Categorizing Network Tra c in Update-Based Protocols on Scalable Multiprocessors

In this paper we categorize the coherence traac in update-based protocols and show that, for most applications , more than 90% of all updates generated by the protocol are unnecessary. We identify application characteristics that generate useless update traac, and compare the isolated and combined eeects of several software and hardware techniques for eliminating useless updates. These techniques include dynamic and static hybrid protocols, a data re-mapping strategy, and coalescing write buuers. Our simulations show that these techniques are eeective for diierent types of useless updates. Overall, software caching (where dynamic data re-mapping is performed under programmer or compiler control) has the potential to signii-cantly increase the percentage of useful traac in applications. When software caching is not applicable, either the static or the dynamic protocol generates the least useless traac. Although coalescing write buuers provide great reductions in the total number of messages transferred, these buuers do not necessarily increase the percentage of useful traac.