poor performance of OpenMP version

I've just downloaded the most recent version of MRCC and built it with the Intel compiler (v.14) and the -pOMP option. Previously I had built an older version with the PGI compiler and had been using that for some time. Running from the Molpro interface (but setting the value of OMP_NUM_THREADS externally and using serial Molpro), the Intel-compiled version sees a significant slow-down when the number of threads are greater than 1 in CCSDT calculations (simple test). This is completely contrary to my previous PGI version where I did get some speedup. I've used both the default Intel build (uses mkl=parallel) and one where I've inserted openmp throughout and then linked to mkl as in Molpro (this is similar to what I had done before with the PGI compiler).

Since my cpus are all Intel SandyBridge processors, I had expected a strong advantage with using the Intel compiler.

thanks in advance, -Kirk

A quick addition to this post. Tonight I also successfully got the newest version of MRCC to build using the PGF compiler (v.12.10) in conjunction with gcc, also using -pOMP. In that case I see reasonable scaling as I increase the number of threads just as in my older 2012 version of MRCC.

-Kirk

Dear Kirk,
It is strange. I have just run a quick test, H2O2 CCSDT/cc-pVTZ, and I have got a speedup of about two on a quad core Intel i7-4700MQ CPU. I use Intel compiler version 14.0.2.144.
Are you sure that the OMP_NUM_THREADS and MKL_NUM_THREADS environmental variables are correctly set?

Dear Mihaly,

actually I've never set MKL_NUM_THREADS, only the OMP one (I thought the latter too precedence). What should the MKL environment variable be set to?

best, -Kirk

Dear Mihaly,

so if I set MKL_NUM_THREADS=1 , a 4-core calculation is 2x slower than a single-core (as specified by using OMP_NUM_THREADS). Obviously I have something not set correctly.

-Kirk

Dear Mihaly,

I think actually everything is fine after all. It seems that Molpro prints the incorrect CPU times. I just happened to see what it prints for wall times and these show the correct scaling, not great but a factor of 2 speed-up when going from 1 core to 4.

sorry for stirring this up for nothing.

-Kirk