A typical problem description for independent task scheduling does not account for the operating system. This gap is a concern, as it makes the application of results difficult in real environments. Operating systems such as GNU/Linux are timesharing systems, which can hide latency (I/O) with computation. This feature should not be ignored (by assigning only one task to a core), but leveraged to produce better schedules. Moreover, the kernel power management features needs to be present in the model as it imposes a precise model for power or energy evaluation. GNU/Linux cpu-freq and the on-demand governor are now included in the scheduling problem model. The self-regulated power management, based on CPU utilization, should also not be ignored (it is typically replaced with a user-set voltage/frequency operating point for each task), as it offers a mean to manage complexity in the data center, and can increase energy savings.

This paper introduces a simple model which nevertheless accounts for these effects. The model also includes previous efforts to model for contention of shared resource, such as memory.

Moreover, the concept of software pipelining is presented as a profitable technique to decompose large applications into smaller tasks. The benefits are the possibility of porting larger applications onto millicomputing based clusters, and the improved quality of the schedules obtained due to the finer grain control over the application.

The model, together with an adapted heuristic designed for the independent task mapping, serves as a simulation tool to explore the performance and energy efficiency of millicomputing for data centers.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £85 +P&P)