Perfomance Question.

[ben_43]

Hi:

We have 4 nclass servers and the worst case scenario is that we have to support 400 users/box(assuming 2 boxes go down). We had some load test and the following resources got to the maximum
nfile - 97%
cpu - 60-70%
we have 6 CPU's and 20 Gig Mem. on all the four boxes.Assuming we have to support 600 users(on one box) what are the additional resources we may require? The users come in thru xemulation (from PC) and load average of the box(output of w command)was around 15 for 400 users. I am also attaching my kmtune output. Any thoughts???

Cheers
Ben.

[Pete Randall]

Ben,

I'd bump maxusers up to 1200 or so. This should also raise nfile (assuming you're still using formulas).

Pete

[ben_43]

Pete:

Thanks. I am not on formula. But the question would be would 600 be possible to support on a single box with these resources? If not what additional resources we may require?

[Martin Johnson]

You should be able to handle 600 users. Response time will be a little slower, but this should not be a permanent situtation as the down systems should be repaired ASAP.

You may need more swap space. Also check current usage of the file table and file locks and make sure there is room enough to handle a 50% increase in load.

HTH
Marty

[Pete Randall]

Ben,

When you say "support 600 users", do you mean on a temporary basis until repairs can be made to other, failed servers? Assuming that's the case, I would say yes, it will support 600 with some reduction in performance. If you're willing to accept that then you don't need to add any resources.

Pete

[Wodisch]

Hi Ben,

your systems seem to be "big" enough, but for the tuning, well, that'll need quite some effort!
- 60-70% cpu with 200 users => you will not be able to stand 400 or 600!
- 97% nfile => immediate problem, core dumping apps, even :-(
- what about disk I/O? To what degree is the bandwidth to/from your disks saturated? That'll be your biggest problem, I guess...
- what about network I/O? similiar to disk I/O...

Just some other $0.02,
Wodisch

[A. Clay Stephenson · Accepted solution]

As a first approximation, nfile, nflocks, and nproc should increase by about 50%. You probably need to increase some IPCS paramters (e.g. shmseg by about 50% as well). Npty, nstrpty, and nstrtel also probably need to be increased. I would say that your biggest unknown is the amount of swap space. You are using pseudoswap now and I assume that you probably are correctly using less swap than memory. If that is the case then 20GB may not be enough for the additional 200 users. You have two options: 1) add more memory or 2) add swap but know that you are going to take a big performance hit if you are swapping. The performance hit may be acceptable if you are only going to be pulling this many users for relatively short periods of time.

[ben_43]

Hi:

We are OK with the disk and memory. The main factor would be the CPU and the kernel parameters. The box can support upto 8 cpu's and in that sense we can go for 2 more if required. The load scenario was -Load average of 15 for 400 users. And we need to laod and soak 600.

Thanks
Ben.

[H.Merijn Brand (procura]

playing with mktune is much more fun if you can see the real value and the formula next to eachother. Then you see the relations.

maxuser increment is good, but influences a lot of other parameters.

use attached kmtune.pl to see them side-by-side and then reread the good advice from all above

[ben_43]

Hi:

I have attached my swapinfo output. I mentioned that the load average on the box was 15. Is it not very high? will it not go up with 600 users?

Thanks
Ben.

[Sandip Ghosh]

You can ad swap on the fly. So you can do as per your requirement. But dont keep both the swap in the same disk. Keep the different swap in different disk so that it should not create any problem during accessing the swap. If both the swap is in the same disk then it will be slow. And Keep the priority of all the swap same. Then it can access different swap simultaneously.

Sandip

[Martin Johnson]

Yes, the load will go up and the result will be slower response time. Again, your senario is losing 2 out of 4 servers at the same time. This should be a very rare event.

Instead of looking at adding CPUs, I would review the service contract on these systems to insure it is at the support level you need.

The focus should be on how quickly can you restore service to the downed systems, not on having excess, unused, capacity available - a waste of money.

HTH
Marty