Suggestions for Locality Optimizations Rochester (SloR) is a locality analysis tool that collects samples of accesses at run time and provides the programmer suggestions to improve both the spatial and temporal locality. SloR is an extension of Slo, a tool originally developed by Kristof Beyls at the University of Ghent. Slo samples individual memory access reuses and provides suggestions to move reuses closer together. With only temporal locality information, it was able to halve execution times of several SPEC2000 programs. Nonetheless, Slo is only concerned with temporal locality of data elements.
SloR preserves all the functionality of Slo and also provides a number of additional features. First, it provides feedback on so-called block sampling. Directly analyzing specific memory accesses leads to a false sense of understanding because cache blocks contain more than a single piece of data. Block sampling originates from this idea. In addition to analyzing specific memory elements, SloR also examines the temporal locality of memory blocks. SloR also provides a spatial locality ranking. It identifies the causes of poor spatial and temporal locality with the precision at least as fine as a single basic block. Finally, SloR provides feedback on the layout of fields within structures based on reference affinity. While full reuse distance would cause a slowdown by a factor in the hundreds, sampling based analysis is typically within a factor of ten of the original execution.