220 likes | 327 Vues
This presentation introduces a fast static scheduling algorithm optimized for Directed Acyclic Graphs (DAGs) on an unbounded number of processors. The algorithm, including EZ, DSC, and Clan methods, aims to minimize total processing time and reduce critical path lengths through clustering and heuristic approaches. We explore tree-structured task graphs, communication overhead, and scheduling classifications, comparing various algorithms across different test graphs to determine efficiency and speedup. The presentation highlights the advantages and complexities of Clan at varying granular levels, offering insights for broader application scenarios. References include seminal works on static scheduling algorithms for parallel processing systems.
E N D
Fast Static Scheduling Algorithm for DAGs on an Unbounded Number of Processors Speaker:Si-Wen Hung Advisor:Dr. Sao-Jie Chen
Outline • Introduction • Scheduling Classification • EZ • DSC • Clan • Comparison
Introduction • Scheduling • Assign all the tasks to PE • Minimize the total processing
Introduction • Tasks graph • Communication overhead
Introduction • Tree-structured task graph
Introduction • Critical path heuristic • In order to reduce the critical path by clustering • DSC • MCP • EZ
Introduction • Assumptions: • Task duplication is not allowed • The number of available processors is unlimited • The task execution is triggered by the arrival of all data and at the completion of its execution the data are send in parallel to successor tasks.
Algorithm.1 • Clustering steps by Sarkar’s algorithm(EZ) • Initially each task is in a separate cluster • Sort all the edges from high cost to low cost • For each edge from the sorted edge list • If set the edge to zero cost would reduce parallel time • If yes , set the two nodes of the edge to the same cluster
Algorithm.1 • Example
Algorithm.2 • Dominant sequence clustering algorithm (DSC) • Partial free list (PFL),Free list (FL) • Select the highest priority of node from PHL and FL • If set the node to the same cluster would reduce parallel time • If yes , set the two nodes of the edge to the same cluster • Otherwise, open new cluster for the node
Algorithm.2 • Example
Clan • Type • Linear • Independent • Parse Tree • Hierarchical view
Clan • Parse Tree • Hierarchical view
Clan • Multi-Stage Decision Graph • Find the shortest path
Comparison • Granularity Analysis • 420 test graph • Speedup < 1
Comparison • Normalized Relative Parallel Time • Average speed up
Comparison • Efficiency of the algorithm
Conclusion • Clan has high speed up ,but more complexity at low granularity • Clan is not better than other at high granularity • Clan suit the cases of wide range of granularity or low granularity
References • A. Gerasoulis and T. Yang, "A Fast Static Scheduling Algorithm for DAGs on a unbounded Number of Processors," Proc. of Supercomputing'91, (Nov. 1991), pp.633-642. • T. Yang and A. Gerasoulis, "A Fast Static Scheduling Algorithm for DAGs on an Unbounded Number of Processors", Proc. Supercomputing '91, pp. 633-642 (1991). • A. A. Khan, C. L. McCreary and Y. Gong, A Numerical Comparative Analysis of Partitioning Heuristics for Scheduling Task Graphs on Multiprocessors, October 21, 1993.