PureApplication Systems Get Power7+, But Not IBM i

February 18, 2013 Timothy Prickett Morgan

You would have to search high and low, across time and back again, to find a more pure application system than the Application System/400 and its forebears in the System/3X line and its great-grandchild, the Power Systems platform running IBM i. But oddly enough, as IBM announced the Power version of its PureApplication converged systems last week, the one thing it did not do is announce a variant of the stack that could support the IBM i operating system.

It is like blade servers all over again, with OS/400 getting such late and kludge support that the one customer set that might naturally pay a premium for the integration benefits of a BladeCenter chassis supporting a mix of Power and X86 blade servers and integrated management were alienated from the market and then invited late to the party. You probably know what I am thinking that Big Blue ought to do before I even write this story. But let’s go through it just to get the thinking clearly outlined and on paper. Well, pixels.

First things first. There is nothing wrong with the new PureApplication W1700 hardware and software stack announced on February 12 if you are an AIX customer running the DB2 database. Or an Oracle or Hewlett-Pakard shop wanting to move off their respective Sparc/Solaris or Itanium/HP-UX platforms. Or even Wintel and Lintel shops wanting to move up or over (depending on how you view the relative platforms) to Power-AIX. Anything that makes Power stronger makes IBM i live longer.

The new PureSystems W1700 appliances are similar to the Xeon E5-based W1500 appliances that IBM announced back in April 2012 when the Flex System modular servers, which merge servers, networks, and storage into a blade-like and rack-like hybrid. The difference is that the W1500 appliances are based on two-socket, half-width Flex x240 server nodes that use Intel Xeon E5 processors and, at least according to the internal IBM documentation I have seen, the W1700 is based on a four-socket, full-width server node that is based on the Power7+ processor from IBM.

But wait, you say? IBM has not announced a four-socket Power7+ server node for the Flex System chassis. Correct. Not as a standalone product, and not in anything you can read about the W1700 appliances in announcement letter 213-091 or on IBM’s Web site. And let me say something here about the good people of Software Group, which are managing the PureSystems launches. Data sheets are supposed to have data in them. Feeds, speeds, slots, and watts. This is a hardware and a system announcement after all. People want to know what the server node is, what switches you are using to link them together, how the storage is integrated, and all of the elements in the rack that comprise the PureApplication system. Saying it has this many cores, that much memory and disk, and not giving a price is not sufficient.

What I know for sure about the hardware in the new W1700 stacks is that Steve Sibley, director of worldwide product management for IBM’s Power Systems division, told me during a Power7+ briefing relating to the Power Systems iron two weeks ago that Big Blue was waiting for the Power7+ processor to launch the Power-based variant of the PureApplication stacks rather than do a Power7 version last year. This probably has to do with the performance, extra cache memory on chip, and expanded main memory (both physical and through much-improved memory compression, which now has its algorithms done on special accelerators on the Power7+ chips rather than by the CPUs) that the Power7+ chips offer. This stands to reason, of course. Moreover, IBM had to lead against Oracle’s Exadata database clusters and Exalogic middleware clusters, which are based on Xeons, with its own Xeon iron. Or at least someone believed that. If Power is so much better, and if I were IBM, I would always lead with Power and have X86 iron as an alternative.

Anyway, the announcement letter tells you how much a W1700 appliance weighs and how loud it is, but it doesn’t tell you what the processor nodes are, how many cores are on the processor, how much main memory can be added to the node, and it says nothing about the networking that links nodes to storage. One document I have seen confirms that it is a full-width node, and that means it is a four-socket box and it almost certainly has to be the forthcoming Flex p460+ server node I expect any day now as a standalone product. The setups come with from 96 to 608 cores, with memory scaling along with cores. The machines have integrated arrays with a set mix of flash and disk that has data protection across multiple RAID sets (with spare drives) so you can recover from storage errors. Again, the documents don’t say exactly what kind of RAID, and by the way, it matters what kind for both performance and availability reasons. Anyway, the system has arrays with 48 TB of disk and 6.4 TB of flash storage, presumably with hierarchical storage management. (Again, IBM doesn’t say and it should.) Here are their four sizes, and the prices of the fully configured W1700 application appliances:

Small: 96 cores and 1.5 TB of main memory; list price $2.97 million
Medium: 192 cores and 3 TB of memory; list price $5.64 million
Large: 384 cores and 6.1TB of memory; list price of $10.73 million
High-Performance: 608 cores and 9.7 TB of memory; list price of $16.73 million

The systems run AIX 7.1 and cannot be deployed with Linux or IBM i, even though technically there is no reason for it. The machinery is licensed to put the PowerVM hypervisor on each node if you want, and you can also put DB2 Enterprise Server Edition 9.7 on all of the nodes. Ditto for the V7.0, V8.0, and V8.5 versions of WebSphere Application Server Hypervisor Edition.

I was guessing that the small configuration is four four-socket p460+ server nodes using six-core Power7+ chips, and then medium doubles it up, and large doubles it again. The top-end machine has to be using eight-core Power7+ chips for the numbers to work, and it has to have 19 server nodes for the processor core count to work. This is just a guess in the absence of data. An IBM spokesman told me that the W1700s use eight-core Power7+ processors running at 3.61GHz. This is the midrange part in the p260+ node announced last November, by the way, and to get the core counts to work, you could have six half-width nodes instead of four full-width nodes using six-core chips to get there. IBM could also be deactivating cores to have them as capacity on demand. Again, IBM was not clear.

The W1700 machines above may sound pricey, but recall that an Exadata X2-2 database cluster would cost $4.47 million at list price with 96 cores. (I don’t have Exadata X2-3 pricing handy.) None of these machines are inexpensive, and as much as IBM talks about it, they are not really aimed at small businesses or even most midrange shops. Just like Exadatas are not. The PureApplications range in price from $30,938 to $27,911 per core, fully burdened, if you do the math. That is less expensive per core than buying just AIX plus DB2 or just IBM i (which includes DB2 for i) for a big bad Power 780 or 795 SMP system–without the price of the iron even accounted for in the SMP machines. So the message is clear, and one that Oracle has been hammering on for years: If you can run your apps on a virtualized cluster instead of on a big SMP machine, you should.

I have been saying this so long that I am blue in the face, but I will say it again. IBM needs to create a baby PureApplication system that runs IBM i and uses the DB2 Multisystem clustering that has been around forever to glue together nodes into a single database image as the PureScale extensions to DB2 and AIX do in the PureData transaction processing appliances. And IBM needs to also make these beefier PureApplication systems available running IBM i with the same relative pricing to big SMP boxes that AIX/DB2. Any strategy that is good to try to steal Oracle and HP customers and grow the AIX/DB2 base is also good to retain and growth the existing IBM i base.

The conventional wisdom is that it takes 10 times as much money and effort to gain one customer than it does to keep one. We’re down something on the order of 125,000 customers–nearly half of the peak OS/400 base in 1999. Keeping the 150,000 we have happy, and giving IBM i shops all the same goodies and benefits, won’t make headlines like taking on Oracle. But it sure would make the lives of Oracle, HP, Dell, and now Cisco Systems a hell of a lot harder when they try to poach IBM i shops.