You are viewing a plain text version of this content. The canonical link for it is here.
Posted to mapreduce-dev@hadoop.apache.org by "Alan Burlison (JIRA)" <ji...@apache.org> on 2015/06/08 14:55:01 UTC

[jira] [Created] (MAPREDUCE-6391) util/Timer.cc completely misunderstands _POSIX_CPUTIME

Alan Burlison created MAPREDUCE-6391:
----------------------------------------

             Summary: util/Timer.cc completely misunderstands _POSIX_CPUTIME
                 Key: MAPREDUCE-6391
                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-6391
             Project: Hadoop Map/Reduce
          Issue Type: Sub-task
    Affects Versions: 2.7.0
            Reporter: Alan Burlison


util/Timer.cc contains the following code:

clock_gettime(_POSIX_CPUTIME, &ts);

I'm not exactly sure what sort of time value this code is attempting to obtain (the process's on-CPU time perhaps?) but it is completely wrong, even on Linux. _POSIX_CPUTIME is a feature test macro and is *not* a valid parameter to clock_gettime. From the clock_gettime manpage:

bq. The symbols _POSIX_MONOTONIC_CLOCK, _POSIX_CPUTIME, _POSIX_THREAD_CPUTIME indicate that CLOCK_MONOTONIC, CLOCK_PROCESS_CPUTIME_ID, CLOCK_THREAD_CPUTIME_ID are available

On  my Linux distro the value of _POSIX_CPUTIME is defined as 0. That corresponds to

bq.  #define __CLOCK_REALTIME0       0       /* obsolete; same as CLOCK_REALTIME */

And CLOCK_REALTIME is

bq. #define CLOCK_REALTIME          3       /* wall clock */

Which I suspect is not what is required here.

Getting cross-platform CPU time is fiddly, the following blog post has a good summary: http://nadeausoftware.com/articles/2012/03/c_c_tip_how_measure_cpu_time_benchmarking (code examples are CC licensed so should be no taint issues)




--
This message was sent by Atlassian JIRA
(v6.3.4#6332)