Abstract

Modern cluster processors have been steadily increasing the number of cores able to execute concurrent threads. Web search engines critically rely on multithreading to efficiently process user queries and document insertions to support real-time search. This requires synchronization of readers and writers which, for large number of threads, poses the question of what concurrency control strategies are capable of scaling to hundreds of cores and more. This paper presents a comparative study of a number of such strategies. To this end, we focus on the development of suitable simulation models for performance evaluation of search algorithms on dedicated single-purpose multi-threaded processors. We validate our model against actual implementations of the multi-threading strategies to then go further on studying performance on very large processors. We conclude that intra-query parallelism scales up more efficiently than inter-query parallelism.