IBM's Exadata2 knock-off could use SVC and flash
Pure Scale InfiniBand-connected SVC
Comment IBM's DB2 Pure Scale project has a storage problem - storage arrays don't make good cluster nodes.
Pure Scale is the rumoured Exadata 2-killer from IBM. It is thought to be a clustered implementation of IBM's DB2 database running on Power-based servers with the AIX O/S, and an InfiniBand cluster interconnect between server and storage nodes.
It will use InfiniBand's Remote Direct Memory Access (RDMA) features to give nodes access to each other's data for processing. There will be a designated head node server to manage the locking of database fields as transactions and queries are processed and the locking/unlocking of memory of all nodes in the cluster.
Fair enough. The problems we at El Reg Analyst Towers see is that a) your bog-standard storage array controller is not a cluster node, b) it doesn't have an InfiniBand (IB) link, and c) Exadata 2 relies on flash for its performance. Step forward IBM's SAN Volume Controller (SVC).
This is the dedicated box that sits in a SAN fabric, virtualises the SAN storage arrays and presents them to accessing servers as a virtual storage pool. In Project Quicksilver, IBM and Fusion-io demonstrated a million IOPS from an SVC set-up accessing lots of Fusion-io's ioDrive NAND flash solid state storage cards, 40 of them in fact.
Texas Memory Systems' RamSan flash storage is also available for the SVC.
That could provide the missing flash ingredient. What about an IB link? IBM'r D S Guthridge produced a paper last year entitled "Scalable, high performance InfiniBand-attached SAN Volume Controller" that described how: "Large read performance from SVC cache exceeds 3GB/sec in a minimal two-node cluster configuration," with an InfiniBand host attached system.
If IBM were to somewhat hurriedly cook up an Oracle Exadata knock-off product then IBM-connected SVC nodes with Fusion-io or TMS flash could be part of a back-of-an-envelope hardware spec. ®
Are you forgetting about XIV?
RE: "a) your bog-standard storage array controller is not a cluster node, b) it doesn't have an InfiniBand (IB) link"
RE: Stephen Channell
Much as I enjoyed your Soreacle-baiting post, the potential problem I see with the IBM approach is that it is Power-based, which will probably make it expensive compared to the "commodity" Galaxy servers that Oracle plan to use in Exadata2. This may be why Larry went with an announcement of an x64-based Exadata2 rather than SPARC, maybe he already knew IBM had something in the works? Maybe he did or maybe it's just a lucky choice, but I tend to think x64 will have an edge in appeal to shops that don't want the burden of paying out for AIX admins when the Oracle offering just looks more of an "appliance". If IBM can get the pricing right, and at the same time resist the impulse to turn each sale into an IBM Global Services engagement, then they may actually have an Exadata2 denter if not an outright killer.
Rush something to counter the EXADATA2 benchmark?
something tells me that it's highly unlikely IBM could do something in the last 6 or 7 weeks to counter whatever Oracle could be announcing. Whatever it is would have either taken years of work prior to the announcement, or just a phantom announcement with no substance.