[lkml]   [2006]   [Jun]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectMerge of per task delay accounting (was Re: 2.6.18 -mm merge plans)
Andrew Morton wrote:
> per-task-delay-accounting-setup.patch
> per-task-delay-accounting-setup-fix-1.patch
> per-task-delay-accounting-setup-fix-2.patch
> per-task-delay-accounting-sync-block-i-o-and-swapin-delay-collection.patch
> per-task-delay-accounting-sync-block-i-o-and-swapin-delay-collection-fix-1.patch
> per-task-delay-accounting-cpu-delay-collection-via-schedstats.patch
> per-task-delay-accounting-cpu-delay-collection-via-schedstats-fix-1.patch
> per-task-delay-accounting-utilities-for-genetlink-usage.patch
> per-task-delay-accounting-taskstats-interface.patch
> per-task-delay-accounting-taskstats-interface-fix-1.patch
> per-task-delay-accounting-taskstats-interface-fix-2.patch
> per-task-delay-accounting-delay-accounting-usage-of-taskstats-interface.patch
> per-task-delay-accounting-delay-accounting-usage-of-taskstats-interface-use-portable-cputime-api-in-__delayacct_add_tsk.patch
> per-task-delay-accounting-documentation.patch
> per-task-delay-accounting-proc-export-of-aggregated-block-i-o-delays.patch
> per-task-delay-accounting-proc-export-of-aggregated-block-i-o-delays-warning-fix.patch
> I just don't know. There are a number of groups who pop up with various
> enhanced accounting requirements and patches (all quite different) but I
> haven't heard a lot of enthusiasm from any of them over this work, which
> attempts to provide an extensible framework for accumulation and querying
> of per-task metrics.
> But then again, we cannot just sit there and wait for everyone to be 100%
> happy. So I'm 51% inclined to push this along.
> Anyone else who has an interest in this sort of thing needs to be aware
> that there will be an expectation that any future statistics submissions
> should use these interfaces. So the time to pay attention is right now.

Hi, Andrew,

Here is a brief summary of the status of the response we have received from
the stakeholders (some of it has been duplicated in previous postings)

Project Response

1. CSA accounting/PAGG/JOB: Has agreed to use taskstats
Jay Lan <> interface

2. per-process IO statistics: None
Levent Serinol <> Needs are subset of CSA

3. per-cpu time statistics: None (email bounced)
Erich Focht <> Needs can be met by taskstats
Statistics not yet submitted

4. Microstate accounting: None
Peter Chubb <> overlap with delay accounting
prefers /proc due to convenience
taskstats can meet the needs

5. ELSA: Guillaume Thouvenin None
<> ELSA is not a direct user
of new kernel statistics
Consumer of CSA/BSD accounting

6. pnotify: Jes Sorensen <> None
(taken over pnotify from Erik Jacobson) Informed over private email
that pnotify replacement is
being worked on. pnotify
or its replacement will
not be concerned with
exporting data to user space
or collecting any statistics.

7. Scalable statistics counters with /proc Not working on it
reporting: anymore
Ravikiran G Thirumalai,
Dipankar Sarma <>

Studying the responses from all stake holders, Jay Lan's was the most
encouraging. Peter Chubb prefers the /proc interface due to the text interface
and ease of parsing. (in our opinion, taskstats can meet the needs easily
and the getdelays utility can provide the same ease for parsing).
The others did not respond.

Some performance numbers of taskstats were posted at The result highlights are included

Results highlights

- Configuring delay accounting adds < 0.5%
overhead in most cases and even reduces overhead
in some cases

- Enabling delay accounting has similar results
with a maximum overhead of 1.2% for hackbench,
most other overheads < 1% and reduction in
overhead in some cases

These statistics are _per task_ and can be extended easily by anyone
who wishes to obtain per task data. An example of per task improved
scheduler statistics was mentioned in
(I am not sure if the email refers to our per-task statistics). If not,
the new statistics could easily use the taskstats interface.

These statistics can be used by software product stacks to monitor
usage information about the various tasks they create and control.
I also informally spoke to a group of students (verbally), who were
excited at the possibility of using the per-task statistics to do
dynamic deadline based power management. They want to use the delay data
(CPU and IO) to predict deadlines for a task and then use these results
for dynamically scaling CPU frequency.

The ability to monitor the CPU run and delay data and IO delay data is useful.

I would request you to consider the inclusion per-task delay accounting into


Balbir Singh,
Linux Technology Center,
IBM Software Labs
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2006-06-06 03:02    [W:0.295 / U:1.848 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site