Reconfigurable Molecular Dynamics Simulator Navid Azizi, Ian Kuon, - PowerPoint PPT Presentation

FCCM 2004 Reconfigurable Molecular Dynamics Simulator Navid Azizi, Ian Kuon, Aaron Egier, Ahmad Darabiha and Paul Chow University of Toronto

FCCM 2004 Why is Molecular Dynamics interesting? � Simulates interaction of atoms over time � Many possible applications – Biomolecules � Computationally intensive to handle >1000 atoms � Large computer clusters used in the past � Can a reconfigurable simulation system do this better? 2 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 What is Molecular Dynamics? � Simulate using classical Newtonian mechanics F = m a � Integrate acceleration to get position and velocity changes � Use a very small timestep ~ 1 femtosecond 3 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Molecular Dynamics Background � Simulation Procedure per timestep – Sum force over all interacting atoms – Calculate acceleration – Integrate acceleration to update atom position and velocity – Repeat for all atoms 4 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Background - Forces � Two types of forces –Bonded – O(n) –No hardware acceleration required –Non Bonded – O(n 2 ) –Needs hardware acceleration 5 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Background – Force Calculation � Lennard-Jones (LJ) potential models interaction 1.0E-10 ⎡ ⎤ 12 6 ⎛ ⎞ ⎛ ⎞ σ σ ⎜ ⎟ ⎜ ⎟ ( ) ⎢ ⎥ φ r = 4 ε − 7.5E-11 LJ ⎝ ⎠ ⎝ ⎠ ⎢ ⎥ r r ⎣ ⎦ Potential 5.0E-11 2.5E-11 0.0E+00 -2.5E-11 Distance � Force on a atom is the gradient of potential 6 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Background – Simulating Large Volumes � Any interesting volume has far too many atoms to simulate � Solution – Periodic Boundary Conditions Replicated Replicated Replicated Box Box Box B’ B’ A’ A’ B’ A’ Box being Replicated Replicated simulated Box Box B’ B A’ A B’ A’ Replicated Replicated Replicated Box Box Box B’ B’ A’ A’ B’ A’ 7 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Architectural Design – System Overview Slope Memory Force PairGen Computer Function Value Memory Particle Sun Acceleration Workstation Update Memory System Verlet Control Update 8 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Architectural Design – Force Computer Function � r 2 from PG used for Value function lookup Memory � Interpolate to obtain a PairGen more accurate force Force Slope magnitude Computer Memory Acceleration Update 9 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Architectural Design – Force Computer � r 2 is larger than 18-bits –Look up table has a 18-bit address 6.00E+09 5.00E+09 Psuedo-Acceleration 4.00E+09 3.00E+09 2.00E+09 1.00E+09 0.00E+00 0 5E-19 1E-18 1.5E-18 2E-18 -1.00E+09 Separation 2 10 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Architectural Design – Force Computer FF...FF Address to lookup tables high middle bits bits Residual for r 2 interpolation low bits 11 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Precision and Scaling Factors � Architecture uses integer operations to reduce complexity � Precision : number of bits used to represent a value � Scaling Factor : the weight of the least significant bit of the value Precision Scaling Factor 12 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Calculating the Precision and Scaling Factors � Calculations made with atoms at varying distances � Scaling Factor = the minimum value � Precision = log 2 of the difference between minimum and maximum Quantity Scaling Precision Factor 2 -64 Position 38 2 -15 Velocity 51 2 -64 Acceleration 37 13 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Simulation Environment is Configurable � Simulation reconfigurability – Change precision, scaling factors, number of atoms. forces – No wasted hardware – No time overhead when precision is reduced � Entire process is automated – One input file controls entire process – Hardware – C program creates appropriate VHDL – Software interface, Software initialization – Always match the hardware 14 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Implementation � Used the Transmogrifier 3 –4 interconnected Virtex-E 2000’s –2MB memories connected to each Virtex-E –Slow by today’s standards 15 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Verification � Tested accuracy of implementation –Compared TM3 results with software 2500 2000 Energy (kJ/mol) TM3 -Total Energy 1500 Software - Total Energy TM3 - Kinetic Energy 1000 Software - Kinetic Energy TM3 - Potential Energy 500 Software - Potential Energy 0 0 100 200 300 400 500 600 700 Timestep 16 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 System Performance � For a 8192 atom MD system running on the TM3 –Frequency: 26 MHz –Timestep Duration: 37 sec � For a 8192 atom software system running on a 2.4GHz Pentium 4 –Timestep Duration: 10.8 sec � MD system is 3.4X slower than software 17 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 How to improve this? � Memory � New FPGA � Parallelism 18 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Memory Requirements of MD System � Acceleration Array (8192 atom system uses 0.17 MB) � Velocity Array (8192 atom system uses 0.17 MB) � Position Array (8192 atom system uses 0.34 MB) � Lookup Tables: 2 MB 19 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Improving Performance – Memory Organization � On TM3 there is only one external SRAM per FPGA � Single SRAM for all atom information causes large slowdowns – Handshaking – Serial reads for x, y and z – Hardware issues Better memory system 2.1 seconds/timestep (5X faster than software) 20 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Improving Performance – Clock Speed � Run on modern FPGA � All possible improvements for clock speed not explored � Expect a factor of 4 increase to a 100MHz Better memory system + Faster Clock Speed 0.51 seconds/timestep (21X faster than software) 21 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Improving Performance – Parallel Architecture Value Mem VA AA PAi PAj LJFC(0) VU(0) AU(0) Slope Mem Value Mem VA AA PAj PAi LJFC(1) Pair VU(1) AU(1) Slope Mem Gen Value Mem VA AA PAi PAj LJFC VU(n-1) AU(n-1) (n-1) Slope Mem Better memory system + Faster Clock Speed + Parallelize 0.51/n seconds/timestep (21n X faster than software) 22 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Cost, Power Benefits � Performance – MD system can deliver a 21X performance benefit over software – Assume a conservative 10X performance advantage � Cost – Microprocessor motherboard-sized board can fit 4 FPGAs – 4 FPGAs (each $200) + board + SRAM + misc. ~ $1500 – Microprocessor Motherboard + CPU + DRAM ~ $1500 23 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Comparison of MD Simulator and Supercomputers Per Board 1 Pentium + 4 FPGAs + 1 GB DRAM 24 MB SRAM Performance 1X 40X Cost ~$1500 ~$1500 Power 106W 40W Metric Improvement of MD System Performance/Power 100X Improvement Performance/Cost 40X Improvement Performance/Space 40X Improvement 24 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Conclusions � Easily reconfigurable MD System designed � Molecular dynamics simulation can be done on FPGAs � Simple enhancements will improve speed – Power, Cost and Space savings over software 25 Reconfigurable Molecular Dynamics Simulator

FCCM 2004 Future Work � Improve accuracy � Target newer FPGA platform � Support new forces Acknowledgements � Funding for the TM3 Project was provided by Micronet and Xilinx 26 Reconfigurable Molecular Dynamics Simulator