As we explained earlier this week, DB2 PureScale makes use of InfiniBand clustering to link multiple server nodes equipped with AIX and DB2 together. The secret sauce is that it has a designated server in a cluster - which functions similar to a head node in a parallel supercomputing cluster - to manage the locking of database fields as transactions are processed, and the locking and unlocking of memory in all of the nodes in the cluster as they seek information from each other as part of the OLTP cranking.
Because the server nodes are linked to each other and to the master PureScale node through the Remote Direct Memory Access (RDMA) features of InfiniBand, the processors are basically cut out of the networking stack, unlike TCP/IP clustering techniques. The central caching server is mirrored so it is not a single point of failure, and radically cuts down on the intra-node communications that normally happen in a parallel database implementation, according to Spang.
Interestingly, the server nodes in a PureScale setup all link to each other through the 12X remote I/O port on the Power Systems servers. The 12X I/O port is a variant of double-data rate InfiniBand that IBM has tweaked to allow remote I/O drawers (laden with disk controllers, disk drives, tape drives, and so on) to be lashed back to central servers.
A decade ago, IBM took a variant of similar remote I/O drawer technology called OptiConnect and a parallel database clustering technology called DB2 MultiSystem to create a parallel cluster of AS/400 proprietary minis that presented a single database view to applications, even though it was running on a cluster. (Hmmm.)
DB2 MultiSystem generated about as many customers as press releases, and IBM stopped talking about it. We'll see if this doesn't happen again with PureScale.
For now, according to Scott Handy, vice president of marketing and strategy for the Power Systems division, IBM is only offering the PureScale feature of DB2 on Power 550 servers, which have from two to eight Power6+ cores in the most recent iterations, and on Power 595 boxes, which have from eight to 64 Power6 cores running at 5GHz. IBM is not offering a blade configuration, which seems odd.
The PureScale feature is only being enabled on the AIX 6.1 operating system and only with the DB2 V9.7 database, and Handy wouldn't say what IBM's plans were for Linux, Windows, or i/OS. Both Linux and i/OS run on Power iron, and DB2 V9.7 runs on Linux and Windows operating systems, so it would seem that Linux should be a snap to support and putting PureScale on i/OS's own implementation of DB2 and enabling it for Windows running DB2 V9.7 would be relatively easy.
According sources that spoke to El Reg who are familiar with IBM's plans, PureScale is going to be ported to Windows and Linux using DB2 databases, but will not be ported to the HP-UX or Solaris versions of DB2.
IBM is being fairly tentative - at least publicly - about its plans for the PureScale clustering technology not just because clusters that look and feel like SMP boxes are relatively new to end users, but no doubt because it is worried about the potential impact clusters will have on the sale of SMP iron.
It has always been cheaper to build and buy a cluster of cheaper servers than an SMP box of equivalent raw oomph, but the cluster administration and application and data tweaking was a costly nightmare. If what IBM says about PureScale is true - that it can scale linearly up to a hundred server nodes or more - then PureScale could take a chunk out of the big Unix and mainframe boxes IBM sells. And equally frightening for the Big Blue sales team is the prospect that customers stop doubling up on their servers for high availability clustering.
Go for configurations
Of course, if IBM doesn't sell the database clusters, Oracle and Sun surely will try. And at many accounts, they may very well succeed in pitching against IBM's big iron, even with PureScale in the mix. It all depends on the bang for the buck and the ease of management.
Spang would not divulge IBM's plans to run benchmark tests on a PureScale setup, but it seems likely that Oracle is going to put Sun's x64 and Sparc clusters through the paces and hammer on IBM.
The DB2 PureScale feature will be available in December. Pricing will be released at that time. All of the AIX iron that it requires is available now. Presumably IBM will cook up some sample configurations for customers to look at as they buy components, and if IBM was smart, it would put together a few configurations with a single product number, as Oracle has done with the Exadata V2 and as IBM has done with the Smart Analytics System. ®
pureScale has been worked on for over 4 years now
Exadata V1 was likely not quite dreamed up yet when the work on pureScale was started - 4+ years ago - just as scotdb said.
It takes years to build this kind of technology into a product - even when you are borrowing the architecture from another platform.
Oracle RAC is yesterdays technology that was designed on a message passing architecture through sockets - using vast quantities of CPU horsepower to service the context switches and interrupts that such communication entails.
Oracle has gone out of their way to try to get vendors to get emerging technologies like InfiniBand fit into their crappy design by insisting on RDS - so that it looks and feels like sockets - so that they don't have to make the deep investment to exploit RDMA and IB in its native form through uDAPL or MPI - from usespace to userspace - no interrupts - no wasted memcpy into the kernel and system calls etc.
RDMA is here to stay and will be mainstream on 10G ethernet as well shortly. Oracle is behind the curve on this on.
Who do you think invented Parallel Sysplex in 1993? The same company that is introducing it in their Power systems.
Not aimed at Exadata but at RAC !!!
Now that this is officially announced, you'll see that this has nothing to do with Exadata but at those who want to move from RAC to something more scaleable.
And I can assure you that this wasn't "cooked up in the last few weeks" since I was at a briefing on this in June during IOD Europe, and at that point they''d been working on this for a couple of years.
The obvious comparison is with DB2 for z/OS Data Sharing. The underlying architecture and even the terminology being used is the same, and no doubt there is a lot of common code too.
The big benefit of Data Sharing on System Z has always been the Coupling Facility which prevents having to broadcast changes to all the members all of the time. That's where the scaleability comes from. Getting this on AIX is awesome.