Presentation is loading. Please wait.

Presentation is loading. Please wait.

Weekly Report- Reduction Ph.D. Student: Leo Lee date: Oct. 30, 2009.

Similar presentations


Presentation on theme: "Weekly Report- Reduction Ph.D. Student: Leo Lee date: Oct. 30, 2009."— Presentation transcript:

1 Weekly Report- Reduction Ph.D. Student: Leo Lee date: Oct. 30, 2009

2 Outline 7 Reduction implementations Matrix multiplication Protein Identification Work plan

3 Five implementation discussed before Interleaved Addressing With “%” Interleaved Addressing Bank conflicts Sequential addressing Perform add when loading Unroll the last warp

4 Implementation 6: Complete Unrolling Specify block size as a function template parameter

5 Invoking template kernels Could block size be a parameter?

6 Results

7 Implementation 7:

8 Results

9 Brief summary Optimization Use efficient operator, not %; Avoid branch divergent in warps; Try to minimize the time of accessing the global memory; Avoid bank conflict in shared memory; Unroll the loop as much as possible;

10 Matrix multiplication Time calculation total time, in Main function double lf = clock(); for(int i=0; i<nMultipTimes; ++i) { RunTest(argc, argv, i, lfTotalTime, true); } total time = clock()-lf; Compute time, in RunTest Function //create and start timer unsigned int nTimer = 0; cutilCheckError(cutCreateTimer(&nTimer)); cutilCheckError(cutStartTimer(nTimer)); matrixMul >>(pDC, pDA, pDB, WA, WB); //copy result from device to host cutilSafeCall(cudaMemcpy(pHC, pDC, nMemSizeC, cudaMemcpyDeviceToHost)); //stop and destroy timer cutilCheckError(cutStopTimer(nTimer));

11 Experiments WA, HA, WB GPU / CPU Comput time (ms)total time (ms) 16,16,16 GPU / CPU45 / 1524678 / 78 32,32,32 GPU / CPU60 / 6227250 / 203 48,80,128 GPU / CPU225 / 86126625 / 1203 128,256,512 GPU / CPU 4249 / 4582935531 / 49328 512,512,512 GPU / CPU 27441 / 364232 70359 / 382062 2048,2048,2048 GPU16979672020968 1697967->20, excluding the results transferring to the host

12 Mass Spectrometry Based Protein Identification Mixed Proteins >ipi|IPI00243451|IPI00243451.6 MDQHQHLNKTAESASSEKKKTRR CNGFKMFLAALSFSYIAKALGGII MKISITQIERRFD… TAESASSEK MFLAALSFSYIAK … Digest Mixed peptides LC-MS/MS Data analyze Protein sequence Peptide sequence Merge Tandem MS

13 Protein Identification Computing intensive Usually identify 50,000 around Mass Spectrum; Each MS will score with at least 10,000 peptides, and some time more than 100,000 peptides; Some experiment costs more than a year, and parallel software is commonly used. Parallel identification related papers are published in JRP, RCM and Bioinformatics. But still no research are based on GPU.

14 Protein Identification Mass spectrum Peptide Mass + tolerance M1 M2 M3 … M200 TAESASSEK Mass P1 P2 P3 … P40 score For each Mi in MS check if it is in P Got a vector MSB (0, 1, 0, 0, …) For each Pi in P check if it is in MS Got a vector PB(1, 1, 0…) Compute the cosine… Keep the top-k score Vector MSVector P

15 400 EVDG 400 AAEE 400 PSTD 631 EMSVPS 699 TLKHLK 699 WDRDL …… Peptide >IQPSKANME TEPDQ… >DEAVPPPAL QLQFN… >RQRAILKVM NTIGGE… … Protein Protein Identification

16 X! Tandem The most efficient open source software Parallel version Not too large!

17 Other works Read parallel computing books; Learn MLP, data mining.

18 Work plan Data mining homework; K-means debug and test.

19 Thanks


Download ppt "Weekly Report- Reduction Ph.D. Student: Leo Lee date: Oct. 30, 2009."

Similar presentations


Ads by Google