Algorithms for scalable synchronization on shared-memory multiprocessors

Reviewer: Roy Levin

The authors have compiled a competent survey of spin-lock and barrier synchronization algorithms, including some new ones of their own. Two sections of the paper present these algorithms with clear pseudocode. Performance comparisons on a BBN Butterfly and a Sequent Symmetry support the conclusions in the subsequent section on architectural implications. The paper closes with a concise list of recommendations that will aid software practitioners who must navigate this sometimes bewildering space of possible algorithms. I was distracted by several small inconsistencies in notation and presentation that should have been picked up in the editorial process (like the references to the nonexistent Algorithms 8 and 11). It also took me a while to realize that a direct correspondence does not always exist between the algorithms whose code is presented and the algorithms whose performance is graphed. Finally, the space devoted to a lame attempt to summarize the correctness proof of the authors' intriguing new spin-lock algorithm would have been better spent explaining how a key race-condition is resolved. These inconsistencies are relatively small flaws in an otherwise well-organized and well-reasoned paper. I encourage designers of parallel architectures and parallel algorithms to read it.