Uncontrolled Interthread Interference in Main Memory Can Destroy Individ- Ual Threads’ Memory-level Parallelism, Effectively Serializing the Memory Requests of a Thread Whose Latencies Would Otherwise Have Largely

......The main memory (dynamic RAM) system is a major limiter of computer system performance. In modern processors, which are overwhelmingly multicore (or multithreaded), the concurrently executing threads share the DRAM system, and different threads running on different cores can delay each other through resource contention. One thread’s memory requests can cause DRAM bank conflicts, row-buffer conflicts, and data/address bus conflicts with another’s. As the number of on-chip cores increases, the pressure on the DRAM system increases, as does the interference among threads sharing the system. Unfortunately, many conventional DRAM controllers are unaware of this interthread interference. They schedule requests simply to maximize DRAM data throughput. For example, the commonly used row-hit-first (FR-FCFS, or first ready, first come, first served) scheduling policy is thread unaware. Uncontrolled interthread interference in DRAM scheduling results in two major problems. First, as previous work showed, a state-of-the-art DRAM controller can unfairly prioritize some threads while starving more important threads for long time periods, as they wait to access memory (see the ‘‘Related Work on Memory Controllers’’ sidebar). For example, FR-FCFS unfairly prioritizes threads with high row-buffer hit rates over those witho low row-buffer hit rates. Similarly, an oldest-first scheduling policy implicitly prioritizes memoryintensive threads over memory-nonintensive ones. In fact, it is possible to write programs to deny DRAM service to more important programs running on the same chip, as we showed in our previous work. Such mmi2009010022.3d 31/1/09 16:11 Page 22