CARL: Compiler Assigned Reference Leasing

Data movement is a common performance bottleneck, and its chief remedy caching. Traditional cache management transparent to the workload: data that should be kept in are determined by recency information only, while program information, i.e., future reuses, not communicated cache. This has changed new design named Lease Cache . The control passed lease compiler technique called Compiler Assigned Reference (CARL). collects reuse interval distribution for each reference uses it compute assign value reference. In this article, we prove CARL optimal under certain statistical assumptions. Based on optimality, miss curve convexity, which useful optimizing shared cache, sub-partitioning monotonicity, simplifies compilation. We evaluate potential using scientific kernels from PolyBench show insertions of up 34 leases code achieve similar or better utilization (in variable size cache) than fixed-size caching policy, been unattainable with automatic but now within programming all tested programs most sizes.