Tracking disk IO by PID or user

[Steven Sauve]

Having used sar to track disk IO, I'm now trying to find out who/what processes are causing it. I know that glance has a real time "most active process" line, but I'm looking for something I can dump into a file so I can play with it a bit. Does anyone know of a command that will give me IO by process or even the top ~10 processes (IO wise)?
Thanks
Steve

[Anthony Goonetilleke]

It can be a bit messy but you can start of by doing something like this

fuser /dev/vg00/lvol1

this gives you all the pid's on that lv you can use this to check the IO utilisation of each proccess and then use perl or something to do a bit of math..

[Andy Monks]

Measureware can do that. It's runs all the time (uses the midaemon glance uses) and captures the data for later analyses.

It's a purchasable product however.

The only other way, is to write your own code using pstat() calls.

[Jim Welch · Accepted solution]

If you want to use glance, you could use it in adviser mode. For example, create a file /tmp/proc.adviser containing:
PhysIoThreshold=0
FirstTime=FirstTime
if FirstTime == 0 then
{
print " "
FirstTime=1
print " Phys"
print " Proc Proc IO"
print " Date Time Interval Name ID Rate"
print "------------------------------------------------------------"
}
process loop
{
if (PROC_DISK_PHYS_IO_RATE > PhysIoThreshold) then
{
print GBL_STATDATE," ",GBL_STATTIME," ",GBL_INTERVAL," ", PROC_PROC_NAME," ", PROC_PROC_ID," ",PROC_DISK_PHYS_IO_RATE
}
}

(note that last print statement is all one long line)
Then run glance like:
/opt/perf/bin/glance -adviser_only -syntax /tmp/proc.adviser -j 30 -iterations 10 > /tmp/proc.out

If your glance thresholds are set correctly, you should see something like the following in the proc.out file:


Phys
Proc Proc IO
Date Time Interval Name ID Rate
------------------------------------------------------------
06/13/00 21:02:35 5.0 glance 8159 0.2
06/13/00 21:02:50 5.0 vx_iflush_thread 28 0.2
06/13/00 21:02:50 5.0 glance 8159 0.2
06/13/00 21:02:55 5.0 vx_delxwri_threa 31 0.2
06/13/00 21:02:55 5.0 rpcd 833 0.2
06/13/00 21:02:55 5.0 vi 7948 1.0

(my system is pretty idle and I used a shorter interval and a 0 threshold so I could get some data).

You can set the threshold variable to cut down on the data if you like and you can
add more PROC_DISK_* metrics to the print statement depending on what you are trying to capture... see the /opt/perf/paperdocs/gp/C/metrics.pdf file for metric info and the adviser.pdf file for info on using adviser mode. Also note that the first interval is usually a short one while glance is initializing. I usually ignore the first interval and wait for one with something closer to the requested interval.