Quick-and-dirty memory access tracing with instruction-based sampling

With the increasing deployment of heterogeneous memory architectures, the efficient execution of a workload becomes more sensitive to fine-grained memory placement decisions. To establish a sound information base for such decisions, we must first understand memory access behavior beyond the level of coarse-grained statistics. However, collecting detailed memory traces is a costly process. Therefore, we propose a low-overhead solution based on instruction-based sampling that provides incomplete yet informative access sequences. We showcase the practical value of such sparse traces by analyzing the overhead and comparing workload runs on two memory technologies with distinct characteristics