See Section 6 of the paper for a detailed description of the benchmark applications, experimental parameters, machine specifications and environment conditions.

Following are the complete results for the above benchmarks, including those omitted from the paper.

• Benchmark 1: runtime overheads due to Abstract Transport in the SJ Framework.
  • 1-1. Performance of the new SJ Runtime with Abstract Transport mechanisms against the preceding TCP-coupled SJ implementation.
  • 1-2. Performance of the new SJ Runtime against Java RMI.


• Benchmark 2: impact of SJ noalias types.
• Benchmark 3: transport-independent delegation for cross-transport session migration.


• Benchmark 4: macro benchmarks using parallel algorithm implementations.
  • Parallel Jacobi algorithm for solving the discrete Poisson equation. This application further demonstrates the impact of noalias types. The performance of the SJ implementation is compared against a MPJ Express equivalent.
  • A distributed simulation of the n-Body problem. This application demonstrates implementation portability and deployment flexibility due to transport-independence.

The full source code (including build and execution scripts) for Benchmarks 1, 2 and 3 is here: tar.gz.

The complete results (raw data sets) for Benchmarks 1, 2 and 3 are available here: tar.gz.

Format of the results files. In each file, the results for each parameter configuration are headed by e.g. benchmark1a: depth=0 where depth gives the size of the BinaryTree messages being communicated. A single run of each benchmark corresponds to 5 lines of data: the time (micros) to complete sessions of length 0, 1, 10, 100 and 1000 for the specified message sizes (tree depth). Each benchmark was repeated 1000 times in total for each parameter configuration.

The source code for the parallel algorithm implementations are included in the application source package and in the main SJ distribution. The raw data from the macro benchmarks is available from here.

Back to the main page.

1-1. Performance of the new SJ Runtime with Abstract Transport framework against the preceding TCP-coupled SJ implementation.

Note: for this benchmark, the new SJ Runtime was configured to utilise the same wire protocol and message serialization format as the preceding SJ implementation (to directly measure the overheads due to the Abstract Transport mechanisms).

Back to the top (for benchmark source code and raw data files).

1-2. Performance of the new SJ Runtime against Java RMI.

*Note: the RMI benchmarks do not include the cost of the initial RMI registry lookup. Hence, no measurements are taken for RMI sessions of length 0. For this benchmark, the new SJ Runtime was configured to utilise a more optimal serialization component (different to above) for the communication of large binary tree messages.

Back to the top (for benchmark source code and raw data files).

2. The potential performance gain from the communication of SJ noalias types in shared memory sessions.

Back to the top (for benchmark source code and raw data files).

3. Performance gain in applications featuring session delegation due to (cross-transport) session migration.

Back to the top (for benchmark source code and raw data files).

4. Macro benchmarks using parallel algorithm implementations.

Please see here for implementation details and the complete benchmark results for the parallel Jacobi and n-Body algorithms.

Back to the top or go to the SJ homepage.