PUMA: Pseudo unified memory architecture for single-ISA heterogeneous multi-core systems

Single-ISA heterogeneous multi-core processors have advantages over cost-equivalent homogeneous ones, which integrate cores having the same instruction set architecture (ISA) but offer different performance and power characteristics. When these cores share the off-chip main memory, requests from different cores will interfere with each other, leading to low system performance and unfairness even starvation. Unfortunately, state-of-the-art memory scheduling and thread scheduling algorithms are ineffective at solving these problems. This paper proposes a fundamentally new memory architecture of pseudo unified memory (PUMA), which partitions the memory into regions according cores' different performance, each core mostly requests only one memory region seldom exceeding, reducing interfere among cores while retaining bank level parallelism for improving performance and fairness. We evaluate the design trade-offs involved in our PUMA and compare it against three state-of-the-art memory management methods. Our experimental results show that PUMA improves both system performance and fairness among cores while reducing memory power.