"Parallel processing on X-validation"

wessel

Parallel processing,

When using X-validation and the RM parallel processing plug in installed,
is RM evaluating classifier performance on each X-fold in parallel?

Best regards,

Wessel

land

Hi Wessel,
not automatically. But there exists a X-Validation  (parallel) operator that will use multiple threads for computing it parallelly. Keep in mind, that it will use more memory, since each operator is executed parallely and each instance will create its data structures like for example the kernel cache of SVMs. But from my experience, you will recieve an increase of speed thats nearly linear in the number of cores/threads used.

Greetings,
  Sebastian

dragoljub

I have also noticed a check box for parallelizing training/testing.

When this is checked I experience synchronization problems. Did you guys actually implement some parallel smo algorithms for training? For testing it makes since since data can be split up and classified in parallel.

-Gagi

land

Hi,
this check box is generally available on all sub process providing operators, but I would rather recommend to use it with caution. Using this checkbox the user has to decide if the process is actual paralellizable, if not, these synchronization problems might occur. I would recommend using the dedicated parallel operators to be sure.
We don't have implemented a parallel SMO, at least the X-Validation can be implemented parallel for each learning algorithm:
Since the Predictions are made only once on a single fold, these write accesses don't interfere and the reading of the data for learning is indeed no problem, so it can be carried out parallely, too. So you can speed up a ten fold X-Validation up to factor of ten by using ten threads on a machine with at least ten cores. Beyond this, you would need not only to execute the single threaded learning algorithm ten times, but have to make this paralell, too.

Greetings,
  Sebastian

dragoljub

I have a 4 core system. I have explored preforming 10-fold X-validation using 10 threads and this also gave me problems. I am back to my default of 4 but occasionally if the 10-thread method works its still faster than 4 threads.

Could this also be a sync problem?

Thanks,
-Gagi ;D

land

Hi,
sometimes using more threads than cores available avoids some still existent race conditions that prevent a thread from working all the time.

What problems do you encounter? Could you be a little bit more precise so that I can check if there are still problems?

Greetings,
  Sebastian