Ruud van der Pas
Sun Microsystems

Title: Evaluating the throughput performance of the heavily threaded UltraSPARC T2 multicore processor

The focus of this talk is on the recently announced UltraSPARC T2 processor (http://www.sun.com/processors/UltraSPARC-T2), the successor of the very successful UltraSPARC T1 design.

Compared to its predecessor, UltraSPARC T2 has several major enhancements. One of the new features is that each of the 8 cores is now equipped with a fully fledged floating-point unit, capable of delivering 1.4 Gflop/s. This gives an aggregate floating-point performance of 11.2 Gflop/s on a single processor. Each core also 8 hardware threads, making this a very powerful and interesting processor for HPC.

An obvious question to ask is how much performance this new processor delivers.

To tap the full potential, one has to either run multiple independent jobs simultaneously, or a parallel application, or a mixture.

In our study, we looked at the throughput performance in more detail. To this end, we used an in-house test suite consisting of a variety of typical HPC applications. Several throughput benchmarks were conducted using this suite.

We compared the measured performances with the theoretically predicted values. As will be shown, the results are rather astonishing. They not only clearly demonstrate the hardware threads boost performance, but it also raises the interesting question how to evaluate these kind of heavily threaded multicore designs. A methodology for this is presented. The theoretical estimates obtained are in very good agreement with the observed values.

About Ruud van der Pas

Ruud van der Pas is a Senior Staff Engineer in the Software Organization at Sun Microsystems and operates on a world-wide basis. He has been with Sun since 1998, providing application tuning consultancy to customers and working closely with engineers to discuss enhancements to current and future products. His main expertise is in serial performance, shared memory parallelization using OpenMP and application performance analysis.

Ruud has studied mathematics and physics and has been involved with the performance aspects of High Performance and Technical computing since 1985. His experiences cover a great variety of architectures, including the Cray-1 and Cyber 205 vector supercomputers, clusters of workstations, vector shared memory systems, plus cache based SMP and cc-NUMA architectures. Prior to joining Sun, he worked at SGI, Convex Computer Corporation, the University of Utrecht and Philips.

Ruud regularly gives technical presentations and workshops at conferences and seminars all over the world.