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 default and only exposing work when requested thieves. This is the first paper obtains bounds on are (essentially) independent of total amount work, thus corresponding to great improvement, in both design theory, over state-of-the-art algorithms. Consider any computation with $$T_{1}$$ critical-path length $$T_{\infty }$$ executed P processors using our scheduler. Our analysis shows expected execution time $$O\left( \frac{T_{1}}{P} + T_{\infty }\right) $$ , incurred during at \left( C_{\mathrm{CAS}} C_{\mathrm{MFence}}\right) PT_{\infty where $$C_{\mathrm{CAS}}$$ $$C_{\mathrm{MFence}}$$ respectively, denote maximum cost executing Compare-And-Swap instruction Memory Fence instruction.