Parallel Computing for Bayesian Networks - Bayesian Networks in R: With Applications in Systems Biology - page 117

35

30

25

20

15

1 2 3 4 5 6

number of slaves

Fig. 5.7 Performance of cross-validation for different numbers of slave processes, measured by its
execution time (in seconds)

Hill-climbing and tabu search have comparable error rates (50 4%,

respectively). As was the case for the bootstrap, using two slave processes halves
the execution time.

. 67 % and 50 .

> system.time(bn.cv(hailfinder, 'mmhc', loss = "pred",
+ loss.args = list(target = "CompPlFcst")))
user system elapsed
37.782 0.056 37.836
> cl = makeCluster(2, type = "MPI")

> system.time(bn.cv(hailfinder, 'mmhc', loss = "pred",
+ loss.args = list(target = "CompPlFcst"),
+ cluster = cl))
user system elapsed
0.196 0.032 19.202
> stopCluster(cl)

Adding more slaves to the cluster improves the performance of the cross-validation
further, at least up to cluster of size 6 (see Fig. 5.7 ).

> cl = makeCluster(3, type = "MPI")
> system.time(bn.cv(hailfinder, 'mmhc', loss = "pred",
+ loss.args = list(target = "CompPlFcst"),
+ cluster = cl))
user system elapsed
0.320 0.016 15.726
> stopCluster(cl)
> cl = makeCluster(4, type = "MPI")
> system.time(bn.cv(hailfinder, 'mmhc', loss = "pred",
+

loss.args = list(target = "CompPlFcst"),

+ cluster = cl))