With the advancement of semiconductor technology, the System on Chip (SoC) is becoming more and more complex, so the on-chip communication has become a bottleneck of SoC Design. Since the traditional bus system is inefficient and not scalable, the Network-On-Chip (NoC) has emerged as the promising communication mechanism for complex SoCs. As some systems have specific performance requirements, such as a minimum throughput (for real-time streaming data) or bounded latency (for interrupts, process synchronization, etc), communication with Guaranteed Service (GS) support becomes crucial for predictable SoC architectures. Circuit Switching (CS) is a popular approach to support GS, which firstly has to allocate an exclusively connection (circuit) between the source and destination nodes, and then the data packets are delivered over this connection. However, it is inefficient and inflexible because the resource is occupied by single connection during its whole lifetime, which can block other communications. Hence, two extensions of CS have been proposed to share resources: i) Time-Division Multiplexing (TDM), in which the available link capacity is split into multiple time slots to be shared by different flows in TDM scheme; and ii) Space-Division-Multiplexing (SDM), in which only a subset (sub-channel) of the link wires is exclusively allocated to a specific connection, while the remaining wires of the link can be used by other flows.
The connection allocation is critical for CS, since the data delivery can start only after the associated connection is allocated. In this thesis, we propose a dedicated hardware connection allocator to solve the dynamic connection allocation problem for CS NoCs, which has to i) allocate a contention-free path between source-destination pairs and ii) allocate appropriate portions of link bandwidth (appropriate number of time slots and subsets) along the path. The dedicated connection allocator, called NoCManager, solves the connection allocation problem by employing a trellis-search based shortest path algorithm. The trellis search can explore all possible paths between source node and destination. Moreover, it shall find the requested path in a fixed low latency and can guarantee the path optimality in terms of path length if the path is available.
In this thesis, two different trellis graphs, Forward-Backtrack trellis and Register-Exchange trellis are proposed. The Forward-Backtrack trellis completes the path search in two steps: forward search and backtracking. Firstly, the forward search begins at source node that traverses the network to find the free path. When destination node is reached, the backtrack starts from destination to select the survivor path and collect the associated path parameters. However, Register-Exchange trellis saves the entire survivor path sequences during forward search. Consequently, the backtracking step can be omitted, and thus the allocation time is halved compared to forward-backtrack approaches. Moreover, each trellis graph consists of three categories, unfolded structure, folded structure and bidirectional structure. The unfolded structure can provide high allocation speed while folded structure is more efficient from a hardware point of view. The bidirectional structure starts the search at two sides, source node and destination node simultaneously, so the allocation speed is 2 times faster than previous unidirectional search. Furthermore, in order to address the scalability issue of previous centralized systems, the partitioned architecture (i.e. spatial partitioning technique) is proposed to divide the large system into multiple smaller differentiated logical partitions served by local NoCManagers. This partitioning technique keeps the request load of the manager and manager-node communication overhead moderate. Inside each partition, the path search problem is solved by a local manager with trellis-search algorithm. To establish a path that crosses partitions, the managers communicate with each other in distributed manner to converge the global path.
In order to further enhance the path diversity and resource utilization, we adopt the combined TDM and SDM technique. In combined TDM-SDM approach, each SDM sub-channel is split into multiple time slots so that can be shared by multiple flows. Hence, the number of sub-channels can be kept moderate to reduce router complexity, while still providing higher path diversity than TDM scheme. In order to investigate and optimize TDM-SDM partitioning strategy, we studied the influence of different TDM-SDM link partitioning strategies on success rate and path length that allowed us to find the optimal solution. The dedicated connection allocator using the trellis-search algorithm is employed for TDM, SDM and TDM-SDM CS.
In the end, we present the router architecture that combines the circuit-switching network (for GS communication) and packet-switching network (for best-effort communication).
| Identifer | oai:union.ndltd.org:DRESDEN/oai:qucosa.de:bsz:14-qucosa-229091 |
| Date | 14 September 2017 |
| Creators | Chen, Yong |
| Contributors | Technische Universität Dresden, Fakultät Elektrotechnik und Informationstechnik, Prof. Dr. Gerhard Fettweis, Prof. Dr. Holger Blume, Prof. Dr. Ralf Lehnert, Prof. Dr. Christian Mayr |
| Publisher | Saechsische Landesbibliothek- Staats- und Universitaetsbibliothek Dresden |
| Source Sets | Hochschulschriftenserver (HSSS) der SLUB Dresden |
| Language | English |
| Detected Language | English |
| Type | doc-type:doctoralThesis |
| Format | application/pdf |
Page generated in 0.0041 seconds