Timing-Constrained Global Routing with Buffered Steiner Trees

Abstract

This dissertation deals with the combination of two key problems that arise in the physical design of computer chips: global routing and buffering.<br /> The task of buffering is the insertion of buffers and inverters into the chip's netlist to speed-up signal delays and to improve electrical properties of the chip. Insertion of buffers and inverters goes alongside with construction of Steiner trees that connect logical sources with possibly many logical sinks and have buffers and inverters as parts of these connections. Classical global routing focuses on packing Steiner trees within the limited routing space. Buffering and global routing have been solved separately in the past.<br /> In this thesis we overcome the limitations of the classical approaches by considering the buffering problem as a global, multi-objective problem. We study its theoretical aspects and propose algorithms which we implement in the tool BonnRouteBuffer for timing-constrained global routing with buffered Steiner trees. At its core, we propose a new theoretically founded framework to model timing constraints inherently within global routing. As most important sub-task we have to compute a buffered Steiner tree for a single net minimizing the sum of prices for delays, routing congestion, placement congestion, power consumption, and net length. For this sub-task we present a fully polynomial time approximation scheme to compute an almost-cheapest Steiner tree with a given routing topology and prove that an exact algorithm cannot exist unless P=NP. For topology computation we present a bicriteria approximation algorithm that bounds both the geometric length and the worst slack of the topology. To improve the practical results we present many heuristic modifications, speed-up- and post-optimization techniques for buffered Steiner trees.<br /> We conduct experiments on challenging real-world test cases provided by our cooperation partner IBM to demonstrate the quality of our tool. Our new algorithm could produce better solutions with respect to both timing and routability. After post-processing with gate sizing and Vt-assignment, we can even reduce the power consumption on most instances. Overall, our results show that our tool BonnRouteBuffer for timing-constrained global routing is superior to industrial state-of-the-art tools.