PRM on Service Guard

[sreeram n]

Hi Admins,

I have two node cluster ( N4000 running hpux 11.11) with 4 packages running on node1 and 6 packages running on node2. Due to some performance issue we want to move one package to node2 from node1.

I want to allocate some resource dedicatedly to that package at the time of failover. That is when ever failover happens I need to allocate some resource to that package on node1.

I have gone through some hp docs and found it is possible through PRM.

My question is, is it advisable to do so?. Has any one of you done this before.

If yes how can it be done?

Your expert advice is required for me to proceed further.

I am not familiar with PRM.

Thanks,
Sreeram

[sreeram n]

Gurus,

Please help..

thanks Sreeram

[Steven E. Protter]

Shalom,

PRM is a big pricey tool that can help you manage performance. It has a lot of bells and whistles and a learning curve.

If other performance enhancers like more memory or CPU's are not available you may wish to consider it.

You need to take a cold hard look though whether its not best to just provide your operations department some sudo enabled fail over scripts they can use when top or glance looks sluggish.

SEP

[TwoProc · Accepted solution]

You can do this with PRM. I used PRM for a while, but in the end - I de-configured it from running on the servers.

The reason:
Let's say you want to allocate 25% of processing to a particular application and it's set of related tools. If you turn on the "CAP" in which the program can never take 25% percent of available resources, then even if the rest of the machine is idle, you'll only consume 25% of the total CPU. If you don't turn on the "CAP" then you're programs assigned to 25% may exceed the 25% limit, if there are unused cycles left on the box. So, if you put this in "NONCAP" mode, your program could theoretically use 100% of the server, if the rest of the server was unused. Of course, if it was 50% unused then only that portion would be available to you.
So you get to pick one of the modes, CAP or UNCAPPED. After testing it out, you'll probably pick UNCAPPED of the two. Just because it seems silly to have extra processing horsepower waiting around not use it.

In either case, once you're either "CAP"mode or you're exceeding the capacity of the box in UNCAPPED mode, and thus, in either case, PRM is now working for you.

The problem becomes, if you're allocation is 25% for this program - you're NOT going to get 25% of your 8 cpus in the box, which would be two processors available to program. Nu-uh, what you're going to get is 25% of EVERY processor! Sounds good right? Maybe and Maybe not. Since you've got an N4000, the fastest chips for that thing were, I believe, 550 Mhz. So 1/4 of 550 is about 138 Mhz. So, if your program is single threaded, then the single thread will be effectively running on a 138Mhz proc. That is, ... v e r y s l o w l y. And the rest of your 1/4% of processing would be ... idle, or not depending on the setting of CAP.

So, using PRM you'd think you're getting 25% of 8 cpus, which would be two 550 Mhz cpus! The problem is - you're not. You're going to get 1/4 of 8 CPUS.

In other words, you're processor allocation becomes 8 SLOW processors, instead of two fast ones.

What's wrong with this? Well, it depends, 8 slow procs may be good for some applications, like maybe a medium to light weight web server. But it's terrible for a database, a java servlet, etc. It's not good for anything that best utilizes any single full speed processing.

So, if the process you're moving has numerous processes that come up and go away very quickly, then PRM works fine. If your process is one that spawns several very hard working processes that stay active for extended lenghts of time, then you'll be dissapointed.

PRM is a great tool and it has its place, but sometimes it is difficult to place it correctly.

[sreeram n]

Thanks Steven and twoproc for your reply. Let me talk to my management with these details.

Thanks again.

[Emil Velez]

If you have the Mission Critical OE you might have the entitlement for PRM.

Do not use CPU Capping as outlined above.

PRM is good for serviceGuard environments. Lets say that system A has a application that uses 80% of the CPU. System B has a application that uses 70% of the CPU. IN a failover you want application A to get 70% of the CPU and application B to get 20% of the CPU. When there is extra capacity before failover without CPU capping the applications can use the extra CPU and memory that is not allocated so PRM really only benefits when you are at 90% or more CPU and there is a contention for resources.

I hope this helps.