1. Ants in the Ocean: System Design Techniques for Underwater Sensing Applications Ryan Kastner Dept. of Electrical and Computer Engineering University of California, Santa Barbara Computer Engineering Seminar Northwestern University May 24, 2004

2. Ecological Research Programs  Santa Barbara Channel Long Term Ecological Research (SBC LTER)  Partnership for Interdisciplinary Studies of Coastal Oceans (PISCO)  Goals  Focuses on understanding the nearshore ecosystems of the west coast  Time/space variation of individual organisms, populations, and ecological communities

3. Ecological Studies  Importance of land vs. ocean processes in giant kelp forests  How do different nutrients effect ecosystem?  How/when are nutrients delivered?  Runoff during storms very important time  Rough conditions, high surf, undertow  How to measure?  Quantify larval transport to nearshore habitats  Red tide  Conditions for larval transport - temperature, salinity, chemicals?  Marine management tools  When/where to make protected zones?  What is effect of environmental factors on marine life?

4. Enabling Ecological Research  Many studies done over limited time frames  Drop sensors, sit on boat  Storm comes, run out to boats and throw sensors into water  Alternatively, leave sensors unattended  Malfunction loses months of data  Sensors get lost/stolen Ideal situation - real-time, adaptive sampling techniques

5. CoastalNet Temp/Depth Sensors 802.11 AP Conductivity Sensor Acoustic Link Low Rate/FSK Acoustic Link Acoustic Modem/Array Signal Processor Directional Antenna (“Pringles Can”) 802.11Access Point To Internet Wi/Fi Link up to 7 miles OFDM/DS Acoustic Link

6. CoastalNet Challenges  Underwater communication  Water more complex medium than air - severe multipath problems, doppler shifts and long latencies  Commercial modems ~2400 baud – fine for simple sensors. What about higher data rates?  Sensors  Variety of different types, sizes – video, salinity, pressure, temperature, …  System design issues - equip with batteries, antennas, waterproof

7. Applications Sample Kalman Filter z -1 Weight Update MMSE Detector Sample Kalman Filter z -1 Weight Update MMSE Detector Multiuser Detection EKF Underwater Acoustic Receivers Radiolocation (GSIC) Filters (FIR, ARF, EWF)

8. System Design  Goal: Map application specification to system architecture  Subject to always increasing constraints – power, energy, latency, cost, size, … © Sangiovanni-Vincentelli

9. System Design and Architecture  Problem – take application code and map it to some system platform (e.g. reconfigurable device)  System platforms are extremely (and increasingly) complicated, multiprocessing computing systems  Mix of hardware and software components  Microprocessors – RISC, DSP, network, …  Logic level (FPGA) Reconfigurable logic  Specs for current high performance FPGA (Xilinx Virtex II)  3K to 125K logic cells,  Four PowerPC processor cores  Complex memory hierarchy - 1,738 KB block RAM, external memory, local memory in CLBs  Possibility of soft core processors – DSP  Custom hardware - embedded multipliers, fast carry chain logic, etc.  Large amount of performance improvement possible, IF there is a good mapping How do we best represent the application for mapping?

10. Obligatory Design Flow Slide.c program Syntactic/Semantic Analysis AST Parallelizing compiler transforms SUIF Function Level SSA CFG Generation SSA CFG Machine SUIF  Proc Backend x86RISC x86 Code RISC Code Profiler sample inputs PDG+SSA Generation Coarse-grain Optimizations System Partitioning HDL Generation Fine-grain Optimizations SSA CFG AST device architecture description System Compiler Synthesizable HDL Behavioral, Logic and Physical Synthesis Platform Programming Software bitstream Functional Embedded System Backend SSA CFG

11. Design Flow  Application specification  Can be written in C, SystemC, SystemVerilog, linear systems, signal flow graph, CDFGs  Must have front end to task graphs  Focusing first on a C to task graph Signal Flow Graph if(x < y) i = 10; else i = 255; while(i) x = y++; C codeLinear Systems

12. Intermediate Representation val = pred; for(i = 0; i<len; i++) val += diff; if(val > 32767) val = 32767; else if(val < -32768) val = -32768;  Must exploit fine AND coarse-grain parallelism  Ideally want automatic mapping  Need a form that can do synthesis to both hardware/software ?

13. PDG+SSA Representation val = pred; for(i = 0; i<len; i++) val += diff; if(val > 32767) val = 32767; else if(val < -32768) val = -32768; Input Application (in C) CDFG Form

14. PDG+SSA Representation CDFG FormPDG+SSA Form

15. Advantages of PDG+SSA  Exploits parallelism  Explicitly shows control and data dependences  Control structures do not limit data parallelism  Regions are hyperblocks – allows aggressive optimizations  Synthesis to hardware and software Looks complicated! What does it buy us?

16. Comparing CDFG, PDG  Benchmarks – bunch of MediaBench functions  PDG, CDFG 2-3 times faster than sequential execution  PDG about 7% faster than CDFG  PDG, CDFG approx. same area

17. Comparing Different Predicated Forms  Comparison with PSSA, sequential execution  PSSA - predicated static single assignment  Used by several other projects – CASH, Sea Cucumber  PDG+SSA on average 8% faster than PSSA

18. Map Application to HW/SW Cores  Dependence analysis to exploit fine/coarse grain parallelism  Interprocedural dependencies – selective inlining  Control dependencies – loop optimizations, hoisting, if conversion  Data dependencies – arrays, aliasing, liveness  System partitioning  Cluster into coarser grained tasks  Decide how to divide application onto platform

19. System Partitioning  How do you decide where to map different parts of the application?  Hardware or software – which processor, which memory, exact location, etc.  Extremely hard set of problems (NP-Hard)  Must be flexible - different applications/systems have wide variety of models  Fundamental problem - many different heuristic methods have been developed  Simulated annealing  Genetic Algorithms  Tabu Search  Kernighan/Lin  …

20. Task Graph Model  Application synthesis model  Directed Acyclic Graph  Each node is amount of computation  Coarse grained – loops, function calls, summations, filtering  Fine grained – addition, multiplication, comparison, shifting  System Partitioning  Map coarse grain task nodes onto a set of computational cores  Cores – RISC processor, CLBs, Digital Signal Processors, IPs, etc.

21. Our approach – Ant System Heuristic  Inspired by ethological study on the behavior of ants [Goss et. al. 1989]  A meta heuristic  A multi-agent cooperative searching method  A new way for combining global/local heuristics  Extensible and flexible

22. Ant System Heuristic

31. Autocatalytic Effect

32. Formulating Problems Using Ant Search  Problem model – define search space, create decision variables  Pheromone model – used as a global heuristic, distribution of pheromones, evaporation and strengthening strategies  Ant search strategy – local heuristics and solution space traversal  Solution construction – method of creating an answer from decision variables  Feedback – provide assessment of solution quality and adjust pheromones accordingly

33. Ant System Algorithm

34. System Partitioning Problem Model Example:  Task 1, 2, 7 and 8 are assigned to the GPP  Task 3, 4, and 6 onto the configurable logic  The inbound edges are colored accordingly  We don’t care the coloring for virtual nodes t 0 and t n  We don’t care the coloring for edge e 8n  Each task node is assigned a color  Color for each computational core

35. Pheromone Model  Each computing resource is assigned with a color c k  Each edge e ij is associated with a set of global heuristics (pheromone trails)  ij (k) indicating the favorableness for t j to be colored with c k  A coherent coloring is defined as:  Each task node in the DAG is colored  All the inbound edges of a task node have the same coloring as that of the corresponding task node

36. Ant Search Strategy  Each ant traverses the graph in topologically sorted order  Guarantees that each inbound edge to the current node has been already examined  At each node, the ant will:  Make guesses for the coloring of the successor nodes  Make decision on the coloring of the current node

37. Ant Search Strategy  At task node t i, the ant makes guesses the coloring for each of the successor nodes t j :   ij (k) : global heuristic on coloring t j with c k   j (k) : local heuristic on coloring t j with c k

38. Solution Construction  Upon entering a new task node t i, the ant makes a decision on the coloring of t i :  probabilistically based on the guesses made by all the immediate precedents of t i  Inbound edges are correspondingly colored once this decision is made

39. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

40. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

41. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

42. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

43. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

44. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

45. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

46. t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n

47. Find the best and update the pheromone trails based on the solution’s quality t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t 0 t n Next iteration Feedback

48. Experimental Setup  Testing benchmarks:  DAGs of different sizes are generated randomly with average branching factor of 5  Real functions (in C/C++) extracted from the MediaBench suits are mapped onto the task nodes  Tasks are analyzed using SUIF and Machine SUIF tools to achieve detailed CDFG level description  Simplified communication interface between tasks  Major problem: Real applications task graphs  Goal: Find the optimal resource partition that achieves the best worst case execution time under FPGA area constraint

49. Definitive Quality Assessment  91.7% of the results are within the top 3% 77% of the results of AS are within the top 2% 63.5% of the results are within top 0.1%  Comparing AS results with brute force search  Offers definitive measurement for the quality  Gives full solution space (can filter out EASY cases)

50. Result Quality Assessment  33 difficult testing cases  3 25 possible partitions  SA-50 has comparable run time as the AS  SA-500 and SA-1000 runs at 10 and 20 times  Larger testcases – too big for brute force search  Comparison with Simulated Annealing

51. Code Generation  Once task graph is partitioned  Generate code for each task  Create communication protocols  Bus creation & arbitration  Memory hierarchy  Currently assume simple direct communication  Need code generation from every input specification to every computational core  Software – use conventional compiler flow  Hardware – need flow from task to HDL  Scheduling is fundamental problem for both hardware and software synthesis

52. Instruction Scheduling  Given: set of instructions and collection of computational units  Instruction modeled using data flow graph (DFG)  Directed acyclic graph  Each node is instruction  Each edge is a data dependence  Find schedule for instructions to minimize some function (latency, area, power, …) Auto Regressive Filter

53. Instruction Scheduling  NP-hard  Fundamental problem - many different heuristic methods have been developed  ILP  Force directed  Genetic algorithm  Path based  Graph theoretic  Computational geometry  List scheduling + NOP   +< - - 1 2 3 4 v2v1 v3 v4 v5 vn v6 v7 v8 v9 v10 v11

54. List Scheduling  Simple and effective  Greedy strategy  Operation selection decided by criticality  O(n) time complexity  Make a priority list of the instructions based on some measure (mobility, instruction depth, number of successors, etc.)  No single priority function works well over all applications  Highly dependent on problem instance  Priority function quality highly varied

55. Combining Ants and Lists  Ants determine priority list  List scheduling framework evaluates the “goodness” of the list + NOP   +< - - 1 2 3 4 v2v1 v3 v4 v5 vn v6 v7 v8 v9 v10 v11

56. Ant Search Strategy  Every iteration each ant creates a priority list  Fill one instruction at a time  Memory about instructions already selected  At step j ant has already selected j-1 instructions  jth instruction selected probabilistically

57. Ant Search Strategy   ij (k) : global heuristic (pheromone) for selecting instruction i at j position   j (k) : local heuristic – can use different properties  Instruction mobility (IM)  Instruction depth (ID)  Latency weighted instruction depth (LWID)  Successor number (SN)  ,  control influence of global and local heuristics

58. Pheromone Model  Each instruction op i  I associated with n pheromone trails  where j = 1, …, n   ij (k) indicates the favorableness for op i to be positioned at jth position in the priority list  Initially all  set to fixed value  0  Evaporation rate ij

59. ARF Pheromones

60. Experimental Results  ILP (optimal) using CPLEX  List scheduling  Instruction mobility (IM), instruction depth (ID), latency weighted instruction depth (LWID), successor number (SN)  Ant scheduling results using different local heuristics (Averaged over 100 runs)

61. Other Topics of Research  Data & computation layout – simultaneous distribution of data/computation, utilizing on-chip block RAM, exploiting parallelism,  DSP synthesis – optimization of polynomial expressions, synthesizing multiple constant multiplications to hardware  Low power microarchitecture techniques – cache design to minimize leakage current, configurable memory hierarchy to minimize power, prefetching to minimize power

62. ExPRESS Group  ExPRESS - EXtensible, Programmable, Reconfigurable Embedded SystemS  Extensible – customized processors, configurable instruction sets  Programmable – post-manufacturing customization  Reconfigurable – rapid configuration changes ASIC RISC RAM FPGA ARM DSP System On Chip (SOC)

63. More ExPRESS Information…  Members  PhD Students  Wenrui Gong  Anup Hosangadi  Yan Meng  Gang Wang  Undergrads  Daniel Grund  Willis Hoang  Webpage - http://express.ece.ucsb.edu/http://express.ece.ucsb.edu/

64. Extra Slides

65. Radiolocation (GSIC)

66. Multiuser Detection Sample Kalman Filter z -1 Weight Update MMSE Detector Sample Kalman Filter z -1 Weight Update MMSE Detector

67. Underwater Acoustic Receiver EKF

68. Filters Finite Impulse Response

69. Auto Regressive Filter