Fence-free work stealing on bounded TSO processors

Abstract

Work stealing is the method of choice for load balancing in task parallel programming languages and frameworks. Yet despite considerable effort invested in optimizing work stealing task queues, existing algorithms issue a costly memory fence when removing a task, and these fences are believed to be necessary for correctness. This paper refutes this belief, demonstrating work stealing algorithms in which a worker does not issue a memory fence for microarchitectures with a bounded total store ordering (TSO) memory model. Bounded TSO is a novel restriction of TSO -capturing mainstream x86 and SPARC TSO processors -that bounds the number of stores a load can be reordered with.Our algorithms eliminate the memory fence penalty, improving the running time of a suite of parallel benchmarks on modern x86 multicore processors by 7% − 11% on average (and up to 23%), compared to the Cilk and Chase-Lev work stealing queues.

Bandwidth and Locality Aware Task-stealing for Manycore Architectures with Bandwidth-Asymmetric Memory

“…Targeting scheduling systems for task-based programs, a large amount of prior work aims to improve energy-efficiency [38,41], to improve data locality [9,10], or to reduce scheduling overhead [17,29]. However, with the increasing bandwidth requirements of computing tasks, many papers have also conducted related research for efficient bandwidth usage.…”

Temporally Bounding TSO for Fence-Free Asymmetric Synchronization

“…We further adapt the echo method [29] to make non-owner lock acquisition speed comparable to standard locks, assuming the owner acquires the lock frequently. Because our lock does not rely on blocking safe points, it (1) can be used in C/C++ programs, which do not naturally define safe points, and (2) enables non-owner acquisition even if the owner is scheduled out or delayed.…”

“…This signals to T 1 that T 0 is waiting to acquire L, so T 1 can stop the ∆ delay and enter the critical section. To implement this notification we use echoing [29]: We expand the flags to 64-bits, 63 of which are used as version numbers that uniquely identify each writewhenever T 1 writes to flag 1 , it increases flag 1 's version. T 0 uses this version to notify T 1 that it is spinning while trying to acquire L, by writing-or echoing-what it reads from flag 1 into flag 0 (Lines 59-63).…”

“…Thus, in hazard pointers for example, we only add work to the reclamation path and not to every object retirement. (In contrast, the bounded TSO[S] model [29], which does not have a global clock, would require reading A when retiring an object. )…”

“…While all biased locks collapse compared to pthreads in this case, we see the benefit of bounded delay: the FFBL outperform the safe point lock by 50× (with echoes) and by 7× (without echoes). [29]. In contrast, TBTSO's temporal reordering bound facilitates nonblocking synchronization without relaxing semantics, making it more broadly applicable.…”