Per Process Network I/O info needed

[Boyd Johnson]

To what I have seen so far, there are a number of good utilities providing useage of cpu and disk down to a process level.

However, I need to know the input bytes, output bytes, and input rate, output rate of network activity on a per process basis.

I need this information to diagnose performance issues related to one (or more) process hogging the network (or out of control).

What can I do? suggestions? If you suggest glance or gpm, please indicate SPECIFICALLY how to get this information as I have looked and cannot get this information.

A sniffer is only good for an interface card level but again does not get me to the process level.

Thanks

M. D. Parker
UNIX systems manager
michael.parker@am.sony.com

[Sridhar Bhaskarla]

Boyd,

It is a good question. I do have a partial answer. This is not going to give you detailed network staticstics but can help you determine which application/process is waiting on Network.

If you do have measureware, you can configure applications (process/user groups) and let the system collect the data for sometime. Then you can use extract utility to extract the information related to each application. The metric that you will be interested is APP_NETWORK_SUBSYSTEM_WAIT_PCT.

Good examples are under /var/opt/perf/parm file. There are some default applications already defined in the file. You can make use of those entries to configure your own applications. You can be as detailed as you want.

Look at the man page of extract for more information. extract itself has a guided tour that should help you extract the stats.

-Sri

[Boyd Johnson]

There are no network waits visible (all 0%). I did discover what you indicated in your reply previously. But I really need to see specifically the network specific information on a per process level.

I am missing something here...this seems to be a tremendous missing piece of information that glance/measureware is unable to provide.

Mike

[Sridhar Bhaskarla]

Boyd,

You are correct that Glance+/Measureware lacks this information and there were not many complaints also so far.

Now, you can still get per process information again limited to network wait only. You need to remove the thresholds in /var/opt/perf/parm file otherwise only the processes that exceed these thresholds will be logged. Once you get the logging on for all the processes, you can extract PROC_LAN_WAIT_PCT PROC_LAN_WAIT_TIME to give you an idea. You should be looking at
SHIFT-b window in glance and LAN+STREAMS waits to see the waits on the network.

-Sri