Dear all,
I am testing GPU-Hmmer and CUDA-MEME on Tesla card C1060. I am getting the results which does not show the speedup provided on nvidia website. I want to know about this difference and what dataset they have used which gives such speedup i.e 23x faster with CUDA-MEME as compared to MEME and 60x-100x faster with gpu-HMMER.
Please suggest me about such performance.
My few results are as follows:
GPU-Hmmer v/s mpi-HMMER
****************** ON TESLA **************************
$ hmmbuild globins4.hmm globins.sto
-
$ time cuda_hmmsearch --force-GPU 2 globins4.hmm ~/MPIBLAST/uniprot_sprot.fasta >& sprot
real 3m31.565s
user 4m31.885s
sys 5m55.051s -
$ time cuda_hmmsearch --force-GPU 2 globins4.hmm ~/MPIBLAST/uniprot_trembl.fasta >& trembl
real 57m49.711s
user 68m56.729s
sys 86m35.589s
******************* On CPU ***************************
$ hmmbuild globins4.hmm globins.sto
-
$ time mpirun -machinefile ../../mcfile -np 2 ../../path-mpi-hmmer-gnu/bin/hmmsearch --mpi globins4.hmm ~/MPIBLAST/uniprot_sprot.fasta
real 6m30.618s
user 0m0.185s
sys 0m0.073s
- $ time mpirun -machinefile …/…/mcfile -np 2 …/…/path-mpi-hmmer-gnu/bin/hmmsearch --mpi globins4.hmm ~/MPIBLAST/uniprot_trembl.fasta >& trembl
real 87m17.149s
user 0m0.140s
sys 0m0.041s
CUDA-MEME v/s MEME
****************** ON TESLA **************************
1)$ time ./meme ~/MPIBLAST/made284k.fasta -protein -mod oops -maxw 40 -maxsites 50 -maxsize 300500 >& out284
real 37m17.982s
user 37m50.411s
sys 0m7.486s
- $ time ./meme ~/MPIBLAST/made320k.fasta -protein -mod oops -maxw 40 -maxsites 50 -maxsize 300500 >& out320
real 51m45.040s
user 54m26.881s
sys 0m9.267s
******************* On CPU ***************************
- $ time ./path-meme-cpu-gnu/bin/meme ~/MPIBLAST/made284k.fasta -protein -mod oops -maxw 40 -maxsites 50 -maxsize 300500
real 43m47.634s
user 42m13.788s
sys 0m0.244s
- $ time ./path-meme-cpu-gnu/bin/meme ~/MPIBLAST/made320k.fasta -protein -mod oops -maxw 40 -maxsites 50 -maxsize 300500
real 60m16.750s
user 60m15.372s
sys 0m0.304s
Thanks in advance
- Deepti