On stalling in LogP
We investigate the issue of stalling in the LogP model. In particular, we introduce a novel quantitative characterization of stalling, referred to as δ-stalling, which intuitively captures the realistic assumption that once the network’s capacity constraint is violated, it takes some time (at most δ) for this information to propagate to the processors involved. We prove a lower bound that shows that LogP under δ-stalling is strictly more powerful than the stall-free version of the model where only strictly stall-free computations are permitted. On the other hand, we show that δ-stalling LogP with δ = L can be simulated with at most logarithmic slowdown by a BSP machine with similar bandwidth and latency values, thus extending the equivalence (up to logarithmic factors) between stall-free LogP and BSP argued in  to the more powerful L-stalling LogP.
KeywordsCapacity Constraint Local Memory Incoming Message Barrier Synchronization Maximum Delivery Time
Unable to display preview. Download preview PDF.
- 3.M.T. Goodrich. Communication-Efficient Parallel Sorting. In Proc. of the 28th ACM Symp on Theory of Computing, pages 247–256, Philadelphia PA, 1996.Google Scholar
- 5.V. Ramachandran, B. Grayson, and M. Dahlin. Emulations between QSM, BSP and LogP: a framework for general-purpose parallel algorithm design. TR98-22, Dept. of CS, Univ. of Texas at Austin, November 1998. (Summary in Proc. of ACM-SIAM SODA, 1999.)Google Scholar