Abstract
dc:descriptionOver the last two decades, we have witnessed a fundamental paradigm shift in digital system design: the transition to the multi-core realm. Naturally, the multi-core domain has elevated the criticality of the on-chip interconnection fabric, which is now tasked with satisfying amplified communication demands. Owing to their scalability attributes, Networks-on-Chip (NoC) have established their position as the de facto communication medium in multi-core systems. To sustain system scalability into the many-core domain (with potentially hundreds of cores), it is imperative that the NoC's hardware cost is minimized, while not sacrificing network performance.To this end, we propose three alternative architectures that can significantly improve the performance of NoCs, or lead to an overall lower power consumption.The first one is a pipelined router architecture, called ShortPath, that parallelizes - for the first time - the allocation steps involved in the operation of a VC-based router without resorting to speculation. Most importantly, ShortPath is augmented with an always-productive pipeline bypassing mechanism, which skips all stages without contention, and "fast-forwards" the flits to the first encountered point of contention.The other two approaches exploit fast link traversal, after appropriate wire engineering, to rapidly transfer flits between adjacent routers (connected with links of reasonable, short-to-medium length of up to a few millimeters) in half a clock cycle. Under this clocking principle, two design alternatives are explored, which allow for half-cycle and Double-Data-Rate (DDR) link traversal. The proposed approaches can markedly increase network performance, or decrease the area/power cost of the NoC. Although not obvious at first glance, half-cycle link traversal opens up new possibilities for reducing wire capacitance. By harnessing these opportunities, the half-cycle-delay requirement becomes easier to achieve and potentially extends half-cycle traversal capabilities to longer links. To tackle longer links, novel DDR dual-stream elastic buffers are proposed for pipelining the links while still following DDR flow control.As multi/many-core architectures evolve, the demands on the NoC are amplified. In addition to high performance and physical scalability, the NoC is increasingly required to also provide specialized functionality, such as network virtualization, flow isolation, and quality-of-service guarantees. Although traditional architectures supporting Virtual Channels (VCs) offer the resources for flow partitioning and isolation, an adversarial workload can still interfere and degrade the performance of other workloads that are active in a different set of VCs. Motivated by this aspect, we present PhaseNoC, a truly non interfering VC-based architecture that adopts time-division multiplexing at the VC level. Distinct flows, or application domains, mapped to disjoint sets of VCs are isolated, both inside the router's pipeline, and at the network level. Any latency overhead is minimized by appropriate scheduling of flows in separate phases of operation, irrespective of the chosen topology. When strict isolation is not required, the proposed architecture can employ opportunistic bandwidth stealing to further reduce packet latency.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Psarras, Anastasios
- Ψαρράς, Αναστάσιος
Subjects
dc:subject × 14- Δίκτυα σε ολοκληρωμένα κυκλώματα
- Ολοκληρωμένα κυκλώματα
- Πολυπύρηνοι επεξεργαστές
- Συστήματα σε ολοκληρωμένα κυκλώματα
- Networks-on-Chip
- VLSI
- Multi-core architectures
- Chip Multi-Processors
- System-on-chip
- On-chip Interconnection Networks
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/41266