CLAM: Compiler Lease of Cache Memory

Caching is a common solution to the data movement performance bottleneck of today’s computational systems and networks. Traditional caching examines program behavior and cache optimization separately, limiting performance. Recently, a new cache policy called Compiler Lease of cAche Memory (CLAM), has been suggested for program-based cache management. CLAM manages cache memory by allowing the compiler to assign leases, or lifespans, to cached items over a hardware-software interface, known as lease cache. Lease cache affords new performance potential, by way of program-driven cache optimization. It is applicable to existing cache architecture optimizations, and can be used to emulate other cache policies. This paper presents the first functional hardware implementation of lease cache for CLAM support. Lease cache hardware architecture is first presented, along with CLAM hardware support systems. The cache is emulated on an FPGA, and benchmarked using a collection of scientific kernels from the PolyBench/C suite, for three CLAM lease assignment policies: Compiler Assigned Reference Leasing (CARL), Phased Reference Leasing (PRL), and Fixed Uniform Leasing (FUL). CARL and PRL are able to achieve superior performance to Least Recently Used (LRU) replacement, while FUL is shown to serve as a safety mechanism for CLAM. Novel spectrum-based cache tenancy analysis verifies PRL’s effectiveness in limiting cache utilization, and can identify changes in the working-set that cause the policy to perform adversely. This suggests that CLAM is extendable to more complex workloads if working-set transitions can elicit a similar change in lease policy. Being able to do so could yield appreciable performance improvements for large and highly iterative workloads like tensors