A plug-and-play model for evaluating wavefront computations on parallel architectures

Abstract

This paper develops a plug-and-play reusable LogGP model that can be used to predict the runtime and scaling behavior of different MPI-based pipelined wavefront applications running on modern parallel platforms with multicore nodes. A key new feature of the model is that it requires only a few simple input parameters to project performance for wavefront codes with different structure to the sweeps in each iteration as well as different behavior during each wavefront computation and/or between iterations. We apply the model to three key benchmark applications that are used in high performance computing procurement, illustrating that the model parameters yield insight into the key differences among the codes. We also develop new, simple and highly accurate models of MPI send, receive, and group communication primitives on the dual-core Cray XT system. We validate the reusable model applied to each benchmark on up to 8192 processors on the XT3/XT4. Results show excellent accuracy for all high performance application and platform configurations that we were able to measure. Finally we use the model to assess application and hardware configurations, develop new metrics for procurement and configuration, identify bottlenecks, and assess new application design modifications that, to our knowledge, have not previously been explored.

Toward Performance Models of MPI Implementations for Understanding Application Scaling Issues

“…One particular example is that several application models assume that broadcast or allreduce scale with Θ(S log(P )) (e.g., [3,17]) while, as demonstrated in Section 4, a good MPI implementation would implement a broadcast or allreduce with Θ(S + log(P )) [5,13,21]. Generally speaking, performance models for middleware libraries such as MPI depend on the parameters of the network (e.g., bandwidth, latency, topology, routing) and the implemented algorithms (e.g., collective algorithms, eager and rendezvous protocols) and are thus hard to generalize.…”

Predictive analysis of a hydrodynamics application on large-scale CMP clusters

“…An iteration of Hydra employs several parallel functions, each of which consists of a number of the above operations. An aim of this work is to capture the time to solution by modelling the critical-path run-time of the code as demonstrated in previous analytic modelling research (Mudalige et al, 2008). We develop a general analytic model for the first two key operations (local computation, near-neighbour communication) before applying these to specific segments of the Hydra code.…”

“…From the number of inter-and intra-node connections (equations 11 and 12), we can derive a model for near-neighbour communications in a similar fashion to that found in Mudalige et al (2008). This makes the assumption that the communication network is full-duplex and that the time for two nodes to perform a non-blocking send and receive is equivalent to the time for a single blocking send and receive because of Hydra's use of MPI_Waitall.…”

“…The development of such HPC codes, the evaluation of their performance on candidate systems and, sustaining performant execution, is a costly and time consuming exercise. To aid in these activities research has been conducted into developing accurate performance modelling tools and techniques for application analysis (Sundaram-Stukel and Vernon, 1999;Hoisie et al, 2000;Kerbyson et al, 2001;Mathis and Kerbyson, 2004;Mudalige et al, 2008).…”

Predictive Evaluation of Partitioning Algorithms through Runtime Modelling

“…Many HPC centres are therefore turning to alternative tools and methodologies (e.g. predictive performance modelling [1], [2], hardware simulation [3], [4] and mini-applications [5], [6]) to facilitate system evaluation, to aid in the comparison of multiple candidate machines, to investigate optimisation strategies, and to act as a vehicle for porting codes to novel architectures.…”