Number of cpu's vs speed - workload issue

[Geoff Wild]

We just migrated a POS system from a 2 node MC/SG environment on 8 way 440mhz cpu N-classes to a 2 node MC/SG environment on 4 way 875mhz rp7410's.

We are seeing higher cpu loads as well as higher system cpu percentages.

What I'm looking for is some good docs/white papers on explaining system behaviour.

IE: my theory is, we have gone from 8 slow workers to 4 fast workers. The amount of work is still the same - hence the higher cpu usage. I need some concrete evidence/docs to support this.

Thanks...Geoff

[Pete Randall]

Geoff,

Anecdotal, not concrete, but we experienced much the same results when we migrated from a 6-way 440Mhz N-class to a 3-way 875Mhz (it's actually a 4-way but we partition with VPARs). Performance seems to be just about the same. The only difference we've seen is during very CPU-intensive, number crunching jobs. They actually run a bit faster due to the sheer processing power of the new CPUs.


Pete

[Steven E. Protter]

http://www6.itrc.hp.com/service/james/dispDoc.do?docURL=http%3A%2F%2Fsearch.hp.com%2Fredirect.html%3Furl%3Dhttp%253A%2F%2Fforums1.itrc.hp.com%2Fservice%2Fforums%2Fquestionanswer.do%253FthreadId%253D542462%26qt%3D%252Bnumber%2B%252Bcpu%2B%2B%252Bversus%2B%252Bspeed%2B%26hit%3D1&aid=SEARCH_FORUMS&pil=1&serStr=number+cpu+versus+speed&pir=1

http://www6.itrc.hp.com/service/james/dispDoc.do?docURL=%2Fservice%2Fcki%2FdocDisplay.do%3FdocLocale%3Den_US%26docId%3D200000073548851&aid=SEARCH_CKI&pil=4&serStr=number+cpu+versus+speed

http://www6.itrc.hp.com/service/james/dispDoc.do?docURL=%2Fservice%2Fcki%2FdocDisplay.do%3FdocLocale%3Den_US%26docId%3D200000063203050&aid=SEARCH_CKI&pil=5&serStr=number+cpu+versus+speed

http://www6.itrc.hp.com/service/james/dispDoc.do?docURL=%2Fservice%2Fcki%2FdocDisplay.do%3FdocLocale%3Den_US%26docId%3D200000076960118&aid=SEARCH_CKI&pil=6&serStr=number+cpu+versus+speed

http://www6.itrc.hp.com/service/james/dispDoc.do?docURL=%2Fservice%2Fcki%2FdocDisplay.do%3FdocLocale%3Den_US%26docId%3D200000009253765&aid=SEARCH_CKI&pil=10&serStr=number+cpu+versus+speed

Maybe one of the above. title pages look interesting.

SEP

[Tim D Fulford]

Hi

sorry no white papers etc..
o 8x 440 = 3520 MHz
o 4x 875 = 3500 MHz...
Very similar in terms of ticks, so I would have expected a similar AVERAGE CPU% for similar workload. Similarly, if you look at tpmC benchmarks (from what ever source) you will probably see that the rp7410-8x875 is slightly more powerful than N4000-4x440 ..

So I'm guessing you have done one or more of the following
* Also swapped out the storage to a more powerful storage device and or the networking, so you may actually be getting more oomph out of rp7410 system.
* Database usually has #cpu-1 worker or virtual processes, so you would have gone from 7x 440 = 3080MHz to 3x 875 = 2625MHz; or looking at tpmC these may actually be equivalent, thus total average CPU may be the same. (e.g your initial assumption was correct)

Regards

Tim

[TwoProc]

What I'd like to offer is a "build-on" of what Tim's very good model is. While you've got roughly the same amount of CPU cycles available - you've got them in less divisible units. Meaning that while you've certainly got the same CPU cycle bandwidth - you've got only half as many workers. This means that it would take fewer very intense jobs to overrrun 1/2 the system - only 2 to be exact. Where as before, if you had two very intense jobs - they could only take over 25% of the server. Now, those jobs should finish a lot faster, but while you've got those "hypothetical 2" jobs - you are more bottled up for performance than you were before.

*IF* this is your case - and hey, it very well might not even be in the realm - you could probably fix it with HP's PRM product. It would allow you to take jobs of a certain type and never allow over x amount of % of cpu (say 25%) - which happened naturally for you on your slower 8 way box.

I guess another way to illustrate is something that I've been noticing for years when it comes to really small servers. It seems that a two way server runs about 4 times faster than the very same 1 way server. Reason? It only took one big job to shut us down before, and now, even with this, we still had a whole 'nother cpu to do multiple jobs on which were relatively well behaved and short.

[Geoff Wild]

Good answers so far - anyone from HP care to jump in?

Just so you know - storage remained the same. What did change?

N-Class to RP7410
HP-UX 11.0 to 11.iv1
MCSG 11.12 to 11.15

I have been looking at PRM to limit non-essential processess - like gzip for example (we gzip the archive logs).

Looking for detailed inner workings on HP-UX and how it relates to CPU's.

Rgds...Geoff

[Tim D Fulford]

Hi

I was going to ask if you were running the N on 11.0. My PERSONAL experiene is that 11i v1 is the same as 11.0, but uses 5-10% more memory for the same apps .... BUT if you run multi-threadded java apps then 11i v1 should be better.

The internals of the N4000 are slower than rp7410, but again I've never felt we (my company) have stressed either N or rp. It looks like you also have a similar situation as the rp is no "faster".

Regards

Tim