Performance analysis of wavefront algorithms on very-large scale distributed systems

Abstract

DISCLAIMERPortions of this document may be illegible in electronic image products. Images are produced from the best available original document.Abstract. We present a model for the parallel performance of algorithms that consist of concurrent, twodimensional wavefronts implemented in a message passing environment. The model combines the separate contributions of computation and communication wavefronts. We validate the model on three important supercomputer systems, on up to 500 processors. We use data from a deterministic particle transport application taken from the ASCI workload, although the model is general to any wavefront algorithm implemented on a 2-D processor domain. We also use the validated model to make estimates of performance and scalability of wavefront algorithms on 100-TFLOPS computer systems expected to be in existence within the next decade as part of the ASCI program and elsewhere. On such machines our analysis shows that, contrary to conventional wisdom, interprocessor communication performance is not the bottleneck. Single-node efficiency is the dominant factor.

Scalable performance analysis of exascale MPI programs through signature-based clustering algorithms

“…Sweep3D is a stencil code in which each process must wait for boundary information from neighboring processes to the north and west before computing values within its subdomain [8]. Similar to Sweep3D, LU is also a stencil code [18] that creates nine different main clusters.…”

Using automated performance modeling to find scalability bugs in complex codes

“…The literature mentions accurate models [19,35] that describe the performance behavior of wavefront processes as they occur in Sweep3D on various architectures. The LogGP model reported in [19] characterizes the communication time as follows:…”

“…Because the amount of waiting time in Sweep3D, which is responsible for the bulk of the time spent in MPI at larger scales, depends on the progress of the wavefront computation, earlier studies [19,35] concluded that single-node performance is the most serious impediment to the scalability of Sweep3D-and not, for example, the saturation of network resources. To see whether we arrive at the same conclusion using our automated approach, we also conducted experiments on Juropa, whose cores are much more powerful than Juqueen's.…”

GPU-based dynamic search on adaptive resolution grids

“…The work in [5] presents a model for analyzing performance of wavefront algorithms. [6] proposes a focused wavefront expansion for path planning, where the expansion of the algorithm is directed towards the solution instead of being spread across the entire environment.…”