Smoothing non-uniform communication latencies for OLTP.
Danica Porobic · Conference on Innovative Data Systems Research · 2015
Transaction processing applications traditionally run on the high-end servers. Up until recently, such servers had uniform core-to-core communication latencies. Now with multisocket multicores, for the first time we have Islands, i.e., groups of cores that communicate very fast with cores that belong to the same group and several times slower with cores from other groups. In current mainstream servers, each chip is an Island; as the number of cores on a chip increases, however, soon we will identify Islands within a single chip. How do transaction processing systems perform on these non-uniform platforms? Aren’t multisocket multicores just another form of abundant parallelism? We already have transaction processing systems that scale up on multicores and scale out across machines. Shouldn’t one of the existing designs be good enough for hardware Islands?