Features:

INTERVIEW WITH BILL BLAKE, SVP, PRODUCT DVPT., NETEZZA
by Alan Beck, Editor-in-Chief, HPCwire

HPCwire: Thus far, the race between COTS-based cluster supercomputers and those based upon proprietary processors has resembled that of Achilles and the tortoise: the clusters approach ever nearer but never quite succeed in surpassing -- or indeed drawing even with -- their elite contenders. Will this situation ever change? Why or why not? And given the speeds involved, will it still matter?

BLAKE: Lessons learned from the development of proprietary processors have tended to flow down to the industry standard parts, much like the lessons learned on formula 1 racing cars have influenced "commodity cars". And today's crop of 64 bit industry standard processors are certainly catching up to where proprietary processors such as Alpha were just a few years ago. But the key factor may well be the economics, since innovation in areas such as new approaches to on-chip paralellism in microprocessors can be a billion dollar proposition in order to significantly exceed the industry standard parts.

Does it matter? Yes if the industry standard parts do not support the dramatically higher memory bandwidth requirements of supercomputing then it matters a lot. Architectural approaches such as hypertransport are very important to opening up the memory system of the processor to high performance system interconnects that are the lifeblood of scalable parallel systems. At the system level, clusters will clearly dominate as all the hardware and software building blocks are there to deliver significant application performance with very good price/performance characteristics.

Linux clusters are the mainstream, and their elite contenders as you call them are relegated to those specialized cases where highest capability is required. As for capacity computing, the important load sharing software, be it from Platform Computing Inc. or many home grown varieties, is in place to support excellent system utilization is in use everywhere. The key enabling technology for COTS-based cluster supercomputers for parallel compute intensive applications has been tools like MPI for coarse grained message passing plus a lot of work by parallel application developers to deal with explicit parallelism in their codes.

HPCwire: What kinds of networking architectures will provide the principal support for the simplified supercomputers of the future? Will such networks ultimately prove as unwieldy, in their own way, as traditional HPC vector processors? Why or why not?

BLAKE: Myricom and Quadrics are both setting the bar for all others to meet in terms of bandwidth and latency. And I expect them to continue in that mode for the forseeable future especially if they can continue to exploit new high bandwidth memory interfaces such as hypertransport. But there are a number of new startups, such as Alacritech and Amasso, that are trying to improve bandwidth and lower the latency of the standard ethernet stack and if they are successful then they will cannibalize the proprietary schemes. By mid to late decade, the horse race at the high end will be between proprietary all-optical switches in conventional topologies like the fat tree and highly optimized mesh architectures built into the microprocessor itself (but without the overhead of maintain cache coherence in large meshes).

HPCwire: Will the new strategy of simplified computing truly provide HPC power for general use -- or will security concerns eventually eclipse the enormous potential that appears to lie just ahead?

BLAKE: I expect the grid forces to solve the security issues needed to make simplified computing truly available for general use. For example, the success Avaki has had with secure data grids for the pharmaceutical companies is very encouraging.

HPCwire: Any other surprises at the system architectural level?

BLAKE: Absolutely. As cluster supercomputers mature, we are seeing significant innovation at the system architectural level, especially in the specialization of node function. New systems such as the Cray/Sandia Red Storm show an interesting approach to both scalability and serviceability with cluster nodes optimized for compute (very light weight kernel) versus file system (full Linux) versus service and maintenance. By solving the software challenges of heterogeneous and asymmetric node configurations, system performance and functionality will improve. At Netezza we are pursuing that path for the hardware needed to support analytic terascale databases as we couple a front end SMP machine to a highly parallel database engine with nodes optimized for database operations with processing as close to the disk where data resides as possible.