performance stats per process

[john strumila]

howdy gurus,

I've been doing performance for many years on non-unix boxes but now I'm interested in HP-UX.

I want performance stats (eg pstat) from each process running and I want to output to a file. Sort of SAR but for each process.

measureware/perfview costs money and we have many boxes, so it's out. TOP is good but it doesnt seem to output to a file and doesn't include 'system' cpu per process (nor does measureware?). Also, the data available in pstat seems far more comprehensive than what's supplied in measureware/TOP/SAR.

Any ideas? Do I have to write something myself?

[James R. Ferguson]

John:

My vote still goes for Glance. I think it's well worth the cost for all you can see.

...JRF...

[Manju Kampli]

John,

you can buy a Measureware clients for all the unix clients and have a central server where it collects the data. use perfview to see the utilisation/performance of system resources. using measureware and perfview you can also see per process resource utilisation using "Drill down to process option". Ofcourse this will give you the history data ( may be 15 mins behind, since measureware data is updated periodically to the server. you can adjust this time, but the system may spend more time in doing this than some thing important)
If you want to look in to the live process utilisation per process either top or sar may be helpful if you don't want to spend money. But as said above Glance is worth spending the money which gives all the minor detail which can help you in determining what each process is doing.

[john strumila]

thanks guys,

As I said, SAR doesn't give data per process and although GLANCE is good it doesn't give as much detail as is possible. Stuff like RSS broken down by text,data,stack or "jobs in disk/page wait" or "disk enqueue to done" dont seem to be there.

It looks like an opening for another product to use them.

[Victor BERRIDGE]

Have a look at loan system product:sysload
It costs but nothing like glance/perf (true in Europe anyway) you pay for a server plus for clients there are clients for other ux flavor.
This product can do a lot
I had it on trial, but the people who make the decision have imposed BMC PATROL
AHRRRrrr..

They said it cant be that good at that price...

http://www.sysload.com/

Regards
Victor

[Stefan Farrelly · Accepted solution]

The graphical version of glance (gpm) does indeed give you a breakdown of RRS by stack,text,data,private,shared,other etc. (under Process Memory Regions) and also queues/wait states (Process Wait States). You get tons more detail using gpm, and if you need even more detail use PerfView.

[Victor BERRIDGE]

Hi Stefan,
I have nothing agains Glance...
But Im not the one who decides what to buy...All I can do is say I want...
I have glance but only on 2 machines that are critical and I had to fight long for it...
I would put glance on all hps aixs solaris..

Sysload is a good alternative if you cant afford glance (in my opinion)
What interresting about sysload is you can view many clients at the same time!
Its in fact completely custom , I mean do decide what you want to monitor, display etc...

[Anthony deRito]

John, GlancePlus will give you more information about a process than you would ever want. If the money is somewhat not a factor for your decision then you have no excuse NOT to purchase GlancePlus. If it is, then forget it.

But your decision to purchase a product should be based on more than cost. (I know... I am using the word "should" real loosely here)If you have critical production servers that requires quick problem resolution and performance bottleneck identification, then it should be good sell to your leadership.

I would not worry about not having enough stats on your processes with GlancePlus but more on how you will learn to interpret and use the abundant information that is given. Try the product for its 30 day eval and you will see what I mean.

Tony

[john strumila]

thanks again guys,

I dont have the gui so I cant check but I still haven't seen disk "enqueue to done" anywhere - not even sadp.

Of course my basic problem is that the boss aint gonna pay for glance on all boxes (we're not critical).

A likely alternative for me is to butcher TOP to get it to write all the data out.

BTW, syscalls shows pstat but does anyone know what stat64 is?

thanks,
John

[Jim Welch]

You can use the pstat system scalls to collect information on individual processes as well as lots of other data. See man pstat.
Also read /usr/include/sys/pstat.h
And it's free if you don't count the time it takes to write the code...