Power Systems i: Serve's Up
Published: December 14, 2009
by Timothy Prickett Morgan
If anything is clear about the upcoming eight-core Power7 processors from IBM due in the first half of next year, it is that these processors are going to have a pretty big chunk of processing capacity. And as I am spending a few weeks in the lead story of The Four Hundred going over the problems and possibilities of the Power Systems i line, it seems appropriate to continue to talk about the issues facing the platform and then, in the new year, set about designing a new and revised AS/400 product line, complete with new pricing and packaging to make a truly competitive entry, midrange, and enterprise platform that takes advantage of that processing capacity.
I am thinking out loud in this series of articles, brainstorming as well as ranting, and I am not so much interested in putting together a well-structured argument and analysis as I am trying to look at the issue of revamping what we still call the AS/400 from as many angles as is fruitful.
IBM can cram a whole lot of Power7 computing capacity into a fairly small space, as I showed in my review of the Power7 IH supercomputer node, which I saw at the SC09 trade show in November out in Portland, Oregon. If IBM booted the future i 7.1 operating system (what you and I would call OS/400 V7R1) on a single node of this supercomputer, it would have an 800,000 CPW cluster of OS/400 server nodes all crammed into a 2U high rack server (one that is six feet deep and three feet wide, mind you) with 1 TB of main memory and huge amounts of bandwidth. (See How Does 800,000 CPWs in a 2U Server Grab You? for the details on the IH supercomputer node.) That node has eight multichip modules, with each MCM packing four Power7 chips, each chip having eight cores running at a top speed of 4 GHz. Each one of those cores on the Power7 chips would have around 3,125 CPWs of online transaction processing capacity, plus loads of electronics to do floating point math, decimal math, and vector math.
This kind of dense packing of the Power7 processors is not going to be available for general-purpose i, AIX, or Linux boxes, but it was kind of interesting to hold 100,000 CPWs of raw power in the palm of my hand. This is as much oomph as a 32-core Power6+ Power 570 machine delivers, and that is a 16U, four-chassis box.
IBM started talking about the Power7 chip in a bit more detail back in August, first here and then here. I am not going to go through the chip architecture again, except to remind you that with 1.6 billion transistors and a fairly large surface area (the Power7 chip was about one and a half times the size of a Liberty postage stamp when I saw it at SC09), there are going to be plenty of duds. Or more precisely, plenty of parts that would otherwise end up in the scrap heap because one, two, four, or even six or seven cores don't work properly, but the remaining cores can be electronically isolated from the duds, resulting in perfectly usable parts.
IBM's product roadmap already calls for making use of these chips, as I explained back in September. Based on the specs that I have seen and a little guesswork, here's how the two classes of Power7 boxes will probably stack up:
- Entry and midrange blade and rack servers will get Power7s with up to four working cores, with maybe 8 MB or 16 MB of the 32 MB of embedded DRAM L3 cache on the chip activated and only one DDR3 memory controller per chip; and these chips will span two or four processor sockets in a regular SMP configuration. These chips will be used in Power 520 and Power 550 class boxes as well as on J Series blade servers with two or four sockets, I would guess.
- Larger midrange and high-end boxes will get Power7 chips with what I am guessing will be six or eight working cores, both memory controllers turned on, and 16 MB up to 32 MB of L3 eDRAM cache. I would speculate that these Power7 packages will be used in Power 560, Power 570, and Power 595 class machines. My guess is that the top-end Power7 box will have 32 sockets for a total of 256 cores and 1,024 threads. By the way, that is the same number of cores and threads as the single Power7 IH supercomputer node. The chips will be in an SMP setup instead of a parallel cluster, however, and will probably be rated at around 800,000 CPW, but DB2 for i only scales across 32 cores at the moment so it will take eight partitions to get there unless IBM does some work on the i 7.1 integrated database.
There are two things that Big Blue can do for the Power Systems i line, given these constraints on the chips.
First and most importantly, IBM should make a line of machines that are dirt cheap and that use the scrap chips with maybe only one or two cores working and a minimal amount of L3 cache memory. If these parts are going to be tossed out anyway, it is better to make use of them. They could be configured with i 7.1 and offered only to customers with installed vintage iSeries and System i boxes who have modest processing needs. I am talking about a truly inexpensive Power Systems i box here--one that costs the same as an X64 server with the equivalent processing capacity, and less if the X64 boxes have more oomph.
I do not want these budget Power Systems i boxes to be second-class systems, as the Smart Cube appliance servers are. (You buy one, and you can't upgrade out of one into the real Power Systems line.) They are simply the entry point to the product line, just like Celeron and Pentium machines are entry products that feed into bigger and yet compatible Xeon products in the X64 world. And if IBM wants to put multiple mostly-dud cores in a box on the cheap, that's great. These will be roughly analogous to the current Power 520 entry machines, only we can think of them as the Power 505 and Power 510 machines that should have been there all along supporting i/OS.
On entry, midrange, and enterprise boxes that have more cores per chip activated, IBM would be wise to take a page out of its mainframe playbook and designate some engines for i/OS and others to run particular workloads as so-called "specialty engines." These could be configured to run database queries, support Java virtual machines, run PHP applications, support various application modernization front ends from third-party tool suppliers, support Linux for infrastructure workloads, run firewalls, support fault tolerance and disaster recovery mirroring on a single system, or even allow for the harvesting of number-crunching capacity for analytics applications running on workstations.
While I don't like the idea of i/OS engines carrying a premium price, lowering the overall price of a mixed workload system with i/OS and its DB2 for i database at the center of it all is a lot better than having a Power Systems i box that is too expensive to do anything but run the absolutely minimal amount of legacy code. I would prefer that these future Power7 boxes offer better value for the dollar than IBM's AIX-DB2 combo, in fact.
I know. Very funny, right? I saw the pigs with wings on that one myself. . . .
Power Systems i: Thinking Inside the Box
Rolling Thunder Rollout for Power7 Processors Next Year
Start Planning for Power7 Iron Now
IBM to Reveal Power7 Secrets at Hot Chips
Power 7: Lots of Cores, Lots of Threads
AS/400: Still Kicking After 21 Years
With No Power6 QCMs, IBM Waits for Power7
IBM Launches Power6+ Servers--Again
Come On Out, Power6+, You Win
Power vs. Nehalem: Time to Double Up and Double Down
Power vs. Nehalem: Scalability Is So 1995, Cash is So 2009
Post this story to del.icio.us
Post this story to Digg
Post this story to Slashdot