An Asynchronous Time-Division-Multiplexed Network-on-Chip for Real-Time Systems

Abstract

Multi-processor architectures using networks-on-chip (NOCs) for communication are becoming the standard approach in the development of embedded systems and general purpose platforms. Typically, multi-processor platforms follow a globally asynchronous locally synchronous (GALS) timing organization. This thesis focuses on the design of Argo, a NOC targeted at hard real-time multi-processor platforms with a GALS timing organization. To support real-time communication, NOCs establish end-to-end connections and provide latency and throughput guarantees for these connections. Argo uses time division multiplexing (TDM) in combination with a static schedule to implement virtual end-to-end circuits. TDM is a straightforward way to provide guarantees and to share the resources efficiently, and it has an efficient hardware implementation. Argo supports a GALS system organization, and additionally it explores more flexible timing within its structure, to address signal distribution issues, using a network of synchronous routers. NOCs consist of a switching structure of routers connected by links, with network interfaces (NIs) that connect the processors to the switching structure. Argo uses a novel NI design that supports time-predictability, and asynchronous routers that form a time-elastic network. The NI design integrates the DMA functionality and the TDM schedule, and uses dual-ported local memories. The routers combine the router functionality and asynchronous elastic behavior. They also use a gating mechanism to reduce the energy consumption. The combination of the NI design and the router design supports the formation of end-to-end paths in the NOC, from the local memory of a sending core to the local memory of a receiving core. These end-to-end paths do not require any dynamic arbitration, buffering, flow control, or clock synchronization, in the routers or the NIs. This thesis explores the implementation of the individual components of Argo, as well as several complete instances of the Argo NOC. The implementations target both FPGA technology and 65 nm CMOS technology. It is shown that (i) the NI design is scalable and four to five times smaller than previously published NIs for similar NOCs, (ii) the router design is power efficient and two to three times smaller than equivalent router designs, and (iii) the overall Argo NOC is around four times smaller than other TDM NOCs. Argo is an important part of the T-CREST paltform and used in a number of configurations. The flexible timing organization of Argo combines asynchronous routers with mesochronous NIs, which are connected to individually clocked cores, supporting a GALS system organization. The mesochronous NIs operate at the same frequency, possibly with some skew, while the network of asynchronous routers absorbs this skew within certain limits. The elasticity of the asynchronous network is explored, answering the question of how much skew the Argo NOC can absorb. A qualitative analysis studies the parameters affecting the elasticity and its limits. A quantitative analysis models the Argo behavior using timed-graph models and worstcase timing separation of events analysis to evaluate the elasticity of Argo. The results show that the skew absorbed by the network of routers can be two or more cycles, depending on the frequency applied at its endpoints, the NIs. Overall this thesis presents the design and implementation of Argo, and the analysis of its elastic behavior. It shows that Argo provides hard real-time guarantees in a straightforward way, it has an efficient implementation and it is time-elastic.