Asymmetric Chip Multiprocessors: Balancing Hardware Efficiency and Programmer Efficiency

Chip Multiprocessors are becoming common as the cost of increasing chip power begins to limit single core performance. The most power efficient CMP consists of low power in-order cores. However, performance on such a processor is low unless the workload is nearly completely parallelized, which depending on the workload can be impossible or require significant programmer effort. This paper argues that the programmer effort required to parallelize an application can be reduced if the underlying architecture promises faster execution of the serial portion of an application. In such a case, programmers can parallelize only the easier-to-parallelize portions of the application and rely on the hardware to run the serial portion faster. We make a case for an architecture which contains one high performance out-of-order processor and multiple low performance in-order processors. We call it an Asymmetric Chip Multiprocessor (ACMP). Although the out-of-order core in the ACMP makes it less power efficient, it enables the ACMP to produce higher performance gains with less programmer effort