As part of JesperFrimann's challenge with respect to the options available to a company with two aging p690s running Sybase, he suggested that:
It's 256 threads per box on the T5440 versus 16 on the p550, and furthermore the power 550 supports shared pool virtualization.. so it will allow for a much higher utilization. So the price performance for the p550 wil be far superior to the T5440 IMHO.
So today's question is: is this true?
Notice, before we get into this, that while I don't want to make excuses for the Power6 it's important to note that this machine has to carry two enormous burdens for IBM:
- first, it was designed to counter Itanium; not SPARC, and not x64. As a result IBM focused on cache and megahertz to the near exclusion of power efficiency and hardware assist processing (things like Sun's onboard crpytology and packet managers) because those were never going to be issues in competing with stuff from Intel; and,
- second, Power6 is the penultimate transition product - the next generation, the Power7, is planned to have eight cores, each offering four SMT style, switchable "threads" running at 4Ghz, and will be the last of the PowerRISC line. After that (assuming IBM can make and hold to the decision) the project Quasar CPUs, built on the 16 SPU Cell2 processor and expected to appear at 4Ghz with up to four processor sets (G6+, SAP, and 16 SPUs) on a single die, will take over.
On its own, however, the Power6 is very nice processor I never hesitate to recommend to IBM customers or as an alternative to Itanium -but the errors IBM made in the strategic assumptions underlying its development show up in odd places: including a fairly desperate scarcity of auditable benchmarks run against SPARC or higher end Opteron (and, soon, Nahelem) gear.
One important one does exist: a SPECjAppServer(R)2004 result featuring the Apache based Websphere with J2EE and DB2 under IBM AIX 5L v5.3. Here's the hardware summary:
|J2EE Application Server||Database Server|
|Model||p570 (4.7 GHz, 4 core, SMT on)||p550 (2.1Ghz, 4 cores, 2 chips, 2 cores/chip (SMT on))|
|Internal disk||4 x 146 GB 15K RPM SAS Disk Drive||73.4 GB 10,000 RPM Ultra320|
|External disk||2 x DS4700 Express Model 70 (2 GB Cache) on FC, each 14 x 36.4 GB/15K|
The comparable Sun result, using a 32GB T5220 with one CPU (8 cores) at 1.4Ghz, running the J2EE application server against Oracle Application Server 10g on a 64GB T5120 (1 CPU, 8 cores, at 1.2Ghz) is 2,000.92 SPECjAppServer2004 JOPS@Standard - 1.67 x IBM's number.
IBM does not publish detailed P570 pricing. However, they do participate in the TPC/C benchmark Sun has considered obsolete since the mid ninties as both too unrealistic to be meaningful and too simple to game - and one result there includes cost details for the p570.
Based on the detail shown my estimate for the discounted price of just the p570 used in the J2EE test is $249K versus the $71,200 shown by Sun's on-line store for the T5220 and T5120 units combined - giving Sun a six times performance advantage per dollar if you could buy the p570 and get the p550 and disk free.
Interestingly, the details published by IBM for the TPC report include this bit:
4.7GHz POWER6 -2 Core Processor Card, 0-core active $11,500 (each, times 8) One Processor Activation for Processor Feature #7380 $23,000 (each, times 16)
suggesting that the cost of just the four licensed cores in the P570 used as the J2EE server is 1.6 times the nominal list price of the two Sun servers combined: delivered, set-up, and fully warranted for three years - and because the Sun combo is also rather a lot faster, the real bottom line here is that only a truly committed IBMer would ever buy a p550/570 combination for something a T2 will do.