Yeah Yeah Yeah..
"Who the heck cares about scaling?"
Well I do.. Scaling is a very important. Which is why it is something that should be looked at when comparing systems and processors. Specially if you are comparing Itanium/POWER/SPARC64 versus x86. Cause Itanium/POWER/SPARC64 will do from 4-256 cores with the same chip. Intel x86-EP will only natively go to 2 chips and 8-12 Cores. Hence when you pull out your Nehalem-EP numbers.. then they are only valid for workloads that does not require more scalability than the Nehalem-EP chip can deliver. For larger workloads you have to look at Nehalem-EX. It's actually pretty simple.
If I need 300 specINT_rate2006, then it's kind of stupid to be looking at a Nehalem-EP x5570 processor cause it'll only take you to about 266. Where as the nehalem-EX X7560 will take you to.. well.. 1400 or so..
So that whole scalability comes at a price, and that is the per core throughput. Again the two above processors in 2 socket configs:
http://www.spec.org/cpu2006/results/res2010q3/cpu2006-20100621-11923.html
and
http://www.spec.org/cpu2006/results/res2010q1/cpu2006-20100315-09857.html
That is 266 and 33.25 per core specINT_rate2006 for the Nehalem-EP and 385 and 24,06 per core for the Nehalem-EX. So the 45% increase in chip throughput comes at a price of a 38% drop in per core throughput.
And you talked about TPC-C
Then
Nehalem-EP (Submit date 04/08/10 ) 631766 tpmc and 78971 tpmc/core:
http://www.tpc.org/results/individual_results/HP/HP_DL370_G6_OEL_TPCC_ES.pdf
And
Nehalem-EX (Submit date Aug 27, 2010) 1807347 tpmc and 56489 tpmc/core
http://www.tpc.org/results/individual_results/HP/HP_ProLiant_DL580G7_2.26GHz_es_100830_Energy_v2.pdf
And finally
POWER6 (Submit date May 21, 2007) 1616162 tpmc and 101010 tpmc/core
http://www.tpc.org/results/individual_results/IBM/IBM_570_20070522_ES.pdf
Again Nehalem-ex is 40% slower per core than -EP. And the picture is clear POWER6 is faster per core, (on this benchmark) than Nehalem.
And also notice that the POWER6 submission is from 2007, not the fastest POWER6 made, and the Nehalem are using SDD drives (which IMHO is a pretty big factor).
Now a current POWER7 system like this one:
http://www.tpc.org/results/individual_results/IBM/IBM_780_TPCC_20100719_es.pdf
does 1200011 tpmc and 150001 per core. Which is almost x3 per core of Nehalem-EX and almost twice that of Nehalem-EP per core performance.
And the sheer fact that you can come with such a statement, only shows that you still have much to learn.
"No, it is you that dont get it. Let me ask you, if you needed the highest performance in the world, which company would you have to go to, IBM or Oracle? Oracle! How the heck can IBM and IBMers claim they are still fastest in the world, because "IBM cores gave 4.7 times more tpmC per core"?? That is an outright lie, that technically ignorant executives might believe. FUD and lies, again. This is a lie, Jesper. Dont you see the lie?"
I wouldn't go to Oracle that's for sure. It's not the fastest machine.. it's a cluster .. CLUSTER. And if you read the pricing information on their TPC-C benchmark you'll see that you don't even buy the software... you lease it.. For the exact amount of years that the benchmark has to do TCO on. You have to pay the listed amount of money every 3 years... And there is NO upgrade protection.
So who is trying to Bull who ? On the POWER 595 benchmark you at least buy the software and then only has to pay (which can be expensive enough btw.) software maintenance and support.
So buying 2 years of extra support for DB2 on the (going from 3 year -5 year TCO) would cost you ... 203,827 USD-Discount. On the Oracle solution using the pricing scheme that they use on the benchmark it would be.... 7,872,000 USD for 3 years.
Now who is b*llsh*tting who ?
You are so drunk on the Oracle/SUN Koolaid that you don't care to read the fine print.
"In short, when I show links or benchmarks, you immediately dismiss them as FUD, lies, and amateurs. When you show links or benchmarks, I accept them. What does that tell you, about Jesper Frimann?"
That I'm much much better at coming up with links than you are ?
It's not my fault that IBM doesn't release any benchmarks on Mainframes. Go complain to them. I just know the numbers that we use. Would a 8 core Nehalem-EX chip faster than a 4 core Mainframe chip on benchmark like specINT_rate2006. Jup sure no argument from me.
But 8 times faster, is so far off the target that it can only be described as FUD, or well.. stupidity.
Also here the type of workload plays a big part. Mainframe Cores have never been known for their ability to crunch numbers, Moving data on the other hand they are pretty good at, and that is also one of the secrets why they can run at such high utilization. Which is also a factor that you have to look at in RL.
"That is hilarious. And proves the z196 is an abomination that should have been killed off and never left the laboratory."
You really really don't get it. Your fanaticism is scary.
"You make it difficult for me. Sometimes you say one thing, the other time you say the opposite.."
That is most likely cause you don't understand what I say.
Now as for response time. I've only pointed out differences in that on the benchmark and said that you can trade response time for throughput.
On the TPC-C benchmark you are quite right there is a big difference in response times, between the POWER 595 and the T5440 clustered benchmark.
On that particular benchmark then one of the key factors is surely the use of SDD on the T5440, which wasn't available back when the POWER 595 benchmark was made. It's a pretty big difference.
But the T5440 clustered benchmark do seem to have idle processing power that could have been used to increase the throughput at the cost of response time. How much is hard to say, why well according to a friend of mine who's a certified Oracle RAC dude, then it's cause they are pushing the limits for scalability on the benchmark, hence they aren't hitting the optimal per core throughput.
// Jesper
Biting the hand that feeds IT
