Multi-Core Parallel Computing with ISIS

ISIS can perform a number of model fitting tasks in parallel, often decreasing the associated run-time in proportion to the number of available compute cores.

ISIS achieves this parallelization by distributing the computation across multiple CPU cores in a way that requires no additional effort on the part of the user, no additional software dependencies, and is compatible with legacy codes.

Along with the built-in support for parallel fitting, a simple interface is provided so that users can write parallel scripts.

For more realistic applications of multi-core parallelism, look in the ISIS source code distribution at the implementation of the parallel Levenberg-Marquardt optimizer, plm, in the file share/plm.sl, or at the implementation of conf_loop, in the file share/conf_loop.sl.

The degree of parallelism that can be achieved with this method is effectively limited by the number of compute cores available in a single machine. Larger scale parallelism can be achieved using the PVM module, but the effort required to set up and run such a calculation can be non-trivial.

ISIS automatically performs the following tasks in parallel, using all available cpus:

• Model fits using plm, a parallel implementation of the Levenberg-Marquardt optimization method.
• Monte-Carlo parameter search, using fit_search
• Sets of single parameter confidence limit searches, with automatic refits upon discovery of a new best-fit, using conf_loop
• 2D confidence contour maps, using conf_map_counts or conf_map_flux.
• parallel_map provides the same interface as array_map, except that the user-provided function executes in parallel.

Here are a few usage examples:

 % Select a parallel fit method,
 % then perform a fit, using all available cores:

  isis> set_fit_method ("plm");
  isis> fit_counts;

 % Select a serial fit method, then compute single-parameter
 % confidence limits for parameters 2,4,7 and 10.
 % The necessary fits will be performed in parallel
 % on all available cores:

  isis> set_fit_method ("lmdif");
  isis> (pmin, pmax) = conf_loop ([2,4,7,10]);

 % Repeat the conf_loop calculation using only 2 slave
 % processes, and running the slaves at reduced priority:

  isis> (pmin, pmax) = conf_loop ([2,4,7,10]; num_slaves=2, nice=10);

The figure below shows the CPU time required to generate each NxN confidence contour map, with the serial times in black and the parallel times in red. For this calculation, on this machine, the parallel computation is roughly 3x faster than the corresponding serial computation:

Note that the speed-up obtained by parallel processing is somewhat problem-dependent and will also depend on the available computer resources, both memory and cpu.