A Communication Ordered Task Graph Allocation Algorithm

The inherently asynchronous nature of the data ow computation model allows the exploitation of maximum parallelism in program execution While this computational model holds great promise several problems must be solved in order to achieve a high degree of program performance The allocation and scheduling of programs on MIMD distributed memory parallel hardware is necessary for the implementation of e cient parallel systems Finding optimal solutions requires that maxi mum parallelism be achieved consistent with resource limits and minimizing communication costs and has been proven to be in the class of NP complete problems This paper addresses the problem of static allocation of tasks to distributed memory MIMD systems where simultaneous computation and communication is a factor This paper discusses similarities and di erences between several recent heuristic allocation approaches and identi es common problems inherent in these approaches This paper presents a new algorithm scheme and heuristics that resolves the identi ed problems and shows signi cant performance bene ts A Communication Ordered Task Graph Allocation Algorithm John D Evans and Robert R Kessler April Abstract The inherently asynchronous nature of the data ow computation model allows the exploitation of maximum parallelism in program execution While this computational model holds great promise sev eral problems must be solved in order to achieve a high degree of program performance The allocation and scheduling of programs on MIMD distributed memory parallel hardware is necessary for the imple mentation of e cient parallel systems Finding optimal solutions requires that maximumparallelism be achieved consistent with resource limits and minimizing communication costs and has been proven to be in the class of NP complete problems This paper addresses the problem of static allocation of tasks to distributed memory MIMD systems where simultaneous computation and communication is a factor This paper discusses similarities and di erences between several recent heuristic allocation approaches and identi es common problems inherent in these approaches This paper presents a new algorithm scheme and heuristics that resolves the identi ed problems and shows signi cant performance bene tsThe inherently asynchronous nature of the data ow computation model allows the exploitation of maximum parallelism in program execution While this computational model holds great promise sev eral problems must be solved in order to achieve a high degree of program performance The allocation and scheduling of programs on MIMD distributed memory parallel hardware is necessary for the imple mentation of e cient parallel systems Finding optimal solutions requires that maximumparallelism be achieved consistent with resource limits and minimizing communication costs and has been proven to be in the class of NP complete problems This paper addresses the problem of static allocation of tasks to distributed memory MIMD systems where simultaneous computation and communication is a factor This paper discusses similarities and di erences between several recent heuristic allocation approaches and identi es common problems inherent in these approaches This paper presents a new algorithm scheme and heuristics that resolves the identi ed problems and shows signi cant performance bene ts Introduction A primary goal of program allocation and scheduling is to minimize program run time The goals of program allocation vary however with the circumstances of hardware and software requirements Memory size limits restrictions on external I O access and processor failure are several considerations that may take precedence over execution time In addition the existence of shared memory distributed memory and the possible duplication of computations to eliminate communication may signi cantly alter the nature of the allocation problem The inherently asynchronous nature of the data ow computation model o ers promising improvement in performance over the control ow programming model Data ow research has focused on developing high speed data ow multiprocessors on recognizing the potential parallelism in sequential programs achieving parallel program performance developing parallel languages Recent research has also ad dressed the extension of data ow concepts to the execution of programs on non data ow architectures The data ow computation model describes computations as a collection of work units called tasks Individual tasks are sequential segments of program execution which can run from start to completion without synchronization Tasks require input data to begin execution and the computational results of tasks are then used as the input to subsequent tasks A task is eligible for execution when all of its inputs are ready Precedence relationships are de ned between pairs of tasks if the results from one task are