Wednesday, January 02, 2013

XSPEC v12.8 - Parallelization

XSPEC v12.8 is now available as part of HEAsoft 6.13.

This is the first of several posts on new features in v12.8. The parallel command tells XSPEC to use multiple cores on your machine. At the moment we have parallel processing in two places: in the fit command and in the error command. The standard Levenberg-Marquardt algorithm requires the calculation of the numerical derivative of the statistic with respect to each of the parameters. This is an obvious place to split the calculation among multiple cores and we do this if the command parallel leven N is given, where N is the number of cores to use. The more compute-intensive the model, the more time will be saved. In general however, this does not produce improvements linear in the number of cores.

The error command can be run on multiple parameters at the same time and using the parallel command here with parallel error N does produce a linear gain in N provided there are at least N parameter errors being calculated. The one current drawback with the parallel error calculation comes if the error finds a new minimum. At present all that will happen is that XSPEC will write out that a new error has been found but not update the best fit. To actually find the new best fit the error command must be repeated on individual parameters. We do intend to fix this.

The next command for which we intend to add parallel processing is steppar. We welcome other suggestions and feedback.

No comments: