The documents distributed by this server have been provided by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.
Publications of SPCL
|G. Bauer, T. Hoefler, W. Kramer, B. Fiedler:|
|Analyses and Modeling of Applications Used to Demonstrate Sustained Petascale Performance on Blue Waters|
(Presentation - presented in Stuttgart, Germany, May 2012, Cray User Group )
AbstractThe sustained petascale performance of the Blue Waters system will be demonstrated using a suite of several applications representing a wide variety of disciplines important to the scientific community of the US National Science Foundation. The geometric mean of the measured floating point rates for these applications running scientific problems of current interest at scale is used to compute the sustained petascale performance (SPP), which is a key acceptance metric for the system. In this report, we discuss the performance of these applications on Cray XE hardware. Our elemental modeling methodology splits each of the codes into a small set of performance-relevant kernels (typically around 5-7). We analyze those kernels in detail on the AMD Interlagos architecture to determine the achieved memory bandwidths and latencies, communication bandwidths and latencies, and floating point rates. This allows us to conclude, in a form similar to the Roofline model, how well each kernel performs with regards to each of the parameters. For example, if a kernel is mediocre in peak performance but utilizes 100% of the memory or network bandwidth, we can conclude that the kernel is utilizing the architecture well and is memory- (or communication-) bound. While a low floating point rate and a low bandwidth utilization suggests optimization opportunities. Our analyses should also provide us with insight into application performance on future systems.