Multicore-enabling the MPJ express messaging library

With the transition to multicore processors almost complete, the parallel processing community is seeking efficient ways to port legacy message passing applications on shared memory and multicore processors. MPJ Express is our reference implementation of Message Passing Interface (MPI)-like bindings for the Java language. Starting with the current release, the MPJ Express software can be configured in two modes: the multicore and the cluster mode. In the multicore mode, parallel Java applications execute on shared memory or multicore processors. In the cluster mode, Java applications parallelized using MPJ Express can be executed on distributed memory platforms like compute clusters and clouds. The multicore device has been implemented using Java threads in order to satisfy two main design goals of portability and performance. We also discuss the challenges of integrating the multicore device in the MPJ Express software. This turned out to be a challenging task because the parallel application executes in a single JVM in the multicore mode. On the contrary in the cluster mode, the parallel user application executes in multiple JVMs. Due to these inherent architectural differences between the two modes, the MPJ Express runtime is modified to ensure correct semantics of the parallel program. Towards the end, we compare performance of MPJ Express (multicore mode) with other C and Java message passing libraries---including mpiJava, MPJ/Ibis, MPICH2, MPJ Express (cluster mode)---on shared memory and multicore processors. We found out that MPJ Express performs signicantly better in the multicore mode than in the cluster mode. Not only this but the MPJ Express software also performs better in comparison to other Java messaging libraries including mpiJava and MPJ/Ibis when used in the multicore mode on shared memory or multicore processors. We also demonstrate effectiveness of the MPJ Express multicore device in Gadget-2, which is a massively parallel astrophysics N-body siimulation code.

A benchmark-driven modelling approach for evaluating deployment choices on a multicore architecture

The complexity of current and emerging high performance architectures provides users with options about how best to use the available resources, but makes predicting performance challenging. In this work a benchmark-driven performance modelling approach is outlined that is appro- priate for modern multicore architectures. The approach is demonstrated by constructing a model of a simple shallow water code on a Cray XE6 system, from application-specific benchmarks that illustrate precisely how architectural char- acteristics impact performance. The model is found to recre- ate observed scaling behaviour up to 16K cores, and used to predict optimal rank-core affinity strategies, exemplifying the type of problem such a model can be used for.

Dynamic group communication for large-scale parallel data mining

Exascale systems are the next frontier in high-performance computing and are expected to deliver a performance of the order of 10^18 operations per second using massive multicore processors. Very large- and extreme-scale parallel systems pose critical algorithmic challenges, especially related to concurrency, locality and the need to avoid global communication patterns. This work investigates a novel protocol for dynamic group communication that can be used to remove the global communication requirement and to reduce the communication cost in parallel formulations of iterative data mining algorithms. The protocol is used to provide a communication-efficient parallel formulation of the k-means algorithm for cluster analysis. The approach is based on a collective communication operation for dynamic groups of processes and exploits non-uniform data distributions. Non-uniform data distributions can be either found in real-world distributed applications or induced by means of multidimensional binary search trees. The analysis of the proposed dynamic group communication protocol has shown that it does not introduce significant communication overhead. The parallel clustering algorithm has also been extended to accommodate an approximation error, which allows a further reduction of the communication costs. The effectiveness of the exact and approximate methods has been tested in a parallel computing system with 64 processors and in simulations with 1024 processing elements.