lower-power ones, as in Sun’s Niagara. So in terms of providing the optimal performance/power, an SMP is ideally suited for some applications but not all. Having cores of different types in a single processor enables optimizing performance per watt for a wider range of workloads. Having cores of different types on an AMP enables us to employ specialization (that is, we can use each type of core for the type of computation where it delivers the best performance/ energy trade-off). Specialization enables maximizing the overall efficiency of the system and as a result delivers better performance per watt and per area.

Although single ISA AMP systems are not yet being built, much of the lit-

erature investigating the properties of these systems has originated from major hardware players such as Intel1, 7 and HP, 5, 6 indicating that within the industry there is interest in this architecture. Furthermore, existing SMP systems can be configured to be asymmetric if one wants to reduce the amount of power consumed by the processor. For example, configuring some of the cores of an SMP to run at a lower-than-maximum voltage and frequency (via the dynamic voltage and frequency scaling facilities available on most modern CPUs) makes the CPU consume less power and makes the system asymmetric. In that case it is crucial to understand how to get the maximum performance

on this asymmetric system in order to minimize performance losses associated with running some cores at a lower Hertz.

Given the emergence of AMPs, how can you best exploit them to maximize power efficiency (performance per watt)? Our goal is to shed light on some of the challenges software developers will likely face while trying to achieve this, and to provide some practical advice on how to maximize exploitation of AMPs as a result. To that end, we provide several examples demonstrating how to employ specialization on AMP systems in order to maximize performance. One area that will be of particular importance is the design of scheduling algo-