Electronic systems, such as microprocessors, are composed of different units sharing common resources. When designing such systems, the system architect must decide how to distribute the available limited resources, such as area and power, among all the units.

In this dissertation, we introduce an analytical optimization framework that addresses this issue of resource sharing among the different system components. Our technique relies on performance modeling of each unit as a function of the resources allocated to it, and providing the optimal allocation of resources given the workload and total resources available. We name this framework MultiAmdahl.

We demonstrate the use of MultiAmdahl for three case studies. First, we present an optimal solution to the previously-presented problem of resource allocation inside an asymmetric multicore. Later, we present the distribution of area among a CPU and a number of identical accelerators. Finally, using real-life benchmark performance values, we model distribution of area or power in a heterogeneous chip that contains specific and programmable accelerators.

Using MultiAmdahl, we show that the serial or least scalable unit is more important in a system where resource budget (e.g. available power) is high, and less important on systems with a tight budget. We also discuss “efficiency” and “flexibility” of hardware units, and show the two properties are not independent of each other.