CMSC 858M/AMSC 698R. Fast Multipole Methods. Nail A. Gumerov & Ramani Duraiswami. Lecture 20. Outline

Transcription

1 CMSC 858M/AMSC 698R Fast Multipole Methods Nail A. Gumerov & Ramani Duraiswami Lecture 20 Outline Two parts of the FMM Data Structures FMM Cost/Optimization on CPU Fine Grain Parallelization for Multicore Nodes Graphics Processors (GPUs) Fast Matrix-Vector Multiplication on GPU with On-Core Computable Kernel Change of the Cost/Optimization scheme on GPU compared to CPU Other steps Test Results Profile Accuracy Overall Performance Heterogeneous algorithm 1

2 Publication The major part of this work has been published: N.A. Gumerov and R. Duraiswami Fast multipole methods on graphics processors. Journal of Computational Physics, 227, , Q. Hu, N.A. Gumerov, and R. Duraiswami Scalable Fast Multipole Methods on Distributed Heterogeneous Architectures, Proceedings of the Supercomputing 11 Conference, Seattle, November 12-18, (The paper is nominated for the best student paper award). Two parts of the FMM Generate data structure (define the matrix); Perform Matrix-vector product (run); Two different types of problems: Static matrix (e.g. for iterative solution of large linear system): Data structure should be generated once, while run should be performed many times with different input vectors; Dynamic matrix (e.g. N-body problem) Data structure should be generated each time, when matrixvector multiplication is needed. 2

3 Data Structures We need O(NlogN) procedure to index all the boxes, find children/parent and neighbors; Implemented on CPU using bit-interleaving (Morton-Peano indexing in octree) and sorting; We generate lists of Children/Parent/Neighbors/E4-neighbors and store in global memory; Recently, graduate students Michael Lieberman and Adam O Donovan implemented basic data structures on GPU (prefix sort, etc.) and found of order 10 speedups compared to the CPU for large data sets. The most recent work on data structures on GPU is performed by Hu Qi (up to 100 times speed up compared to the standard CPU). In the present lecture we will focus on the run algorithm, assuming that the necessary lists are generated and stored in the global memory. FMM on CPU 3

4 Flow Chart of FMM Run Level l max Levels l max -1,, 2 Level 2 Levels 3,,l max 1. Get S-expansion coefficients (directly) 2. Get S-expansion coefficients from children (S S translation) 3. Get R-expansion coefficients from far neighbors (S R translation) 4. Get R-expansion coefficients from far neighbors (S R translation) Start End Level l max 7. Sum sources in close neighborhood (directly) Level l max 6. Evaluate R-expansions (directly) 5. Get R-expansion coefficients from parents (R R translation) Direct Sparse Matrix- Vector Multiplication Approximate Dense Matrix- Vector Multiplication FMM Cost (Operations) Points are clustered with clustering parameter s, means not more that s points in the smallest octree box. Expansions have length p 2 (number of coefficients). There are O(N) points and O(N boxes )=O(N/s) in the octree. O(Np 2 ) O(N boxes p 3 ) O(p 3 ) O(N boxes p 3 ) Level l max Levels l max -1,, 2 Level 2 Levels 3,,l max 1. Get S-expansion coefficients (directly) 2. Get S-expansion coefficients from children (S S translation) 3. Get R-expansion coefficients from far neighbors (S R translation) 4. Get R-expansion coefficients from far neighbors (S R translation) Start End Level l max 7. Sum sources in close neighborhood (directly) Level l max 6. Evaluate R-expansions (directly) 5. Get R-expansion coefficients from parents (R R translation) O(Ns) O(Np 2 ) O(N boxes p 3 ) 4

5 FMM Cost/Optimization (operations) Since p=o(1), we have TotalCost = AN+BN noxes +CNs =N(A+B/s+Cs) Total Cost Cs s opt =(B/C) 1/2 B/s s opt s Fine Grain Parallelization for Multicore PC s Serial Algorithm Constructor Step 1 Loop 1 Step n Allocate Global Memory Parallel Loop 1 Parallel Algorithm Allocate Constructor 1 Step Step n Global Memory Fork parallelization Open MP C++, Fortran 9x Also auto vectorization of loops (SSE) Loop n Destructor Deallocate Parallel Loop n Destructor Deallocate Cannot write to the same memory address! 5

6 Some test results for OMP parallelization (4 cores) GPUs 6

7 A Quick Introduction to the GPU Graphics processing unit (GPU) is a highly parallel, multithreaded, many-core processor with high computation power and memory bandwidth GPU is designed for single instruction multiple data (SIMD) computation; more transistors for processing rather than data caching and flow control A few cores Hundreds cores Control ALU ALU ALU ALU Cache DRAM CPU DRAM GPU NVIDIA Tesla C2050: 1.25 Tflops single 0.52 Tflops double 448 cores Is It Expensive? Any PC has GPU which probably performs faster than the CPU GPUs with Teraflops performance are used in game stations Tens of millions of GPUs are produced each year Price for 1 good GPU in range $ Prices for the most advanced NVIDIA GPUs for general purpose computing (e.g. Tesla C2050) are in the range $1K-$2K Modern research supercomputer with several GPUs can be purchased for a few thousand dollars GPUs provide the best Gflops/$ ratio They also provide the best Gflops/watt 7

8 Floating-Point Operations for CPU and GPU Is It Easy to Program A GPU? For inexperienced GPU programmers Matlab Parallel Toolbox Middleware to program GPU from Fortran Relatively easy to incorporate to existing codes Developed by the authors at UMD Free (available online) For advanced users CUDA: a C-like programming language Math libraries are available Custom functions can be implemented Requires careful memory management Free (available online) Local memory ~50 kb GPU global memory ~1-4 GB Host memory ~4-128 GB 8

9 Availability Fortran-9X version is released for free public use. It is called FLAGON. FMM on GPU 9

10 Challenges Complex FMM data structure; Problem is not native for SIMD semantics; non-uniformity of data causes problems with efficient work load (taking into account large number of threads); serial algorithms use recursive computations; existing libraries (CUBLAS) and middleware approach are not sufficient; high performing FMM functions should be redesigned and written in CU; Low fast (shared/constant) memory for efficient implementation of translation operators; Absence of good debugging tools for GPU. Implementation Main algorithm is in Fortran 95 Data structures are computed on CPU Interface between Fortran and GPU is provided by our middleware (Flagon), which is a layer between Fortran and C++/CUDA High performance custom subroutines written using Cu 10

11 Time (s) 11/10/2011 Fast Kernel Matrix-vector multiplication on GPU High performance direct summation on GPU (total) 1.E+04 1.E+03 Computations of potential y=ax 2 CPU: 2.67 GHz Intel Core 2 extreme QX 7400 (2GB RAM and one of four CPUs employed). 1.E+02 1.E+01 CPU, Direct y=bx 2 GPU: NVIDIA GeForce 8800 GTX (peak 330 GFLOPS). GPU, Direct 1.E+00 1.E-01 1.E-02 y=bx 2 +cx+d b/a=600 1.E-03 3D Laplace 1.E-04 1.E+03 1.E+04 1.E+05 1.E+06 Number of Sources Estimated achieved rate: 190 GFLOPS. CPU direct: serial code; no use of partial caching; no loop unrolling; (simple execution of nested loop) 11

12 Receiver Data Thread registers Algorithm for direct summation on GPU (Kernel matrix multiplication) 1 2 n Global GPU memory Block of Data Shared Coalesced (fast) read to thread register (fast memory) Source Data Global GPU memory Coalesced (fast) copy to the shared (fast) memory Compute and add contribution of the j-th source to the sum handled by the i-th thread (performed by the i-th thread) i-th thread Write result to global GPU memory Block of Data Block of threads Change of Cost/Optimization on GPU 12

13 Time (s) 11/10/2011 Algorithm modification for direct summation in the FMM 1. Each block of threads executes computations for one receiver box (Blockper-box parallelization). 2. Each block of threads accesses the data structure. 3. Loop over all sources contributing to the result requires partially non-coalesced read of the source data (box by box). Direct summation on GPU (final step in the FMM) Computations of potential, optimal settings for CPU 1.E+02 1.E+01 Sparse Matrix-Vector Product s max =60 y=bx CPU: Time=CNs, s=8 -lmax N 1.E+00 1.E-01 y=ax 2 CPU 4 l max =5 y=cx GPU: Time=A 1 N+B 1 N/s+C 1 Ns 1.E-02 1.E GPU b/c=16 y=8 -lmax dx 2 +ex+8 lmax f read/write float computations 3D Laplace 1.E-04 1.E+02 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 Number of Sources access to box data These parameters depend on the hardware 13

14 Optimization Sparse M-V product Time Dense M-V product Time Time CPU: + = 0 s 0 s 0 s opt s Time Time Time GPU: + = 0 s1 0 0 s s s 1 s opt s Direct summation on GPU (final step in the FMM) Compare GPU final summation complexity: and total FMM complexity: Cost =A 1 N+B 1 N/s+C 1 Ns. Cost = AN+BN/s+CNs. Optimal cluster size for direct summation step of the FMM s opt = (B 1 /C 1 ) 1/2, and this can be only increased for the full algorithm, since its complexity Cost =(A+A 1 )N+(B+B 1 )N/s+C 1 Ns, and s opt = ((B+B 1 )/C 1 ) 1/2. 14

15 Time (s) 11/10/2011 Direct summation on GPU (final step in the FMM) Computations of potential, optimal settings for GPU 1.E+03 1.E+02 Sparse Matrix-Vector Product s max =320 y=bx 1.E+01 y=ax 2 CPU l max =4 y=cx 1.E E-01 1.E-02 2 GPU b/c=300 1.E-03 y=8 -lmax dx 2 +ex+8 lmax f 3D Laplace 1.E-04 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 Number of Sources Direct summation on GPU (final step in the FMM) Computations of potential, optimal settings for CPU and GPU N Serial CPU (s) GPU(s) Time Ratio E E E E E E E E E E E E E E E E E E

16 Other steps of the FMM Important conclusion: Since the optimal max level of the octree when using GPU is lesser than that for the CPU, the importance of optimization of translation subroutines diminishes. 16

17 Other steps of the FMM on GPU Accelerations in range 5-60; Effective accelerations for N=1,048,576 (taking into account max level reduction): Test results 17

18 Profile Accuracy Relative L 2 norm error measure: CPU single precision direct summation was taken as exact ; 100 sampling points were used. 18

19 L2-relative error 11/10/2011 What is more accurate for solution of large problems on GPU: direct summation or FMM? 1.E-02 1.E-03 1.E-04 1.E-05 Error in computations of potential p=4 FMM p=8 FMM FMM Error computed over a grid of 729 sampling points, relative to exact solution, which is direct summation with double precision. 1.E-06 p=12 Possible reason why GPU the GPU error in direct 1.E-07 Direct summation grows: CPU systematic roundoff 1.E-08 error in computation of function 1/sqrt(x). Filled = GPU, Empty = CPU (still a question). 1.E-09 1.E+02 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 Number of Sources Performance N=1,048,576 (potential only) serial CPU GPU Ratio p= s s 33 p= s s 56 p= s s 48 N=1,048,576 p=4 p=8 p=12 (potential+forces (gradient)) serial CPU GPU s s s s s s Ratio

20 Run Time (s) Run Time (s) Run Time (s) 11/10/2011 Performance p=4 p=8 p=12 1.E+02 3D Laplace y=ax 2 y=bx 2 y=cx 1.E+02 3D Laplace y=ax 2 y=bx 2 y=cx 1.E+02 3D Laplace y=ax 2 y=cx 1.E+01 1.E+00 CPU, Direct CPU, FMM y=dx 1.E+01 1.E+00 CPU, Direct CPU, FMM GPU, Direct y=dx 1.E+01 1.E+00 CPU, Direct CPU, FMM y=bx 2 y=dx GPU, FMM GPU, Direct 1.E-01 1.E-01 1.E-01 1.E-02 GPU, FMM a/b = 600 c/d = 30 1.E-02 GPU, FMM a/b = 600 c/d = 50 1.E-02 GPU, Direct a/b = 600 c/d = 50 p=4, FMM error ~ E-03 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 Number of Sources p=8, FMM error ~ E-03 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 Number of Sources p=12, FMM error ~ E-03 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 Number of Sources Performance FMM GPU Computations of the potential and forces: Peak performance of GPU for direct summation 290 Gigaflops, while for the FMM on GPU effective rates in range Teraflops are observed (following the citation below). dir CPU FMM M.S. Warren, J.K. Salmon, D.J. Becker, M.P. Goda, T. Sterling & G.S. Winckelmans. Pentium Pro inside: I. a treecode at 430 Gigaflops on ASCI Red, Bell price winning paper at SC 97, direct 20

21 Heterogeneous Algorithm ODE solver: source receiver update GPU work Single node algorithm CPU work particle positions, source strength data structure (octree and neighbors) source S-expansions translation stencils local direct sum upward S S downward S R R R receiver R-expansions final sum of far-field and near-field interactions time loop 21

22 Advantages CPU and GPU are tasked with the most efficient jobs Faster translations: CPU code can be better optimized which requires complex translation stencil data structures Faster local direct sum: many cores on GPU; same kernel evaluation but on multiple data (SIMD) The CPU is not idle during the GPU-based computations High accuracy translation without much cost penalty on CPU Easy visualization: all data reside in GPU RAM Smaller data transfer between the CPU and GPU Single node tests 22

23 The billion size test case Using all 32 Chimera nodes Timing excludes the partition 1 billion runs potential kernel in 21.6 s Actual performance: TFs To compute without FMM for the same time: 1 EFs performance 23