Presentation is loading. Please wait.

Presentation is loading. Please wait.

HPEC 2004 Sparse Linear Solver for Power System Analysis using FPGA Jeremy Johnson, Prawat Nagvajara, Chika Nwankpa Drexel University.

Similar presentations


Presentation on theme: "HPEC 2004 Sparse Linear Solver for Power System Analysis using FPGA Jeremy Johnson, Prawat Nagvajara, Chika Nwankpa Drexel University."— Presentation transcript:

1 HPEC 2004 Sparse Linear Solver for Power System Analysis using FPGA Jeremy Johnson, Prawat Nagvajara, Chika Nwankpa Drexel University

2 HPEC 2004 Goal & Approach To design an embedded FPGA-based multiprocessor system to perform high speed Power Flow Analysis. To design an embedded FPGA-based multiprocessor system to perform high speed Power Flow Analysis. To provide a single desktop environment to solve the entire package of Power Flow Problem (Multiprocessors on the Desktop). To provide a single desktop environment to solve the entire package of Power Flow Problem (Multiprocessors on the Desktop). Solve Power Flow equations using Newton- Raphson, with hardware support for sparse LU. Solve Power Flow equations using Newton- Raphson, with hardware support for sparse LU. Tailor HW design to systems arising in Power Flow analysis. Tailor HW design to systems arising in Power Flow analysis.

3 HPEC 2004 Algorithm and HW/SW Partition Data InputExtract Data Ybus LU Factorization Forward SubstitutionBackward Substitution Jacobian Matrix Post Processing Update Jacobian matrix Mismatch < Accuracy HOST YES NO

4 HPEC 2004 Results Software solutions (sparse LU needed for Power Flow) using high-end PCs/workstations do not achieve efficient floating point performance and leave substantial room for improvement. Software solutions (sparse LU needed for Power Flow) using high-end PCs/workstations do not achieve efficient floating point performance and leave substantial room for improvement. High-grained parallelism will not significantly improve performance due to granularity of the computation. High-grained parallelism will not significantly improve performance due to granularity of the computation. FPGA, with a much slower clock, can outperform PCs/workstations by devoting space to hardwired control, additional FP units, and utilizing fine-grained parallelism. FPGA, with a much slower clock, can outperform PCs/workstations by devoting space to hardwired control, additional FP units, and utilizing fine-grained parallelism. Benchmarking studies show that significant performance gain is possible. Benchmarking studies show that significant performance gain is possible. A 10x speedup is possible using existing FPGA technology A 10x speedup is possible using existing FPGA technology

5 HPEC 2004 Benchmark Obtain data from power systems of interest Obtain data from power systems of interest Source # Bus Branches/BusOrderNNZ PSSE1,6481.582,98221,682 PSSE7,9171.6414,508108,024 PJM10,2781.4219,285137,031 PJM26,8291.4350,092361,530

6 HPEC 2004 System Profile # Bus # Iter #DIV#MUL#ADD NNZ L+U 1,648643,8761,908,0821,824,380108,210 7,9179259,38818,839,38218,324,787571,378 10,27912238,34314,057,76613,604,494576,007 26,829770,51490,556,64389,003,9261,746,673

7 HPEC 2004 System Profile More than 80% of rows/cols have size < 30 More than 80% of rows/cols have size < 30

8 HPEC 2004 Software Performance Software platform Software platform UMFPACK UMFPACK Pentium 4 (2.6GHz) Pentium 4 (2.6GHz) 8KB L1 Data Cache 8KB L1 Data Cache Mandrake 9.2 Mandrake 9.2 gcc v3.3.1 gcc v3.3.1 # Bus Time FP Eff 1,648 0.07 sec 1.05% 7,917 0.37 sec 1.33% 10,278 0.47 sec 0.96% 26,829 1.39 sec 3.45%

9 HPEC 2004 Hardware Model & Requirements Store row & column indices for non-zero entries Store row & column indices for non-zero entries Use column indices to search for pivot. Overlap pivot search and division by pivot element with row reads. Use column indices to search for pivot. Overlap pivot search and division by pivot element with row reads. Use multiple FPUs to do simultaneous updates (enough parallelism for 8 – 32, avg. col. size) Use multiple FPUs to do simultaneous updates (enough parallelism for 8 – 32, avg. col. size) Use cache to store updated rows from iteration to iteration (70% overlap, memory  400KB - largest). Can be used for prefetching. Use cache to store updated rows from iteration to iteration (70% overlap, memory  400KB - largest). Can be used for prefetching. Total memory required  22MB (largest system) Total memory required  22MB (largest system)

10 HPEC 2004 Architecture SDRAM MemorySRAM Cache SDRAM ControllerCACHE Controller Processing Logic FPGA

11 HPEC 2004 Pivot Hardware Translate to virtual Index reject Read colmap Physical index Virtual index Memory read FP compare Column value Pivot index Pivot value Pivot columnPivot Pivot logic

12 HPEC 2004 Parallel FPUs

13 HPEC 2004 Update Hardware Update Row FMUL Column Word Merge Logic FADD SelectWrite Logic colmap Update Memory read row Submatrix Row

14 HPEC 2004 Performance Model C program which simulates the computation (data transfer and arithmetic operations) and estimates the architecture’s performance (clock cycles and seconds). C program which simulates the computation (data transfer and arithmetic operations) and estimates the architecture’s performance (clock cycles and seconds). Model Assumptions Sufficient internal buffers Sufficient internal buffers Cache write hits 100% Cache write hits 100% Simple static memory allocation Simple static memory allocation No penalty on cache write-back to SDRAM No penalty on cache write-back to SDRAM

15 HPEC 2004 Performance …

16 GEPP Breakdown Cycle Count By # of Update Units 1416 Pivot259401259401259401 Divide964399643996439 Update2409312642662248295


Download ppt "HPEC 2004 Sparse Linear Solver for Power System Analysis using FPGA Jeremy Johnson, Prawat Nagvajara, Chika Nwankpa Drexel University."

Similar presentations


Ads by Google