Scheduling computations with provably low synchronization overheads

We present a Work Stealing scheduling algorithm that provably avoids most synchronization overheads by keeping processors’ deques entirely private by default and only exposing work when requested by thieves. This is the first paper that obtains bounds on the synchronization overheads that are (essentially) independent of the total amount of work, thus corresponding to a great improvement, in both algorithm design and theory, over state-of-the-art Work Stealing algorithms. Consider any computation with work T1 and critical-path length T1 executed by P processors using our scheduler. Our analysis shows that the expected execution time is O T1 P + T1 , and the expected synchronization overheads incurred during the execution are at most O ((CCAS + CMF ence) P T1), where CCAS and CMF ence respectively denote the maximum cost of executing a Compare-And-Swap instruction and a Memory Fence instruction.

Performance Analysis of Work Stealing in Large-scale Multithreaded Computing

Randomized work stealing is used in distributed systems to increase performance and improve resource utilization. In this article, we consider randomized work stealing in a large system of homogeneous processors where parent jobs spawn child jobs that can feasibly be executed in parallel with the parent job. We analyse the performance of two work stealing strategies: one where only child jobs can be transferred across servers and the other where parent jobs are transferred. We define a mean-field model to derive the response time distribution in a large-scale system with Poisson arrivals and exponential parent and child job durations. We prove that the model has a unique fixed point that corresponds to the steady state of a structured Markov chain, allowing us to use matrix analytic methods to compute the unique fixed point. The accuracy of the mean-field model is validated using simulation. Using numerical examples, we illustrate the effect of different probe rates, load, and different child job size distributions on performance with respect to the two stealing strategies, individually, and compared to each other.

About optimal managment of work-stealing deques in two-level memory

В работе рассмотрена задача оптимального управления work-stealing деком (англ. –- deque) в двухуровневой памяти. Предполагается, что известны вероятности параллельных операций с деком и временные характеристики памяти для двух уровней. Задача состоит в нахождении оптимального числа элементов с двух сторон дека, которые при перераспределении дека должны быть оставлены в быстрой памяти. В качестве критерия оптимальности рассмотрены минимальные средние затраты на перераспределение памяти, которые возникают в случае переполнения или опустошения быстрой памяти. Такой критерий позволяет учитывать конкретные скорости доступа к уровням памяти и применять разработанные методы к разным сочетаниям быстрой и медленной памяти. Построены математическая и имитационная модели процесса работы с деком, представлены результаты численных экспериментов.