Petascale computing with accelerators

Abstract

A trend is developing in high performance computing in which commodity processors are coupled to various types of computational accelerators. Such systems are commonly called hybrid systems. In this paper, we describe our experience developing an implementation of the Linpack benchmark for a petascale hybrid system, the LANL Roadrunner cluster built by IBM for Los Alamos National Laboratory. This system combines traditional x86-64 host processors with IBM PowerXCell™ 8i accelerator processors. The implementation of Linpack we developed was the first to achieve a performance result in excess of 1.0 PFLOPS, and made Roadrunner the #1 system on the Top500 list in June 2008. We describe the design and implementation of hybrid Linpack, including the special optimizations we developed for this hybrid architecture. We then present actual results for single node and multi-node executions. From this work, we conclude that it is possible to achieve high performance for certain applications on hybrid architectures when careful attention is given to efficient use of memory bandwidth, scheduling of data movement between the host and accelerator memories, and proper distribution of work between the host and accelerator processors.

Optimizing Linpack Benchmark on GPU-Accelerated Petascale Supercomputer

“…Since the same matrix between tasks can be reused, the order of the four tasks is like T 0, T 1, T 3, T 2 by using the "bounce corner turn" [18] method. When T 1 is executed, matrix A 1 does not need to be transferred, neither do B 2 for T 3 and A 2 for T 2.…”

“…Using Cell accelerators [26] , in 2008 IBM built the first heterogenous petascale supercomputer called Roadrunner [18] . This system was very different than a GPU-accelerated system.…”

Accelerating LINPACK with MPI-OpenCL on Clusters of Multi-GPU Nodes

“…A node in such heterogeneous clusters was typically built with multicore CPUs and a single accelerator (e.g., a GPU). Recently, more and more cluster systems have started to have multiple accelerators per node to deal with large size problems [1], [2], [3]. Multiple accelerators per node may enlarge the benefits of a heterogeneous system, especially for massively data-parallel applications.…”

“…Moreover, it is usually used as a yardstick of the performance of supercomputers because the TOP500 supercomputer list [10] ranks supercomputers by their performance on the LINPACK benchmark. The LINPACK benchmark requires 2 3 n 3 þ Oðn 2 Þ double-precision floatingpoint operations to solve a system of linear equations of order n. Reducing the operation count (e.g., using the Strassen algorithm for matrix multiplication) is not allowed. Under this constraint, any optimizations can be applied to the algorithm in order to achieve the best performance for the target system.…”

Efficient SIMD solution of multiple systems of stiff IVPs

“…processing elements (SPEs). The CBE is currently used in scientific computing on both large [2][3][4] and small scales [5,6] due to its high floating-point throughput. The CBE allows SIMD instructions to be used without resorting to assembly language and provides a great deal of programmer control over memory management.…”