Izboljšava algoritmov statičnega razvrščanja z nestrogim proženjem vozlišč programskega grafa

Extended abstract. Obtaining maximum performance from parallel distributed memory machines with a large number of processors depends on partitioning parallel programs into modules and scheduling these modules for the shortest possible execution time. The scheduling problem is known to be NP-hard [2] and heuristic algorithms have been proposed to obtain optimal and sub optimal solutions [1,3,10,11]. In this paper, we present three efficient static algorithms for scheduling modules to the processing units of a parallel computer system. They are based on CPM [14], VL [5] and DSH [6,1] scheduling algorithms. We improved them with nonstrict trigerring of coarse grained program graph nodes. The partitioning algorithm [12,9] partitions an application into tasks with an appropriate grain size and represents them in the form of a program graph (PG). If the sum of communication delays for transferring operands between nodes is greater than the sum of their execution times, it is possible to achieve faster execution time by joining them in a larger program grain. It is interesting that joining small grains into larger ones enables the processing element to start execution of a new grain without all input operands. We call this kind of execution a nonstrict triggering of program graph nodes. An example of a nonstrictly triggered node is shown in Fig. 1. We introduced two new attributes for each communication arc. CS represents the operand’s relative sending time. It defines the time from the moment when the operand is sent to the end of a node execution. CR represents the operand’s relative receiving time. It defines the time from the beginning of a program graph node execution to the moment when this operand must be present to avoid delaying the node’s execution. We present our model of a macro dataflow computer (MMDC, Fig. 2) which supports nonstrict triggering of coarse