Introduction

Model-T is a preliminary investigation into scheduling pipelined applications -- such as Streamline -- on manycore processors. Its main feature is an assembly-line style scheduling algorithm that maximizes L2 DCache hitrate by

1. breaking up applications into pipelines
2. co-scheduling adjacent filters on adjacent nodes
3. monitoring their throughput
4. adjusting runqueues to make filters have overlapping working sets

The last step increases DCache hitrate. At the same time, it identifies the slowest links in the pipeline. Iteratively increasing their timeslots or parallelizing their computation removes individual bottlenecks resulting in higher global throughput.

Further reading

check out the Model-T paper on our publications page.