Background. Multicore applications share cache. Composable analysis is needed to see how programs interact with dynamic, usage based cache management. Miss ratio/rate doesn’t compose.
Xiaoya’s work. Footprint composes but assumes uniform interleaving.
Jake et al. Common logical time in CCGrid 2015 handles non-equal length component traces, eg. one thread accesses memory 100 times more frequent than another thread. But we still assume uniform interleaving.
Hao repots several advances.
Now we can compose threads that have any type of interleaving.
Cache associativity modeling
The Smith method is the dominant solution for nearly 40 years but assumes equal probability of access in all cache sets. Hao’s new solution removes this assumption and uses time-preserving decomposition to also allow non uniform interleaving.
GPU texture cache
Modeled as a sector cache to give composable performance for all cache sizes and associativity as for normal cache.
Space efficient algorithms for shared footprint analysis.
Possibly memcached or io traces.
Static locality analysis.
Locality aware scheduling