Evaluating Application Performance on Big Iron

Four years ago, a group of scientists from the U.S. Department of Energy's Lawrence Berkeley National Laboratory and Princeton Plasma Physics Laboratory began to undertake what would become one of the most comprehensive performance evaluations of supercomputers. The project, led by Lenny Oliker at Berkeley Lab, has produced findings that not only illustrated the strengths and weaknesses of various high-performance computing architectures, but also pinpointed bottlenecks in scaling applications for petascale computing down the road.

The team ran six scientific codes that represented a wide range of research disciplines on each of the six supercomputers. The systems were Bassi and Jacquard from Lawrence Berkeley National Laboratory, Jaguar and Phoenix from Oak Ridge National Laboratory, Blue Gene/L from Argonne National Laboratory and another Blue Gene/L from IBM Thomas J. Watson Research Center.

Bassi is an IBM Power5 system with 888 compute processors (111 8-way nodes) that runs on AIX. Jacquard contains 640 single-core AMD's Opteron processors (320 2-way nodes), running Linux 2.6.5., while Jaguar features 14,400 dual-core Opteron processors (5,200 2-way nodes) and running Catamount. Pheonix is a vector-based Cray X1E system containing 768 processors (96 8-way MSP nodes) and runs UNICOS.mp. The Blue Gene/L at Argonne is an IBM PowerPC 440-based system with 2,048 processors (1024 2-way nodes) and runs SuSE Linux OS (SLES9). The Blue Gene/L at IBM's research center contains 40,000 processors.

The codes had been used in six areas of research: magnetic fusion (GTC), fluid dynamics (ELBM3D), astrophysics (Cactus), high-energy physics (BeamBeam3D), materials science (PARTEC) and AMR gas dynamics (HyperCLaw).

After publishing in journals and conferences and garnering kudos along the way, Oliker and his team is set to win another recognition. The IEEE International Parallel and Distributed Processing Symposium plans to present the scientists with a Best Paper award in the applications track on March 28 in Long Beach, California.

The paper's co-authors are Andrew Canning, Jonathan Carter, Costin Iancu, Michael Lijewski, Shoaib Kamil, John Shalf, Hongzhang Shan and Erich Strohmaier from Berkeley Lab, Stephane Ethier from the Princeton Plasma Physics Laboratory and Tom Goodale from Louisiana State University.

Oliker, whose other research interest includes examining the possibility of using embedded processor technologies to develop energy-efficient supercomputers, sat down recently to talk about the research and what got him started on evaluating these high-performance systems.

HPCwire: What contributions does your research make to the HPC community?

Lenny Oliker: If you go to the literature today and look for comparisons on how the machines differ from one another, there is surprisingly little information out there. Detailed application analysis is critical as it reveals algorithmic limitations to scalability and cases where a particular code is poorly suited to an architectural feature. Head-to-head comparison of applications across architectures provides invaluable insight into the suitability of a given type of machine for a particular class of scientific method. Our current goal is to understand limitations of scaling into the petaflop regime, as this is a very important thrust in HPC today.

HPCwire: Challenges along the way?

Oliker: In order to have a good experiment, we needed to have realistic scientific applications. Ones that scientists say, yes, this is how we would run it. It's a moving target when you work on different applications which are constantly being developed and refined. This is an effort that requires a large group of folks from disparate backgrounds in order to make the numbers meaningful. A key part of this is in gathering codes from a variety of computational areas. Also, you want to feel like you are doing justice to each of the different flavors of architectures. As you move codes from machine to machine, it's not just a process of running the codes and entering different numbers. You are looking at problems that can be addressed for each type of the machines and thinking about how to best use those resources.

HPCwire: What prompted you to undertake this project?

Oliker: The way this project got started was with Earth Simulator. When it went on line in 2002, it was the most powerful computer in the world and stayed on the TOP500 list five consecutive times. It was a shakeup of the United States supercomputing community. I was part of the U.S. team that was able to gain access to Earth Simulator, and I took several trips to Japan. It started a great collaboration with a number of application scientists here at Berkeley Lab as well Livermore Lab and Princeton. Once we generated baseline data based on these evaluations, we've continued exploring the latest supercomputing trends, with numerous publications along the way.

HPCwire: What are the key findings?

Oliker: We have numerous results demonstrating the tremendous potential of modern parallel vector systems. But we also found that the imbalance in speed between scalar and vector core on the Cray X1 has a big negative impact on the performance of some other applications. In terms of superscalar systems, our studies show encouraging data that the slide in the sustained performance of microprocessor cores is not irreversible if architects are willing to invest the effort to address the bottlenecks of scientific applications. For instance, the IBM Power5 often shows execution efficiency over its Power3 and Power4 predecessors, thanks to dramatically improved memory bandwidth and increased attention to latency hiding through advanced prefetch features.

HPCwire: What's next?

Oliker: Our work is a moving target, as there is constant evolution in the application algorithms and supercomputing systems. Currently, we are in the midst of a major paradigm shift, as multi-core processors are increasingly used as HPC building blocks. Understanding how to best utilize vast numbers of simpler processors will be a major challenge to both high-end computing and the commercial world. We are also interested in broadening our code base to include the increasingly complex numerical approaches being employed, such as adaptive mesh refinement techniques. Although many methods used today rely on regularly structured computations, emerging multi-scale applications will require irregular and dynamically evolving simulations. We plan to continue our research to help characterize the behavior of these emerging techniques.

Read more about Oliker's research.