IBM’s Smartie and Pizzazz Clusters–Still i-Less
April 12, 2010 Timothy Prickett Morgan
IBM and its peers in the IT racket have a way of making even neat technology sound boring with the public product names they choose, even if they do show a little flair with product names from time to time when those products are in development. So it is with data analytics and online transaction processing clusters that Big Blue rolled out last week, which I am rechristening the Smartie and Pizzazz systems because, well, because I just can’t stand typing Smart Analytics System and PureScale Application Server every time I mention the machines.
It is no secret that Oracle has been pushing clusters as an alternative to midrange and high-end servers based on tighter symmetric multiprocessing (SMP) or non-uniform memory access (NUMA) clustering, technologies that have been widely deployed in RISC servers for nearly two decades and for nearly that long in X86 and X64 machines. Creating a shared memory space for operating systems, databases, and applications to frolic in is expensive, and high availability means doubling up the systems and clustering them. So you end up with a cluster for mission-critical applications no matter what you do.
When Compaq was desperate for money in the wake of buying Digital Equipment more than a decade ago, it licensed the key clustering glue that made its VAXcluster and TruCluster Unix variant interesting, unique, and valuable to Oracle, which used that technology to take a pretty awful clustered implementation of its database, called Oracle Parallel Server, and turn it into Oracle Real Application Clusters. Which, by the way, is useful and really does support both data warehousing and OLTP workloads with some scalability limits on the latter, according to people I know who know these things. Oracle’s RAC extensions to its 11g database are at the heart of the Exadata V2 database appliance, which I told you about last September. Basically, it is a bunch of Sun Microsystems X64 blade servers and storage arrays mixing lots of flash and lots of disk drives married to Oracle RAC plus some storage management software called Exadata from Oracle. (Oracle owns Sun now, so it is all Oracle technology at this point.)
Rather than cook up a single product to take on a much more serious Oracle, which owns servers, storage, and switches and intends to use them to compete with IBM, Hewlett-Packard, Dell, Cisco Systems, and others, IBM is taking a two-prong approach with the Smartie and Pizzazz clusters.
The Smartie systems are clusters of Big Blue’s servers (or clusters of logical partitions on mainframes, in the case of one machine announced last week) equipped with its InfoSphere Warehouse data warehousing software and its Cognos 8 data analytics tools. The Smartie boxes do not yet have the predictive analytics software IBM gained last year with its $1.2 billion acquisition of SPSS, but that software should be bolted onto the Smartie clusters soon. (It is a wonder that it hasn’t happened already, in fact, but mergers of companies and roadmaps take time.) The original Smartie boxes from last summer clustered up Power6-based Power 550 servers running AIX 6.1, DS3500 storage arrays, and InfiniBand switches. Last week, IBM gave this product a more formal name, the Smart Analytics System 7600, and promised that it would soon update the box with its new Power7-based systems and, presumably, the SPSS tools. IBM did not say what Power7 machines would go into the Smartie 7600s, but my guess is the four-socket Power 750s, not the eight-socket Power 770 and 780 boxes.
IBM also rolled out the Smartie 9600, a mainframe variant of the data analytics setup that can be acquired new atop a System z10 Business Class mainframe in six different performance ranges. IBM plunks z/OS and DB2 in one partition and the mainframe-Linux versions of the InfoSphere Warehouse and Cognos tools in another partition. IBM is also allowing customers to buy just the software and create a Smartie 9600 setup on partitions on existing machines. Big Blue rebranded an existing appliance, known as the InfoSphere Balanced Warehouse D5100, as the Smartie 5600, which became “smart” because the cluster of System x3650 M2 servers (using last year’s quad-core Xeon 5500 processors and running Linux) now has Cognos 8 in the stack.
Pricing is rarely divulged on such things as the Smartie systems, and last week was no exception. What I can tell you is that installation and configuration come in the price tag, whatever that might be.
What I can also tell you is that there is just an outside chance that IBM will listen to reason and create variations of the future Smartie 7600 systems based on i 7.1 and the integrated DB2 for i database and give OS/400 and i shops a little love. It would be nice if the InfoSphere Warehouse and Cognos software could run natively in conjunction with DB2 for i and on i 7.1, but if not, put them on Linux or AIX partitions on a Power Systems machine. The key customers in banking, finance, insurance, wholesale and retail distribution, telecom, and manufacturing that rely on the IBM i platform have a need for data and predictive analytics, and there is no good reason they should have to run another database to get the job done.
The same holds true for the PureScale Application System–or PSAS, or “pizzazz” if you want to be polite about how to pronounce that abbreviation–that was announced by Big Blue last week as well. This is a cluster of Power7 machines running AIX 6.1 and the forthcoming PureScale clustering features that were to debut with DB2 V9.8 but which were apparently backcast into DB2 V9.7 for AIX last September for delivering in December. The point is, DB2 for i has DB2 Multisystem and DB2 SMP extensions since the mid-1990s, and there is no good reason why IBM can’t create PureScale clusters that employ these technologies. In fact, I suspect, based on the description of the PureScale clustering technology that I gave you last fall when the AIX product was announced running on clustered Power 550s, PureScale borrows rather heavily from the DB2 Multisystem/OptiConnect clustering that OS/400 shops had to build big data warehouses and OLTP setups long before IBM’s DB2 for AIX and Windows could even contemplate it.
In any event, it was with a certain amount of wry amusement I saw the Application System tacked onto the PureScale moniker, given the AS/400 is the Application System/400. As I complained when IBM rushed out the PureScale database clustering feature for AIX 6.1 and DB2 V9.7 on Power 550 and Power 595 servers last September, trying to steal some thunder from the Oracle-Sun Exadata V2 launch, IBM was missing the whole point in that it was selling a software feature loaded on specific machines and Oracle was peddling a complete system, ready to go for OLTP and data warehousing applications to run on.
IBM seems to have listened to the reasoned arguments that many of us made, no doubt including a lot of sales reps and resellers who have to make a living selling IT solutions and who carry a lot more weight than smartass editors of newsletters. (We all do our part in this ecosystem, mind you.) So now the Pizzazz is a box, not a feature, with a specific price (which has not been announced yet, of course).
Because OLTP workloads behave in different ways, IBM is basing the Pizzazz cluster on two of the new Power 770s, linked by an InfiniBand switch through the 12X remote I/O ports and making use of InfiniBand’s Remote Direct Memory Access (RDMA) feature to create a database locking mechanism that acts like a funky system backplane to turn the Power 770s into what, as far as applications are concerned, is a single database. (Just like the DB2 Multisystem/OptiConnect links did on the AS/400 so long ago.)
The base Power 770 in the Pizzazz cluster comes with two sockets fully populated with eight-core Power7 chips running at 3.1 GHz, but only four cores (two per socket on the system board) are activated. The Power 770s, like their Power 570 predecessors, can be scaled from one to four nodes using NUMA clustering and up to 512 GB using current memory cards from IBM and as far as 2 TB with memory cards IBM hopes to ship in November. If the database behind an AIX/DB2 OLTP workload likes fat nodes in a clustered configuration, then the Power 770 nodes can have processors turned on and server nodes added. Additional Power 770 nodes can be added to the cluster to scale out the OLTP workload, and it can scale very far indeed.
According to Bernie Spang, director of product strategy for database software and systems at IBM, IBM has been able to show that 64 Power Systems nodes using the PureScale database extensions have been able to scale to 95 percent efficiency, and that even on 128-node setups, the cluster runs at about 85 percent efficiency. These numbers mean that a 64 node cluster delivers the same OLTP throughput as 61 individual nodes would if you aggregated their performance mathematically (obviously, you need some sort of clustering technology to accomplish this technically); a 128-node system is yielding the performance of about 109 nodes measured individually. This is the kind of scalability that SMP and NUMA systems with supposedly tighter coupling can deliver.
This is why I have been advocating for IBM to stop just building big SMP systems and to also build clusters using DB2 Multisystem to give OS/400 and now i shops a competitive alternative to big and expensive SMP boxes and something to sell against cheap Windows boxes in the midrange. Giving customers a choice seems pretty obvious to me, but then IBM would not be able to push customers into more expensive iron, just like Oracle is saying IBM, Hewlett-Packard, Sun, Fujitsu and other big iron players always want to do.
Oracle is right, at least for a certain set of customers, and the sooner IBM’s Power Systems people realize this and give i shops a clustered alternative, as it is giving AIX shops, the sooner it will get some of its business back. I would start with the Power 720s for a clustered Pizzazz box running i 7.1, and maybe put a clustered Power 710 into the field with Pizzazz configurations, too. In this midrange market, you want to start small, not large.
For the Smart Analytics Systems:
IBM turns back on server history: To give and to hybrid (The Register)
For the DB2 PureScale Clusters: