IBM Keeps the Power 595 at 254 Partitions, For Now
April 21, 2008 Timothy Prickett Morgan
Sometimes, server makers can be perplexing. Every server maker on the planet, and their chip partners if they use them for all or some of their product lines, are keen on showing that they have hardware-assisted virtualization electronics built into their chips. IBM started adding virtualization features into the hardware to better deploy logical partitions a decade ago and a decade earlier in mainframes, so this is old hat to Big Blue even though people make a big deal about it these days on the X64 platform or various RISC or Itanium platforms.
In October 2006, I had a long chat with Brad McCredie, the chief architect of the Power6 processor who works in IBM’s Systems and Technology Group, about the virtualization capabilities IBM was putting into the Power6 chips. As we all know, the Power4 and Power5 chips could support a maximum of 10 logical partitions per processor core, up to a maximum of 254 partitions on the high-end, 64-core System i 595 and System p 595 servers. In theory, there was no reason why these top-end machines could not support 640 partitions (or maybe 638, saving two of them for whatever purpose IBM has, since it is clearly doing this on the 595 machines already). But IBM wanted the partitions to have enough processing oomph to be useful on the bigger machines, and with the SMP overhead, the company capped partitions at a little less than half of what the theoretical maximum was for the Power4 and Power5 architectures.
But with the Power6 chips, IBM clearly had other ideas, and I was flabbergasted when I heard McCredie tell me that IBM had woven support for 512 logical partitions per Power6 core into the architecture of the chip. You will remember from IT Jungle’s coverage back in the fall of 2006, that I did a little math to show that this worked out to a stunning 32,768 partitions per 595-class server, and assuming a Power 595 could handle about 8 million transactions per minute, that would leave about 244 TPM per partition, or about as much oooph as a 25 MHz or 33 MHz 80486 server from the dawn of time. I reckon that a useful partition needs about 20 times more performance than that for running a modest workload for a handful of users, but that still works out to over 1,600 logical partitions for such a big box. And you will also remember that McCredie made no promised that IBM would implement that many partitions in what has become the PowerVM hypervisor.
As it turns out, the Power 595 announced on April 8 will indeed only support the same 254 maximum logical partitions as its predecessors. In fact, whatever the underlying Power6 hardware can do, the Power Systems family of machines is still capped at 10 processors per core and still has caps on the 595 box. Why, you ask? I dunno. My guess is that IBM has to tweak the PowerVM hypervisor so it can make use of the new virtualization hardware in the Power6 chips, and I will guess further that future operating system releases–call them i 6.2 and AIX 6.2–will include an updated set of PowerVM hypervisors that allow for more logical partitions. As a minimum, IBM really needs to boost the count to 15 per core, and really should double or triple the maximum limits imposed at the system level. This is necessary to better position the Power Systems line for server consolidation. Other X64 hypervisors do not have any such limits, although it is a rare box that has more than a handful of partitions on it anyway for now. That will change as virtual and logical machine partitioning becomes the norm, not some exotic thing leading-edge companies do for production workloads.