The R Project SVN R

Rev

Rev 73228 | Rev 81140 | Go to most recent revision | Show entire file | Ignore whitespace | Details | Blame | Last modification | View Log | RSS feed

Rev 73228 Rev 73309
Line 462... Line 462...
462 
at once, to ensure that the number of forked jobs does not exceed the number
 462 
at once, to ensure that the number of forked jobs does not exceed the number
463 
of selected CPUs. As soon as a child process has finished, the next
 463 
of selected CPUs. As soon as a child process has finished, the next
464 
available job for the CPU is forked.
 464 
available job for the CPU is forked.
465 
 
 465 
 
466 
The following code example demonstrates how the execution
 466 
The following code example demonstrates how the execution
467 
time can be reduced by assigning the elements of \code{X} to specific CPUs.
 467 
time can be reduced by assigning the elements of \code{X} to specific CPUs,
- 
 
 468 
not on Windows, where \code{mc.cores} must remain at 1.
468 
 
 469 
 
- 
 
 470 
%%     setting eval=TRUE  would cost ca. 1.5 minutes every time this is built
469 
%%         set eval=FALSE  to save ~ 40 seconds
 471 
%%     also,   .Platform$OS.type == "Windows"  cannot use mc.cores = 2
470 
<<affinity-ex, eval=TRUE>>=
 472 
<<affinity-ex, eval=FALSE>>=
471 
## Exemplary variance filter executed on three different matrices in parallel.
 473 
## Exemplary variance filter executed on three different matrices in parallel.
472 
## Can be used in gene expression analysis as a prefilter
 474 
## Can be used in gene expression analysis as a prefilter
473 
## for the number of covariates.
 475 
## for the number of covariates.
474 
 
 476 
 
475 
library(parallel)
 477 
library(parallel)
476 
n <- 300   # observations
 478 
n <- 300   # observations
477 
p <- 5000  # covariates (orig. 20000 is more realistic; limit time here)
 479 
p <- 20000 # covariates
478 
 
 480 
 
479 
## Different sized matrices as filter inputs
 481 
## Different sized matrices as filter inputs
480 
## Matrix A and B form smaller work loads
 482 
## Matrix A and B form smaller work loads
481 
## while matrix C forms a bigger workload (2*p)
 483 
## while matrix C forms a bigger workload (2*p)
482 
library(stats)
 484 
library(stats)
Line 499... Line 501...
499 
## CPU mapping: A and B run on CPU 1 while C runs on CPU 2:
 501 
## CPU mapping: A and B run on CPU 1 while C runs on CPU 2:
500 
affinity <- c(1,1,2)
 502 
affinity <- c(1,1,2)
501 
system.time(
 503 
system.time(
502 
  mclapply(X = list(A,B,C), FUN = varFilter,
 504 
  mclapply(X = list(A,B,C), FUN = varFilter,
503 
           mc.preschedule = FALSE, affinity.list = affinity))
 505 
           mc.preschedule = FALSE, affinity.list = affinity))
- 
 
 506 
##   user  system elapsed
- 
 
 507 
## 34.909   0.873  36.720
- 
 
 508 
 
504 
 
 509 
 
505 
## mclapply without affinity.list
 510 
## mclapply without affinity.list
506 
system.time(
 511 
system.time(
507 
  mclapply(X = list(A,B,C), FUN = varFilter, mc.cores = 2,
 512 
  mclapply(X = list(A,B,C), FUN = varFilter, mc.cores = 2,
508 
           mc.preschedule = FALSE) )
 513 
           mc.preschedule = FALSE) )
- 
 
 514 
##   user  system elapsed
- 
 
 515 
## 72.893   1.588  55.982
- 
 
 516 
 
509 
 
 517 
 
510 
## mclapply with prescheduling
 518 
## mclapply with prescheduling
511 
system.time(
 519 
system.time(
512 
   mclapply(X = list(A,B,C), FUN = varFilter, mc.cores = 2,
 520 
   mclapply(X = list(A,B,C), FUN = varFilter, mc.cores = 2,
513 
            mc.preschedule = TRUE) )
 521 
            mc.preschedule = TRUE) )
- 
 
 522 
##   user  system elapsed
- 
 
 523 
## 53.455   1.326  53.399
514 
@
 524 
@
515 
 
 525 
 
516 
Instead of using \code{affinity.list} for runtime optimization,
 526 
Instead of using \code{affinity.list} for runtime optimization,
517 
it can also be used to simply restrict the computation
 527 
it can also be used to simply restrict the computation
518 
to specific CPUs as in the following example.
 528 
to specific CPUs as in the following example.