I am currently given the dubious task to make an assumption on sizing our current p5 environment in relation to new p7 machines. Is there any kind of a vague factor how to calculate this?
I found some official IBM benchmark results with long lists of different Power machines and their scores but since I do not know the benchmarks themselves they were kinda useless for me.
Is there a simple kind of formula how to guess what number of p7 CPUs you might need to compensate a number of p5 CPUs?
what kind of apps are you running (DBs, Websphere ...) and are your p5 virtualized or capped And will your cpu pool be generous or rather low.
I am doing the a lot of sizing for my company - at the moment we are consolidating 38 small p5 frames (550/570ies) into 4 780ies ... what I find is that from a pure performance perspective a few rules of thumb still work perfectly fine - we have generous pools so I do not really care about cpu entitlements when I do the initial setup. If I have a sybase DB, I use one virtual per engine (backup server and repserver count as engines) + 1 as count for virtual cpus, on Oracle I am checking how many parallel processes (not connections) I have running on the server I want to migrate - that will be the amount of parallel threads I want to allow on the new box too - 1 virtual = 4 threads + 1 virtual for the OS. My virtuals are generally worth 0.1 physical cpu - as they can grow to a physical cpu during peaks that is just fine - what is not needed goes back into the pool. Monitor about 4 weeks after go live closely and than amend the amount of entitled cpu to what you usually need (so in average, NOT during peaks) and that is what they get permanently. Regarding memory - well its pretty much the same.
For websphere boxes we usually go with half as many virtuals as we were using on p5. Again - monitor closely and amend if required.
In my experience you have usually a lot of cpus in the pool and most frames do not have the peak times of all their lpars same time so you should usually be good.
For DBs you can go if you like as well another way: p5 to p6 = 3:2, p5 to p7 = 7:4 - that is what our engineering came up with - they assume high frame usage and small pool
I have the official IBM numbers in the office - can post tomorrow ...
@DGPicket
For those specInt ratings I have no clue as it sounds a bit abstract to me; will have to read about it to be able to interpret them.
@zxmaus
We have a lot different partitions, some capped, some uncapped and apart from saving licenses there is often no real reason why they are capped or not, some shared some dedicated. I will keep it in mind but I am maybe not the one who is deciding this and people here are often resistant to advices.
Application wise we are having a lot of Oracle DBs, a few Informix DBs and some boxes with Weblogic, which should be comparable to Websphere; there are also some minor applications but they are not that hungry.
For the generousity of pools, I do not expect much here. But that is another topic
I would really appreciate those official IBM infos, thanks a lot in forward ^^
Neat! Remember that for the same total nominal spec-int or whatever, fewer processors is really faster. You only get full speed from the first, and then for each additional one sharing the same RAM, bus's, L2cache, L1cache, you get less. 4 cores of 20 > 8 cores of 10.
The standard for doing this work is to compare the rPerf number in the link that zxmaus gave you. This is an IBM standard where a p640 has a value of 1. This is an analytical model of commercial workloads rather than fancy benchmarks so has good application in the field.
Note that it doesn't look at I/O (Network and disk). Also your IBM representative (or BP) should be able to provide you with assistance that your strategy for consolidation is correct.
Don't forget memory overheads for LPARS, VIO etc etc.
Yep, that was the official thing I mentioned in my 1st post and I think I have to take a closer look at those values and descriptions.
Currently I might simply have a look for our processors ie.
p570, p5 1,9GHz
p590, p5 2.1GHz
and just compare them with some middle sized p7 boxes to get some kind of ratios.
Tbh I am not sure if we use more int or fp base calculations. That's something I never (or anyone in our company) lost a thought about.
Well, for fp to get interesting, someone would be modeling airflow on a wing or something like that, with huge matrixes of fp (more at home on a SIMD box but $$). Even applications with considerable fp still use many int operations to calculate addresses. Applications more in strings and RDBMS get very little from fp prowess.
to make an assumption on sizing our current p5 environment in relation to new p7 machines. Is there any kind of a vague factor how to calculate this?
And will your cpu pool be generous or rather low.