[R] performance of svm (e1071) on windows 64 bits vs 32 bits

[Frederico Arnoldi]

Dear R users,
I´ve been using SVM (e1071) on different 64 bits linux distributions very successfully. 
Very recently I had to migrate some of our codes to a server with xeon e7 processors, 65 gb of RAM, running Windows. Very surprisingly, the peformance of the same code, in the same server, on R 32 bits was much better than on R 64 bits (~5 times).   I confirmed these results in another windows machine, now a notebook, to exclude the possibility of a problem specific to that server.
In both cases, R 32 bits is using more RAM memory (~900 mb) than 64 bits (~600 mb). Changing cache size of SVM doesn´t impact it.
As I need to use de 64 bits version, I went to internet looking for ways of improving the perfomance of 64 bits version. Reading about possible causes, I arrived to Blas and Lapack issues. I have tried different pre-compiled Rblas.dll available on the internet, including SurviveGotoBLAS-3.14. Also, I tried to use R from Revolution Analytics (Revolution R Open) that has a optimized Blas (Intel MKL) , without any visible.
Could you give any help or share similar experience with this issue? 
Thank you very much for your help.
Frederico Arnoldi

 		 	   		  
	[[alternative HTML version deleted]]