Exploring coordinated software and hardware support for hardware resource allocation

Multithreaded processors are now common in the industry as they offer high performance at a low cost. Traditionally, in such processors, the assignation of hardware resources between the multiple threads is done implicitly, by the hardware policies. However, a new class of multithreaded hardware allows the explicit allocation of resources to be controlled or biased by the software. Currently, there is little or no coordination between the allocation of resources done by the hardware and the prioritization of tasks done by the software. This thesis targets to narrow the gap between the software and the hardware, with respect to the hardware resource allocation, by proposing a new explicit resource allocation hardware mechanism and novel schedulers that use the currently available hardware resource allocation mechanisms. It approaches the problem in two different types of computing systems: on the high performance computing domain, we characterize the first processor to present a mechanism that allows the software to bias the allocation hardware resources, the IBM POWER5. In addition, we propose the use of hardware resource allocation as a way to balance high performance computing applications. Finally, we propose two new scheduling mechanisms that are able to transparently and successfully balance applications in real systems using the hardware resource allocation. On the soft real-time domain, we propose a hardware extension to the existing explicit resource allocation hardware and, in addition, two software schedulers that use the explicit allocation hardware to improve the schedulability of tasks in a soft real-time system. In this thesis, we demonstrate that system performance improves by making the software aware of the mechanisms to control the amount of resources given to each running thread. In particular, for the high performance computing domain, we show that it is possible to decrease the execution time of MPI applications biasing the hardware resource assignation between threads. In addition, we show that it is possible to decrease the number of missed deadlines when scheduling tasks in a soft real-time SMT system.