1. SIMD and Associative Computational Models Part II: Associative Models

2. Associative Models Topics Introduction –References for Associative Computing Models –Motivation for the MASC model –The MASC and ASC Models –Programming characteristics for ASC Model –Overview of existing ASC algorithms and programs ASC Algorithms –Minimal Spanning Tree –Perhaps Graham Scan –String-Matching (Shannon Steinfadt) MASC simulations with other models (esp. MMB) Timing Justifications for Associative Operations A proposed IS control system for MASC (Wittaya Chantam)

3. Slides that Overlap with PDC Some slides in the Introduction will overlap with PDC Some of these slides are essential for this course. Other slides are included to give you a broader understanding of related information On overlaps slides, if they are not central to PDA course, then they will be covered fairly quickly. Students who have not had PDC course should plan to spend more time studying these slides.

4. Associative Computing References (listed in order of expected use) Jerry Potter, Johnnie Baker, et. al., Associative Computing - A Programming Paradigm for Massively Parallel Computers, Plenum Publishing Company, 1992. –Initial Assignment – Read material related to slides Maher M. Atwah, Johnnie W. Baker, and Selim Akl, An Associative Implementation of Graham's Convex Hull Algorithm,, IASTED International Conference on Parallel and Distributed Computing and Systems, October 1995. Mary Esenwein and Johnnie Baker, VLCD String Matching for Associative Computing and Multiple Broadcast Mesh, IASTED International Conference on Parallel and Distributed Computing and Systems, 1997. Johnnie Baker and Mingxian Jin, Simulation of Enhanced Meshes with MASC, a MSIMD Model, Proc. of the Eleventh IASTED International Conference on Parallel and Distributed Computing and Systems, Nov. 1999, 511-516.

5. Associative Computing References (cont.) Mingxian Jin & Johnnie Baker, Relating the Power of the MASC Model with that of Reconfigurable Bus-Based Models, International Parallel and Distributed Processing Symposium (APDCM Workshop), April 2007. Mingxian Jin, Wittaya Chantamas, Johnnie Baker, forthcoming paper on Comparing the Power of the MASC model to Reconfigurable Bus- Based Models. 2007-8. Mingxian Jin, Johnnie Baker, and Kenneth Batcher, Timings for Associative Operations on the MASC Model, International Parallel and Distributed Processing Symposium, IEEE Workshop on Massively Parallel Processing San Francisco, April 2001 Wittaya Chantamas, Johnnie Baker, and Michael Scherger, Compiler Extension of the ASC Language to Support Multiple Instruction Streams in the MASC Model using the Manager-Worker Paradigm, International Conference on Parallel and Distributed Processing Techniques and Applications (PDPTA’06), June 2006. Wittaya Chantamas and Johnnie Baker, A Multiple Associative Model to Support Branches in Data Parallel Applications using the Manager-Worker Paradigm, International Parallel and Distributed Processing Symposium (WMPP Workshop), April 2005.

6. WEBSITE FOR PAPERS http://www.cs.kent.edu/~parallel Follow the pointer to “papers” Note: The preceding papers and many others on ASC and MASC are available at the above website.

7. Introduction to Associative Models ASC and MASC

8. ASC & MASC are KSU Models Several professors and their graduate students at Kent State University have worked on models The STARAN and the ASPRO fully support the ASC model in hardware. The MPP supports it partly in hardware and partly in software. –Prof. Batcher was chief architect or consultant Dr. Potter developed a language for ASC Dr. Baker works on algorithms for models and architectures to support models Dr. Walker is working with the hardware design of the machine. Dr. Batcher and Dr. Potter are currently advisors

9. Motivation The STARAN Computer (Goodyear Aerospace, early 1970’s) and later the ASPRO provided an architectural model for associative computing embodied in the ASC model. ASC extends the data parallel programming style to a complete computational model. ASC provides a practical model that easily supports massive parallelism. MASC provides a hybrid data-parallel, control parallel model that supports associative programming. Descriptions of these models allow them to be compared to other parallel models

10. Parallel and Associative Computing Lab Parallel and Associative Research Group Associative Models of Computation Parallel Runtime Environments Parallel and Associative System Software Parallel and Associative Applications Associative and Parallel Algorithms

11. Associative Models of Computation Parallel Runtime Environments Parallel and Associative System Software Parallel and Associative Applications Associative and Parallel Algorithms Parallel and Associative Research Group ASC Processor Research Group FPGA-Based ASC Processor MASC Processor Structure Codes, ASC-centric Implementations Pipelined ASC w/ Reconfigurable Network Multithreaded ASC Processor

12. The ASC (Associative Computing) Model Computers supporting ASC model in hardware are Goodyear Aerospace’s STARAN USN ASPRO CELLNETWORKCELLNETWORK Memory CELLS  PE Memory PE Memory IS PE Instruction Stream

13. The ASC Model of Computation The ASC name is derived from ASsociative Computing) ASC Model: A SIMD model with certain additional constant time features. – Constant time features identified on next slide –These constant time features can be supported (less efficiently) in software by a traditional SIMD –The name “associative” is due to its ability to locate items in the memory of PEs by content rather than location. Uses associative features to simulate an associative memory Does not have an associative memory

14. The ASC Constant Time Properties Broadcast data in constant time Constant time global reduction of –Boolean values using AND/OR –Integer values using MAX/MIN Constant time associative search (see next slide) Responder processing –An IS can detect if a data test is satisfied by any of its cells in constant time (i.e., any-responders) –An IS can select one arbitrary responder in constant time (i.e., pick-one) Above properties can be supported in hardware with broadcast and reduction networks (see [2] below ). References: 1.Potter, Baker, Scott, Bansal, Leangsuksun, Asthagiri, ASC: An Associative Computing Paradigm. IEEE Computer, Nov. 1994, 19-25 2.M. Jin, J. Baker, and K. Batcher, Timings of Associative Operations on the MASC model, Workshop of Massively Parallel Processing, IPDPS ’01.

15. Dodge Ford Make Subaru Color PE1 PE2 PE3 PE4 PE5 PE6 PE7 red blue white red Year 1994 1996 1998 1997 Model Price On lot 1 1 0 0 0 0 1 Busy- idle 1 0 1 1 0 0 1 IS The Associative Search IS asks for all cars that are red and on the lot. PE1 and PE7 respond by setting a mask bit in their PE.

16. Detailed List of ASC Properties Instruction Stream –The IS has a copy of the program and can broadcast instructions to cells in unit time Cell Properties –Each cell consists of a PE and its local memory –All cells listen to the IS –A cell can be active, inactive, or idle Inactive cells listen but do not execute IS commands until reactivated Idle cells contain no essential data and are available for reassignment Active cells execute IS commands synchronously

17. Detailed List of ASC Properties (cont.) Responder Processing –The IS can detect if a data test is satisfied by any of its responder cells in constant time (i.e., any- responders?). –The IS can select an arbitrary responder in constant time (i.e., pick-one).

18. Constant Time Global Operations (across PEs) –Logical OR and AND of binary values –Maximum and minimum of numbers –Associative searches Communications –There are at least two real or virtual networks PE communications (or cell) network IS broadcast/reduction network (which could be implemented as two separate networks) Detailed List of ASC Properties (cont.)

19. –The PE communications network is normally supported by an interconnection network E.g., a 2D mesh –The broadcast/reduction network(s) are normally supported by a broadcast and a reduction network (sometimes combined). See posted paper by Jin, Baker, & Batcher (listed in associative references) Control Features –PEs and the IS and the networks all operate synchronously, using the same clock

20. The Associative Computing (ASC) Model (An Alternate Diagram) Cells

21. The MASC Model

22. MASC (i.e., Multiple ASC) is a multiple ASC model A Multiple SIMD (or MSIMD) model with more than one Instruction Stream (IS) Each IS can execute a separate data-parallel task –These threads execute to completion without interacting or interruption Dynamically reconfigurable –Each cell listens to only one IS –Cells can switch ISs, based on a data test. –Cells can switch between being active, inactive, or idle Each IS with its cells is an ASC model Job/functional parallelism is used to control the ISs using the IS network.

23. The MASC Model Basic Components –An array of cells, each consisting of simple PEs (or ALUs) and their local memory –An interconnection network between the cells –One or more instruction streams (ISs) –A control unit for the ISs MASC is a MSIMD model that supports –data parallel threads that execute to completion –Job parallelism used to control these threads –Uses associative programming techniques

24. MASC Basic Properties Each cell can listen to only one IS Cells can switch ISs in unit time, based on the results of a data test. Each IS and the cells listening to it follow rules of the ASC model. Control Features: –The PEs, ISs, and networks all operate synchronously, using the same clock –Restricted job control parallelism is used to coordinate the interaction of the multiple ISs.

25. Characteristics of Associative Programming Consistent use of style of programming called data parallel programming Consistent use of global associative searching and responder processing Usually, frequent use of the constant time global reduction operations: AND, OR, MAX, MIN Broadcast of data using IS bus allows the use of the PE network to be restricted to parallel data movement.

26. Characteristics of Associative Programming Tabular representation of data – think 2D arrays Use of searching instead of sorting Use of searching instead of pointers Use of searching instead of the ordering provided by linked lists, stacks, queues Promotes an highly intuitive programming style that promotes high productivity Uses structure codes (i.e., numeric representation) to represent data structures such as trees, graphs, embedded lists, and matrices. We’ll see examples of the above. –Ref: Nov. 1994 IEEE Computer article. –Also, see “Associative Computing” book by Potter.

27. Languages Designed for the ASC Professor Potter has created several languages for the ASC model. ASC is a C-like language designed for ASC model ACE is a higher level language that uses natural language syntax; e.g., plurals, pronouns. Anglish is an ACE variant that uses an English-like grammar (e.g., “their”, “its”) An OOPs version of ASC for the MASC was discussed (by Potter and his students), but never designed. Language References: –ASC Primer – Copy available on parallel lab website www.cs.kent.edu/~parallel/ www.cs.kent.edu/~parallel/ –“Associative Computing” book by Potter [11] – some features in this book were never fully implemented in ASC Compiler

28. Dodge Ford Make Subaru Color PE1 PE2 PE3 PE4 PE5 PE6 PE7 red blue white red Year 1994 1996 1998 1997 Model Price On lot 1 1 0 0 0 0 1 Busy- idle 1 0 1 1 0 0 1 IS Typical Data Structure for ASC Model Make, Color – etc. are fields the programmer establishes Various data types are supported. Some examples will show string data, but they are not supported in the ASC simulator.

29. ASC Algorithms and Programs A wide range of algorithms implemented in ASC without the use of the PE network: –Graph Algorithms minimal spanning tree shortest path connected components –Computational Geometry Algorithms convex hull algorithms (Jarvis March, Quickhull, Graham Scan, etc) Dynamic hull algorithms

30. ASC Algorithms and Programs (not requiring PE network) –String Matching Algorithms all exact substring matches all exact matches with “don’t care” (i.e., wild card) characters. –Algorithms for NP-complete problems traveling salesperson 2-D knapsack. –Data Base Management Software associative data base relational data base

31. ASC Algorithms and Programs (not requiring a PE network) –A Two Pass Compiler for ASC – not the one we will be using. This compiler uses ASC parallelism. first pass optimization phase –Two Rule-Based Inference Engines for AI An Expert System OPS-5 interpreter PPL (Parallel Production Language interpreter) –A Context Sensitive Language Interpreter (OPS-5 variables force context sensitivity) –An associative PROLOG interpreter

32. Associative Algorithms & Programs (using a network) There are numerous associative programs that use a PE network; –2-D Knapsack ASCAlgorithm using a 1-D mesh –Image processing algorithms using 1-D mesh –FFT (Fast Fourier Transform) using 1-D nearest neighbor & Flip networks –Matrix Multiplication using 1-D mesh –An Air Traffic Control Program (using Flip network connecting PEs to memory) Demonstrated using live data at Knoxville in mid 70’s. All but first were developed in assembler at Goodyear Aerospace

33. ASC Algorithms Minimal Spanning Tree Graham Scan (postponed) String Matching

34. Example 1 - MST A graph has nodes labeled by some identifying letter or number and arcs which are directional and have weights associated with them. Such a graph could represent a map where the nodes are cities and the arc weights give the mileage between two cities. A B C D E 3 52 5 4

35. The MST Problem The MST problem assumes the weights are positive and the graph is connected. Its goal is to find the minimal spanning tree, – i.e. a subgraph that is a tree 1, that includes all nodes (i.e. it spans), and –where the sum of the weights on the arcs of the subgraph is the smallest possible weight (i.e. it is minimal). Why would an algorithm solving this problem be useful? Note: The solution may not be unique. 1 A tree is a set of points called vertices, pairs of distinct vertices called edges, such that (1) there is a sequence of edges called a path from any vertex to any other, and (2) there are no circuits, that is, no paths starting from a vertex and returning to the same vertex.

36. An Example DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 As we will see, the algorithm is simple. The ASC program is quite easy to write. A SISD solution is a bit messy because of the data structures needed to hold the data for the problem

37. An Example – Step 0 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 We will maintain three sets of nodes whose membership will change during the run. The first, V1, will be nodes selected to be in the tree. The second, V2, will be candidates at the current step to be added to V1. The third, V3, will be nodes not considered yet.

38. An Example – Step 0 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 V1 nodes will be in red with their selected edges being in red also. V2 nodes will be in blue with their candidate edges in blue also. V3 nodes and edges will remain black

39. An Example – Step 1 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Select an arbitrary node to place in V1, say A. Put into V2, all nodes incident with A.

40. An Example – Step 2 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Choose the edge with the smallest weight and put its node, B, into V1. Mark that edge with red also. Retain the other edge-node combinations in the “to be considered” list.

41. An Example – Step 3 D E HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Add all the nodes incident to B to the “to be considered list”. However, note that AG has weight 3 and BG has weight 6. So, there is no sense of including BG in the list.

42. An Example – Step 4 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Add the node with the smallest weight that is colored blue and add it to V1. Note the nodes and edges in red are forming a subgraph which is a tree.

43. An Example – Step 5 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Update the candidate nodes and edges by including all that are incident to those that are in V1 and colored red.

44. An Example – Step 6 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Select I as its edge is minimal. Mark node and edge as red.

45. An Example – Step 7 DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Add the new candidate edges. Note that IF has weight 5 while AF has weight 7. Thus, we drop AF from consideration at this time.

46. An Example – after several more passes we have … DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 Note that when CH is added, GH is dropped as CH has less weight. Also, BC is dropped for the same type of reasoning (i.e., it would form a back edge between two nodes already in the MST). When there are no more nodes to be considered, i.e. no more in V3, we obtain the final solution.

47. An Example – the final solution DE HI FC G BA 8 6 5 3 3 2 2 2 1 6 1 4 2 47 The subgraph is clearly a tree – no cycles and connected. The tree spans – i.e. all nodes are included. While not obvious, it can be shown that this algorithm always produces a minimal spanning tree. The algorithm is known as Prim’s Algorithm for MST.

48. ASC-MST Algorithm Preliminaries Next, a “data structure” level presentation of Prim’s algorithm for the MST is given. The data structure used is illustrated in the next two slides. –This example is from the Nov. 1994 IEEE Computer paper cited in the references. There are two types of variables for the ASC model, namely –the parallel variables (i.e., ones for the PEs) –the scalar variables (ie., the ones used by the IS). –Scalar variables are essentially global variables. Can replace each with a parallel variable with this scalar value stored in each entry.

49. ASC-MST Algorithm Preliminaries (cont.) In order to distinguish between them here, the parallel variables names end with a “$” symbol. Each step in this algorithm takes constant time. One MST edge is selected during each pass through the loop in this algorithm. Since a spanning tree has n-1 edges, the running time of this algorithm is O(n) and its cost is O(n 2 ). –Definition of cost is (running time)  (number of processors) Since the sequential running time of the Prim MST algorithm is O(n 2 ) and is time optimal, this parallel implementation is cost optimal. –Cost & optimality will be covered in parallel algorithm performance evaluation chapter (See Ch 7 of Quinn)

50. Graph used for Data Structure Figure 6 in [Potter, Baker, et. al.] a bc d e f 2 8 9 6 3 3 4 7 2

51. Data Structure for MST Algorithm

52. Algorithm: ASC-MST-PRIM(root) 1.Initialize candidates to “waiting” 2.If there are any finite values in root’s field, 3. set candidate$ to “yes” 4. set parent$ to root 5. set current_best$ to the values in root’s field 6. set root’s candidate field to “no” 7.Loop while some candidate$ contain “yes” 8. for them 9. restrict mask$ to mindex(current_best$) 10. set next_node to a node identified in the preceding step 11. set its candidate to “no” 12. if the value in their next_node’s field are less than current_best$, then 13. set current_best$ to value in next_node’s field 14. set parent$ to next_node 15. if candidate$ is “waiting” and the value in its next_node’s field is finite 16. set candidate$ to “yes” 17. set parent$ to next_node 18. set current_best to the values in next_node’s field

53. Comments on ASC-MST Algorithm The three preceding slides are Figure 6 in [Potter, Baker, et.al.] IEEE Computer, Nov 1994]. Figure 6c gives a compact, data-structures level pseudo-code description for this algorithm –Pseudo-code illustrates Potter’s use of pronouns (e.g., them) and possessive nouns. –The mindex function returns the index of a processor holding the minimal value. –This MST pseudo-code is much shorter and simpler than data-structure level sequential MST pseudo- codes e.g., see one of Baase’s textbook cited in references Algorithm given in Baase’s books is essentially the same as this parallel algorithm Next, a more detailed explanation of the algorithm in preceding slide will be given.

54. Algorithm: ASC-MSP-PRIM Initially assign any node to root. All processors set –candidate$ to “waiting” –current-best$ to  –the candidate field for the root node to “no” All processors whose distance d from their node to root node is finite do –Set their candidate$ field to “yes –Set their parent$ field to root. –Set current_best$ = d.

55. Algorithm: ASC-MSP-PRIM (cont. 2/3) While the candidate field of some processor is “yes”, –Restrict the active processors whose candidate field is “yes” and (for these processors) do Compute the minimum value x of current_best$. Restrict the active processors to those with current_best$ = x and do –pick an active processor, say one that contains node y. »Set the candidate$ value of node y to “no” –Set the scalar variable next-node to y.

56. Algorithm: ASC-MSP-PRIM (cont. 3/3) If the value z in the next_node column of a processor is less than its current_best$ value, then –Set current_best$ to z. –Set parent$ to next_node –For all processors, if candidate$ is “waiting” and the distance of its node from next_node is not , then Set candidate$ to “yes” Set current_best$ to the distance of its node from next_node. Set parent$ to next-node

57. Associative Graham Scan Material Postponed this year. See “Graham Scan” Reference for further information

58. Associative String Matching Shannon Steinfadt

59. Shannon’s Slides Shannon’s slides belong here. Currently, they will be posted separately on our webpage.

60. For Additional ASC and MASC Algorithms See Slides for PDC’06 course or other ASC/MASC papers at www.cs.kent.edu/~parallel/ Also, MMB Simulations provide many algorithms for MASC

61. Simulations between MASC and other models (Primarily MMB Model)

62. Simulation References Johnnie Baker and Mingxian Jin, Simulation of Enhanced Meshes with MASC, a MSIMD Model, Proc. of the Eleventh IASTED International Conference on Parallel and Distributed Computing and Systems, Nov. 1999, 511-516. –Reading Assignment & Primary Reference: Read material that is related to slides. Mingxian Jin, Wittaya Chantamas, Johnnie Baker, forthcoming paper on comparing the power of the MASC model to reconfigurable bus-based models. 2007-8. Use following reference until this becomes available. –Mingxian Jin & Johnnie Baker, Relating the Power of the MASC Model with that of Reconfigurable Bus-Based Models, International Parallel and Distributed Processing Symposium (APDCM Workshop), April 2007.

63. Previous MASC Simulation MASC Simulation of PRAM –MASC(n,j) can simulate priority CRCW PRAM(n,m) in O(min{n/j, m/j}) with high probability. –MASC(n,1) [or ASC] can simulate priority CRCW with a constant number of global memory locations in constant time (Avg Case) This result is stronger than it first appears Many CRCW algorithms only require a constant nr of global memory locations –Recent related results of Mingxian Jin mentioned later

64. Previous MASC Simulation (cont) Self-simulation of MASC –Provides an efficient algorithm for MASC to efficiently simulate a larger MASC - with more PEs and/or ISs. –Establishes that MASC is highly scalable –MASC(n,j) can simulate MASC(N,J) in O(N/n + J) extra time and O(N/n + J) extra memory. Best reference on above is Darrell Ulm’s Dissertation

65. The Enhanced Mesh, MMB References: –[Baker & Jin], Reference listed on Slide 19 –Mingxian Jin, Evaluating the power of the parallel MASC model using simulations and Real-Time Applications, KSU Dissertation Aug. 2004, 145 pages. Enhanced meshes are basic mesh models augmented with fixed or reconfigurable buses –At most one PE on a bus can broadcast to remaining PEs during one step.

66. The Enhanced Mesh, MMB Best-known fixed bus example: –Mesh with multiple broadcasting (MMB) –Standard 2-D mesh –Row and column bus enhancements –Broadcasts can occur along only row or column buses (but not both) in one step

67. The Reconfigurable Enhanced Mesh RM For all reconfigurable bus models, buses are created dynamically during execution Best known example: –General Reconfigurable Mesh (RM) –Each PE has four ports called N,S, E, W (often called “NEWS”) –In one step, each PE can set the connections of its ports, based on local data –At most two disjoint pairs of ports can be connected at any time –One such connection is the adjacent pairs, {{N,E}, {W,S}}.

68. Reconfigurable Mesh Architecture

69. Simulation Preliminaries Reasons to simulate other models using MASC –Allows a better understanding of the power of MASC –Provides a simulation algorithm that can be used to convert algorithms designed for the other model to MASC –Provides alternate methods to support MASC. Basic Assumption Used in the Simulations –MASC(n, ) has a mesh PE network with row-major ordering –The enhanced meshes have a 2D mesh with the same size and ordering

70. Simulation Preliminaries (cont.) Basic Assumption Used in the Simulations (cont.) –Each PE in MASC has the same computational power as an enhanced mesh PE –The MASC buses and the buses of the enhanced mesh have the same characteristics –The word lengths of both models are the same and at least  lg(n) . –Each PE in MASC knows its position in the 2D mesh. Words can store the positions of various PEs

71. Simulating MMB using MASC The mapping is between MASC(n, ) and Enhanced meshes of size The mapping assigns a PE in one model to the PE that is in the same position in the 2D mesh in the other model The i-th IS in MASC simulates both the i-th row and the i-th column buses

72. Simulating MMB using MASC

73. Simulating MMB with MASC Since both models have identical 2D meshes, these do not need to be simulated Since the power of PEs in respective models are identical, their local computations are not simulated To simulate a MMB row broadcast on the MASC, –All PEs switch to their assigned row IS –The IS for each row checks to see if there is a PE that wishes to broadcast –If true, the IS broadcasts this value to all of its PEs (i.e., the ones on its assigned row). Simulation of a MMB column broadcast is similar The running time is O(1)

74. MASC More Powerful Than MMB The MASC model is strictly more powerful than the MMB model when both models have the “same size”. –Since MASC can simulate MMB in O(1) time, MASC is as powerful as MMB. –There is a problem which can be solved in constant time using MASC(n, ) with a mesh but which requires  (n log n) time for a MMB to solve. See “Simulation of Enhanced Meshes with MASC, a MSIMD Model” by Baker & Jin (listed in references & posted on web)

75. Simulation of MMB with MASC Theorem 1. MASC(n, j) with a 2-D mesh is strictly more powerful than a MMB for j =  ( ). An algorithm for a MMB can be executed on MASC(n, j) with j=  ( ) and a 2-D mesh with a running time at least fast as the MMB time.

76. Simulation of MASC by MMB PE(1,1) stores a copy of the program and simulates the ISs sequentially. Each instruction stream command or datum is first sent by P(1,1) to the PEs in the first column. Next, the PEs in the first column broadcast this command or datum along the rows to all PEs. Each MMB processor uses two registers, channel and status, to decide whether or not to execute the current instruction. –channel records which IS the processor is assigned to –status records whether PE is active, inactive, idle

77. Simulation of MASC by MMB The simulation of simultaneous broadcasts of ISs takes O( ) time. A local computation, memory access, or a data movement along local links are identical in the two models and require O(1) time. The execution of a global reduction operator OR, AND, MAX, MIN takes O( ) using an optimal MMB algorithm (MMB reduction published by Olariu, et. al.). Since the global reduction operators may be required for O( ) ISs, an upper bound is O( ) or O( ).

78. Simulation of MASC by MMB Theorem 2. MASC(n, ) with a 2-D mesh can be simulated by a MMB in O( ) time with O( ) extra memory

79. Power Comparisons between MASC, RM, and PRAM Here, we allow the size of the simulating model to be a polynomial times larger than the simulated model. – E.g., MASC(P(n),Q(m)) can be used to simulate PRAM(n,m), where P(n) and Q(m) are polynomials in n and m, respectively. Even if the running time of a simulation is small, its “cost” can be huge due to polynomial increase in size of the simulating model. MASC and PRAM can each simulate each other in O(1) time, so they have the same power. RM can simulate both MASC and RM in O(1) time, but neither MASC nor PRAM can simulate RM in O(1) time. RM is more powerful than either MASC or PRAM.

80. Simulation Conclusions MASC is strictly more powerful than an MMB of the same size. Any algorithm for an MMB can be executed on a MASC of the same size with the same running time. In particular, –Optimal algorithms for MMB are also optimal when executed on MASC By using a polynomial times large simulating model, O(1) simulations can be established between MASC and PRAM. A polynomimal times larger RM can simulate MASC and PRAM in O(1) time A polynomial times larger MASC or PRAM can not simulate RM in O(1) time.

81. Timings for Associative Operations Mingxian Jin Johnnie Baker Kenneth Batcher Kent State University

82. OUTLINE The ASC/MASC Constant Time Operations Broadcast/reduction network Discussion on the basic operations Comparison of timings with other models Summary Reference –Mingxian Jin, Johnnie Baker, and Kenneth Batcher, Timings for Associative Operations on the MASC Model, International Parallel and Distributed Processing Symposium, IEEE Workshop on Massively Parallel Processing San Francisco, April 2001

83. Recall ASC & MASC Models PEMemory IS C E L L N E T W O RK I S N E T W O RK PEMemory PEMemory Cells    Broadcast/reduction network

84. Basic Constant-Time Operations Broadcasting Global reduction of logic OR/AND Global reduction of Maximum/Minimum Associative search –Responders –Non-responders The AnyResponder operation The PickOne operation

85. Motivation The basic operations are essential to support associative style of computing Accuracy and fairness of the timings –determine the accuracy of the comparison between MASC and other models –determine the ability of MASC to efficiently support algorithm design and complexity analysis Simulation of MMB has raised some questions about the assigned timings Evidence is needed to justify the correctness of timings Correctness depends on –possible hardware implementations –comparative fairness with respect to other model

86. Broadcast/reduction network Constructed using a group of resolver circuits –R i : a responder bit of PE i –V i = R 0  R 1 ...  R i-1 –A resolver tells whether any earlier PE has R i equal to 1 STARAN-- architectural motivation of MASC –An associative SIMD computer built in 1970’s by Goodyear Aerospace –Possible hardware implementation of the MASC operations

87. A 4-PE resolver PE i RiRi PE i+1 R i+1 PE i+2 R i+2 PE i+3 R i+3 ViVi V i+1 V i+2 V i+3 RiRi R i+1 R i+2 R i+3 V i = R 0  R 1 ...  R i-1 R i  R i+1  R i+2  R i+3 Figure 2. A 4-PE resolver with at most 1-gate delay from any input to any output

88. A 16-PE Resolver PE i PE i+15 4-PE Resolver ViVi R i  R i+1  …  R i+15 Figure 3. A 16-PE resolver with at most a 3-gate delay from any input to any output

89. Gate Delay for Broadcast/Reduction The gate delay on the MASC broadcast/reduction network is at most (2log 4 N –1) If a gate delay takes about 1-5 nanoseconds. –A machine with 2 2 10 processors would have the gate delay at most (2log 4 2 2 10 -1)  5 nanoseconds  5.1 microseconds –A machine with 100 million processors would have the gate delay less than 50 nanoseconds, which is comparable to the time for a memory access in today’s systems. A machine with 2 2 10 processors is impractical –This is greater than the number of atoms in the observable universe

90. The gate delay of (2log 4 N -1) using a regular scale

91. The gate delay of (2log 4 N -1) with the vertical axis using a logarithmic scale

92. Recall Constant Time Associative Operations Broadcasting Global OR and AND operations Associative search The AnyResponder operation The PickOne operation Global maximum and minimum operations

93. Broadcasting 1 bit -- O(1) and  bits -- O(  ) Let  be the length of an instruction or a data item For bus-based architectures, buses normally have bandwidth of  =  logN  where N is the number of PEs, –Allows a processor ID to be stored in a word –Words can be transmitted in one step Separate broadcast network can be built for MASC with the bus bandwidth  For MASC, both the word and instruction broadcasts require constant time

94. Global OR and AND Timing Is computed through the resolver network, like broadcasting However, the tree traversal is in the reverse direction Reasonable to assume this requires constant time

95. Associative Search Operation Timing A IS determines if any of its active PEs contains a (word-length) search pattern The search pattern is broadcast to the PEs by the IS Next, each active PE makes a sequential comparison. Those PEs with matching values set their responder bit and remain active Same tree travel through the network as above Operation requires constant time

96. AnyResponder Operation Timing Usually follows an associative search Returns true if any responder bit is set and false otherwise Essentially a global OR operation

97. PickOne Operation Timing Usually follows an associative search The IS can instruct this PE to broadcast a value After a PE is processed, the responder bit is cleared and the IS can pick another active PE

98. Global Maximum and Minimum Timings Implemented bit-serially with a global AND in the order of the left bit to the right bit. Each global AND keeps active those PEs whose bit value is 1 (or 0 for Minimum) until –one responder is left or –all bits have been processed If all active processors have a 0 value for one bit, then all remain active for the next round. With the word length , this operation takes O(  ) An addition of two word length operands is normally assumed to be an O(1) operation, i.e.,  is a constant Similarly, a global Maximum and Minimum can be justified to be a constant time operation

99. RAM & PRAM Memory Access Time A lower bound established using electronic circuit access –Called a Memory Access Unit (MAU) MAU implemented as a binary tree of switches rooted at each processor When the memory size is M, the memory access time is –  (log M) for RAM –  (log N) for PRAMs with N processors (assuming N=  (M) Both assume the memory access takes constant time

100. Mesh with Multiple Broadcast (MMB) A MMB is a basic mesh enhanced with row and column buses In one time unit, only one processor is allowed to broadcast All other processors read the value being broadcast in constant time More practical compared to PRAMs On a N  N MMB and each PE holding a data item, a global reduction can be computed in O(N ). MMB perform these reductions by designing specific algorithms rather than using circuits Substantially different methods are used to execute these operations on MASC

101. Summary of Associative Operations Timings Operations 1-bit Broadcast Bus  -bit Broadcast Bus Broadcast O(  ) O(1) Addition/Subtractio n O(1) Logic ORO(1) Logic ANDO(1) Associative Search O(  ) O(1) AnyResponderO(1) PickOneO(1) Maximum/Minimu m O(  ) O(1)

102. A Proposed IS Control System for MASC Wittaya Chantamas These slides will be posted separately on webpage.

103. Wittaya’s slides go here This is the location that Wittaya’s slide belong. However, they will be posted separately on the webpage.

104. Slides for Possible Future Use

105. MASC Model Basic Components –An array of cells, each consisting of a PE and its local memory –A PE interconnection network between the cells –One or more Instruction Streams (ISs) –An IS network MASC is a MSIMD model that supports –both data and control parallelism –associative programming Memory PE Interconnection Network IS Network PE Instruc- tion Stream (IS) Instruc- tion Stream (IS) Instruc- tion Stream (IS)