Towards Dynamic Cache and Bandwidth Invasion
State-of-the-art optimizations for high performance are frequently related to particular hardware parameters and features. This typically leads to optimized software for execution on particular hardware configurations. However, so far, the applications lack the ability to modify hardware parameters either statically before execution of a program or dynamically during run-time.
In this paper, we first propose to utilize the flexibility of underlying invasive hardware to adapt to the needs of the software. This enables us to ask for more than just processing power by, e.g., requesting particular cache parameters that correspond to certain application properties. The adaptive hardware architecture therefore is able to dynamically reconfigure itself dependent on the availability of the resources in order to achieve an optimized working point for each application scenario. Secondly, we present requirements for dynamical scheduling of computing resources to resource-competing applications. This becomes mandatory to account for memory-access characteristics of concurrently executed applications. We propose consideration of such characteristics with bandwidth-aware invasion.
With this novel approach, we are able to show that dynamic hardware and software co-design leads to improved utilization of the underlying hardware resulting in higher throughput in means of efficiency such as application-throughput per time-unit.
Keywordsinvasive computing adaptive application-specific microarchitecture reconfigurable cache compute-bound memory-bound HPC
Unable to display preview. Download preview PDF.