The test code was compiled with GCC 4.8.4 on Ubuntu 14.04 and was run
  unmodified on 16.04 (after upgrade and on a second machine after a
  fresh install).

  When the code is compiled with the new GCC 5.4 on Ubuntu 16.04, the
  execution times are approximately the same as with GCC 4.8.4 on Ubuntu
  16.04. Therefore I would not think it is a GCC bug.

  I have a test suite archive for download and test execution prepared:

  http://bauklimatik-dresden.de/downloads/tmp/test_suite.tar.7z

  Run the test suite on 14.04 and on 16.04 and observe the numbers in
  the "New" column, they will differ significantly for most test cases.

  Can you confirm my observation? And if yes, does anyone know how to
  avoid this performance drop?

To manage notifications about this bug go to:
https://bugs.launchpad.net/gcc/+bug/1613996/+subscriptions