← 研踪 PaperTrailarXiv:2611.15111 · v1
Speedup-Aware Rounding for Scheduling Malleable Tasks with Precedence Constraints
首次提交 2026-11-11 14:40
版本更新 2026-11-11 14:40
雷达观测 2026-11-13 11:11
摘要
We study the scheduling of malleable tasks with precedence constraints on identical processors to minimize the makespan, assuming that the processing time of a task does not increase and its work does not decrease with the number of processors, and that the work is a convex function of the processing time. The standard two-phase method solves a linear program, rounds the fractional processing times to allotments, caps the allotments at a threshold , and runs list scheduling. We propose \emph{speedup-aware rounding}, which rounds a task to its slower allotment exactly when this allotment is at most a factor slower than the fractional processing time. We determine the worst-case increase of the work under this rule exactly and show that no per-task rounding rule with time factor does better. With the algorithm achieves the approximation ratio for every , which improves the previous bound ; a single task with linear speedup shows that its ratio is at least . For concave speedup functions, which include the roofline, communication, Amdahl and power-law models, the ratio becomes , which improves the ratio of Jansen and Zhang. We also revisit the strictly decreasing case of our earlier work (Chen and Chu, 2013): the reported ratio relies on an elimination step that is not valid. In this case the algorithm coincides with that of Jansen and Zhang up to its parameters, and with parameters chosen by the corrected analysis it guarantees their ratio . Finally, without convexity no per-task rounding improves the known time--work trade-off, and the analysis cannot go below .
作者备注:31 pages; ancillary files contain the numerical verification programs
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-11 14:40 · 最新观测版本
Speedup-Aware Rounding for Scheduling Malleable Tasks with Precedence Constraints
We study the scheduling of malleable tasks with precedence constraints on identical processors to minimize the makespan, assuming that the processing time of a task does not increase and its work does not decrease with the number of processors, and that the work is a convex function of the processing time. The standard two-phase method solves a linear program, rounds the fractional processing times to allotments, caps the allotments at a threshold , and runs list scheduling. We propose \emph{speedup-aware rounding}, which rounds a task to its slower allotment exactly when this allotment is at most a factor slower than the fractional processing time. We determine the worst-case increase of the work under this rule exactly and show that no per-task rounding rule with time factor does better. With the algorithm achieves the approximation ratio for every , which improves the previous bound ; a single task with linear speedup shows that its ratio is at least . For concave speedup functions, which include the roofline, communication, Amdahl and power-law models, the ratio becomes , which improves the ratio of Jansen and Zhang. We also revisit the strictly decreasing case of our earlier work (Chen and Chu, 2013): the reported ratio relies on an elimination step that is not valid. In this case the algorithm coincides with that of Jansen and Zhang up to its parameters, and with parameters chosen by the corrected analysis it guarantees their ratio . Finally, without convexity no per-task rounding improves the known time--work trade-off, and the analysis cannot go below .
查看此版本解读 →查看此版本原文 ↗