taskstats and /proc/.../io asymmetry?

taskstats and /proc/.../io asymmetry?

[Mark Seger]

If you look at /proc/pid/stat, you can get the total CPU consumed by a 
process.  If you look at /proc/pid/task/tid/stat you can get the cpu 
consumed by a thread and if the tid is that of the parent you only gets 
its consumption as opposed to all its children.

I was surprised to see that the way process I/O is reported doesn't 
follow this model.  There are no /prod/pid/task/tid/io entries but 
rather you need to look in /proc/tid/io.  While I view this as a minor 
inconvenience, I can certainly live with it.  However, /proc/pid/io does 
not show the aggregate I/O numbers for the whole process and that both 
surprises and disappoints.  This means if I have a process that starts a 
bunch of worker threads that do the real work and I want to find the top 
I/O consumers I can't simply walk the /proc/pid tree but rather have to 
look at all the threads of each process and add them up.

Or am I missing something?

-mark

Re: taskstats and /proc/.../io asymmetry?

[Balbir Singh]

Mark Seger wrote:
> If you look at /proc/pid/stat, you can get the total CPU consumed by a
> process.  If you look at /proc/pid/task/tid/stat you can get the cpu
> consumed by a thread and if the tid is that of the parent you only gets
> its consumption as opposed to all its children.
> 
> I was surprised to see that the way process I/O is reported doesn't
> follow this model.  There are no /prod/pid/task/tid/io entries but
> rather you need to look in /proc/tid/io.  While I view this as a minor
> inconvenience, I can certainly live with it.  However, /proc/pid/io does
> not show the aggregate I/O numbers for the whole process and that both
> surprises and disappoints.  This means if I have a process that starts a
> bunch of worker threads that do the real work and I want to find the top
> I/O consumers I can't simply walk the /proc/pid tree but rather have to
> look at all the threads of each process and add them up.
> 
> Or am I missing something?
> 

I looked through the code and your argument seems to be correct. The behaviour
is inconsistent w.r.t. other statistics like utime and stime. We currently
accumulate tgid information in signal_struct, we need to do something similar
for io as well. If nobody gets to it by the time I finish my backlog, I'll try
and get to it.


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL