Using Tracing To Enhance Data Cache Performance in CPUs: The creation of a Trace-Assisted Cache to increase cache hits and decrease runtime

The processor-memory gap is widening every year with no prospect of reprieve. More and more latency is being added to program runtimes as memory cannot satisfy the demands of CPUs quickly enough. In the past, this has been alleviated through caches of increasing complexity or techniques like prefetching, to give the illusion of faster memory. However, these techniques have drawbacks because they are reactive or rely on incomplete information. In general, this leads to large amounts of latency in programs due to processor stalls. It is our contention that through tracing a program's data accesses and feeding this information back to the cache, overall program runtime can be reduced. This is achieved through a new piece of hardware called a Trace-Assisted Cache (TAC). This uses traces to gain foreknowledge of the memory requests the processor is likely to make, allowing them to be actioned before the processor requests the data, overlapping memory and computation instructions. Comparing the TAC against a standard CPU without a cache, we see improvements in runtimes of up to 65%. However, we see degraded performance of around 8% on average when compared to Set-Associative and Direct-Mapped caches. This is because improvements are swamped by high overheads and synchronisation times between components. We also see that benchmarks that exhibit several qualities: a balance of computation and memory instructions and keeping data well spread out in memory fare better using TAC than other benchmarks on the same hardware. Overall this demonstrates that whilst there is potential to reduce runtime via increasing the agency of the cache through Trace Assistance, it requires a highly efficient implementation to be competitive otherwise any potential gains are negated by the increase in overheads.