Hello All,
In general when we use Rprof for performance evaluation on
Multicore systems the output provides the time on the basis of the
"user"
time and the sampling time is equal to the the user time as reported by
system.time. This does not seem right behavior when R is linked to
BLAS/Lapack or other libraries which are optimized for parallel or multicore
architectures as over there user time can be more than the elapsed time and
one would be more interested in just the "elapsed" time taken by
computation returned by gettimeofday() per routine rather than "user"
time
as returned by getrusage().
Could anyone provide any pointers on how to best do R
profiling on parallel and multicore systems.
Regards,
--
Imanpreet Singh Arora
[[alternative HTML version deleted]]
Posting it to R-Devel:
Hello All,
In general when we use Rprof for performance evaluation on
Multicore systems the output provides the time on the basis of the
"user"
time and the sampling time is equal to the the user time as reported by
system.time. This does not seem right behavior when R is linked to
BLAS/Lapack or other libraries which are optimized for parallel or multicore
architectures as over there user time can be more than the elapsed time and
one would be more interested in just the "elapsed" time taken by
computation returned by gettimeofday() per routine rather than "user"
time
as returned by getrusage().
Could anyone provide any pointers on how to best do R
profiling on parallel and multicore systems.
Regards,
--
Imanpreet Singh Arora
--
Imanpreet Singh Arora
[[alternative HTML version deleted]]