Research management and funding
ESC - Efficient Stream Computing by Fitting Computations to Cores
Details
Start date: 01/09/2017
End date: 31/08/2021
Description
tream Computing has been introduced as a paradigm in Cloud Computing and Big Data to emphasise the streaming nature of modern computing applications. Such applications typically pull multiple streams of data and process these under real-time constraints. Recently heterogeneous architectures have been introduced (DSPs, GPUs) to process parts of the workload. The need for such solutions stems from the need to achieve better cost-effectiveness at the given real-time constraints. One of the main metrics of effectiveness is energy costs, which for cloud systems are becoming the major long-term cost. A recent analysis on the web-site SemiAccurate based on numbers from PayPal claims that the move from x86 to ARM/TI/DSPs has resulted in a 9x reduction in acquisition costs, 2x reduction in power consumption per year, and a 7x increase in node density per rack, while keeping the performance the same. Thus for the particular application the use of heterogeneous embedded cores would seem to be very advantageous.
In heterogeneous architectures more energy-efficient specialized compute-units are used for parts of the computations. The design problem then amounts to finding a good decomposition of the problem into tasks and then map the tasks onto the heterogenous architecture to achieve minimal energy consumption at a given performance level. How this mapping should be done, and what parameters influence it (task granularity, memory architecture, processor architecture) is still an open research question in the embedded area.
The ESC project tries to address this mapping problem by introducing the notion of Fitness. Fitness should characterise how much of the micro-architecture the task is able to exploit. A good Fitness between software and microarchitecture means that the overheads of the core are low, while for a bad Fitness the overheads are high. This intuitively corresponds to the idea that a memory bound task fits better and ARM processor than an x86 core, since the task is not able to utilize the complex microarchitecture of the x86.
The goal of the ESC project is to propose a metric for deciding how well an application fits an architecture (a Fitness metric) and a design methodology that supports the use of the metric at different levels of abstractions in the system. We believe that a successful solution of these problems will enable faster design of more energy-efficient Stream Computing Systems.

Last updated on 2017-08-07 at 15:02