Contention Bounds for Locking Computations
Name
Li-liwanlin-meng-eecs-2022-thesis.pdf
Description
Thesis PDF
Size
938.24 KB
Format
Adobe PDF
Checksum (MD5)
28f1d88d66fe142f6f5fc592e628550d
Author(s)
Li, Wanlin
Advisor(s)
Leiserson, Charles E.
Kuszmaul, William
Date Issued
May 2022
Publisher
Massachusetts Institute of Technology
Abstract
This thesis quantifies lock contention in multithreaded programs by expanding the theoretical model of task-parallel execution traces to account for mutual exclusion locks. While lock profiling and contention detection tools abound in software, empirical measurements of contention suffer from wide fluctuations across different executions of the same code due to scheduling variation and processor availability. In this work we present analytical bounds on the maximum possible contention incurred by a given program over all possible execution schedules, even when running alongside other programs in a busy environment or when scheduled by an adversary. Although we show that computing the exact optimum is NP-hard for general task graphs, in the restricted case of fork-join (series-parallel) computations with 𝑛 strands and a single lock we offer a Θ(𝑛²) exact algorithm as well as a Θ(𝑛) 2-approximation for worst case contention. In proving these bounds linking maximum contention to the antichain sizes of a program’s parallel trace, we establish graph-based properties of worst case execution schedules that also apply directly to the related single processor scheduling problem of average response time under task precedence constraints. In addition, our analysis of worst case contention offers improved estimates for the completion time of locking computations under the execution trace model.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright MIT
Persistent DSpace Link
