Synergistic Computing on Heterogeneous Multiprocessors

Abstract:

With the emerging demand for computations on mobile devices, heterogeneous multi-processors are dominating the mobile computing landscape. The heterogeneity present on the SoC enables delicate matching of computational kernels on to the processors that are best suited to perform the computation. While architectural heterogeneity is promising, software development efforts are required to fully benefit from this architectural advancement.

To start with, engaging accelerators into the execution requires redevelopment of the application in accelerator-specific languages, costing substantial time and effort. Early stage cross-platform power-performance prediction therefore facilitates the developer in choosing the accelerator for their specific application requirements before the timely redevelopment and optimization. For the acceleration through GPU, CGPredict is proposed to predict the performance of a computational kernel on an embedded GPU architecture from un-optimized, single threaded high-level C code.

Furthermore, the ever-increasing processing requirements impose higher pressure on the mobile devices, which possess limited processing capability due to the architectural limits in the embedded domain. Executing an application on a single component may not sustain the application requirements, while other components that can potentially be used are not utilized. We show that by careful orchestration of multiple components, performance and energy benefits can be obtained. On the other hand, the inherent use-cases of heterogeneous processors demand stringent power and thermal limits as compared to server systems, especially because of the lack of active cooling. OPTiC is proposed to anticipate the thermal behaviour of the heterogeneous multi-processors at compile time and predict the configuration of workload splitting across heterogeneous cores and voltage-frequency settings to achieve optimal performance.

Lastly, mobile machine learning applications call for evermore powerful on-chip computing resources. Understanding the reason behind the different performance and power behaviours of different components is important not only in choosing the best matching component for execution, but also further optimizations. Synergistic execution demonstrates opportunity to match the workload with resources present in different components. Pipe-it is proposed for ARM big.LITTLE CPU multi-core architecture, which employs a pipelined design to split the convolutional layers across different components to achieve higher throughput performance.

The goal of this dissertation is to embrace the heterogeneity by synergistic computing on multiple components to unleash the full potential of heterogeneous multiprocessors towards high-performance energy-efficient mobile computing.