The Performance of Applications and Operating Systems on Simple IRAM Architectures

Computer memory systems are increasingly a bottleneck limiting the performance of applications and operating systems. IRAM architectures, which integrate a CPU with DRAM main memory on a single chip, have the potential to liberate systems from this bottleneck. In order to better understand the performance trade-offs of a variety of simple IRAM architectures we used the SimOS simulation environment to evaluate seven possible IRAM configurations and one conventional machine configuration. For each configuration, we analyzed the behavior of several standard integer and floating-point benchmark applications as well as memory and operating system microbenchmarks. Our simulation results for these benchmarks indicate that simple IRAM systems of 1998 ought to be as much as 40% faster for memory-intensive applications and nearly twice as fast for memory-intensive operating system services compared to a hypothetical conventional machine of 1998