An Experimental Analysis of Counting Networks∗

Counting networks are highly distributed data structures that provide low contention solutions to distributed coordination problems. Counting networks are structured with balancers interconnected in a networked fashion. We study experimentally (with simulations and real implementation) the performance of counting networks according to three criteria: (i) The way the balancers are distributed among the processors in the system, where we show that in most cases the best performance is to assign a connecting chain of balancers to each system node; (ii) The various sizes of counting networks, where we examine the tradeoffs between size and number of nodes in the system; (iii) The different kinds of counting networks, where we compare the bitonic, periodic, and BM counting networks where we find that for high loads the BM performs better than all the other networks. Key–Words: Distributed data structures, Counting networks, Load balancing, Bitonic, Periodic.