lkml.org 
[lkml]   [2010]   [Dec]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch v2 4/4] taskstats: Export "cdata_wait" CPU times with taskstats
* Michael Holzheu <holzheu@linux.vnet.ibm.com> [2010-12-06 13:37:37]:

> Hello Balbir,
>
> On Fri, 2010-12-03 at 13:03 +0530, Balbir Singh wrote:
> > * Oleg Nesterov <oleg@redhat.com> [2010-12-01 19:51:28]:
> >
> > > On 11/29, Michael Holzheu wrote:
> [snip]
> > > Say, taskstats_exit()->fill_stats()->bacct_add_tsk(). Every thread
> > > does this, including the group_leader. But, it is possible that
> > > group_leader exits first, before other threads. IOW, what
> > > stats->ac_cXtime actually mean?
> > >
> >
> > stats->ac_* time was designed only for tgid's to begin with,
>
> You mean for tids (threads/tasks), no? stats->ac_* time is only reported
> for threads in bacct_add_tsk() and not for tgids.

Sorry, I meant tid's or pid

>
> > so I am
> > not sure if ac_cXtime makes sense for threads
>
> I would suggest to do it the same way as /proc/<tgid>/tasks/<tid>/stat.
> It reports the (same) cumulative time for each thread. See
> do_task_stat() in fs/proc/array.c. So IMHO also for taskstats it makes
> sense to include the cXtime for all threads and not only for the thread
> group leader.
>
> Also I would include tgid in taskstats so that userspace can group the
> tasks into thread groups.
>
> I am not sure regarding the aggregation for tgids in
> TASKSTATS_CMD_ATTR_TGID and exit events with group_dead=1. Currently
> only the delay accounting numbers are aggregated. If we would do it
> like /proc/<tgid>/stat (do_task_stat() with whole=1) we also could
> aggregate also the other values e.g. the CPU time.
>
> I think the following tgid data would make sense to be returned for
> TASKSTATS_CMD_ATTR_TGID and exit events with group_dead=1:
>
> bacct_add_tsk():
> ----------------
> ac_pid thread group leader (tsk->tgid)
> ac_etime thread group leader
> ac_btime thread group leader
> ac_nice thread group leader
> ac_sched thread group leader
> ac_uid thread group leader
> ac_gid thread group leader
> ac_ppid thread group leader
> ac_comm thread group leader
> ac_exit_code thread group leader
> ac_flags thread group leader
> ac_utimescaled ??
> ac_stimescaled ??
> ac_utime sum for all live threads + cdata_threads
> ac_stime sum for all live threads + cdata_threads
> ac_minflt sum for all live threads + cdata_threads
> ac_majflt sum for all live threads + cdata_threads
>
> new:
> ac_cutime cdata_wait
> ac_cstime cdata_wait
> ac_tgid thread group leader
>

This seems to be make sense.

> xacct_add_tsk():
> ----------------
> coremem ?
> virtmem ?
> hiwater_rss thread group leader
> hiwater_vm thread group leader
> read_char sum for all live threads + tsk->signal.ioac
> write_char sum for all live threads + tsk->signal.ioac
> read_syscalls sum for all live threads + tsk->signal.ioac
> write_syscalls sum for all live threads + tsk->signal.ioac
> read_bytes sum for all live threads + tsk->signal.ioac
> write bytes sum for all live threads + tsk->signal.ioac
> cancelled_write_bytes sum for all live threads + tsk->signal.ioac
>
> If we leave everything as it currently is for tgid accounting, we
> probably should also not include the cXtime for tgids and just include
> the data for threads (TASKSTATS_CMD_ATTR_PID and task exit events).
>
> What do you think?
>

This makes sense to me, we'd need to bump up the API version if we are
going to make this change.

--
Three Cheers,
Balbir


\
 
 \ /
  Last update: 2010-12-07 07:01    [W:0.049 / U:0.036 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site