29
Interconnection Networks in Network-on-Chip
throughput, but their energy consumption and silicon area overheads are
much higher than both mesh and BFT networks. Folded torus shows almost
similar results as mesh does. In deep submicron era where low-power design
is a major goal, for a NoC designer, it is always preferable to choose a topology with lower average energy profile. Thus, although the performance of
mesh network is comparatively inferior to that of SPIN- and Octagon-based
networks, it is widely used in the industry (Vangal et al. 2008; Wentzlaff
et al. 2007).
2.3 Switching Techniques
Switching techniques determine when and how internal switches of the
network are set to connect router inputs to outputs for transferring the messages. They can be classified as circuit switching and packet switching. In
circuit switching, a physical path from source to destination is reserved prior
to the transmission of the data. The base latency of a circuit-switched message is determined by the time to set up a path and the time to transmit the
data. Banerjee et al. (2007) designed a circuit switching-based NoC router.
However, this switching technique is inefficient as it produces excessive
blocking, which in turn affects the network bandwidth and also leads to
excessive communication latency.
In the packet switching technique, each message is partitioned into fixed
length packets and the packets are transmitted without reserving the entire
path. Packet-switched networks can further be classified as store-and-forward
(SAF), virtual cut-through (VCT), and wormhole. In SAF switching, a packet is
completely buffered at each immediate node before it is forwarded to the
next node. Therefore, it needs huge silicon area. The latency in communication depends on the size of the packet. CLICHÉ is an example of SAF switching (Kumar et al. 2002). In VCT switching, a packet is forwarded to the next
router as soon as there is enough space to store the packet. VCT switching
overcomes the latency penalty of SAF switching but also requires huge silicon area to store the entire packet. In both SAF and VCT switching, message
flow control is performed at packet level. In Proteo network, VCT switching
technique has been adopted (Tortosa et al. 2004).
In wormhole switching, packets are divided into flow control units (flits)
such as header flit, payload flit, and tailer flit. The header flit contains information about source and destination addresses. The payload flit consists
of data, whereas the tailer flit contains the end of packet information. The
buffers are expected to store only a few flits. As a result, the buffer space
requirement in the switches can be small, compared to SAF and VCT switching. Header flit decoding enables the switches to establish the path, whereas
payload and tailer flits simply follow this path in a pipelined fashion. If a
Interconnection Networks in Network-on-Chip
throughput, but their energy consumption and silicon area overheads are
much higher than both mesh and BFT networks. Folded torus shows almost
similar results as mesh does. In deep submicron era where low-power design
is a major goal, for a NoC designer, it is always preferable to choose a topology with lower average energy profile. Thus, although the performance of
mesh network is comparatively inferior to that of SPIN- and Octagon-based
networks, it is widely used in the industry (Vangal et al. 2008; Wentzlaff
et al. 2007).
2.3 Switching Techniques
Switching techniques determine when and how internal switches of the
network are set to connect router inputs to outputs for transferring the messages. They can be classified as circuit switching and packet switching. In
circuit switching, a physical path from source to destination is reserved prior
to the transmission of the data. The base latency of a circuit-switched message is determined by the time to set up a path and the time to transmit the
data. Banerjee et al. (2007) designed a circuit switching-based NoC router.
However, this switching technique is inefficient as it produces excessive
blocking, which in turn affects the network bandwidth and also leads to
excessive communication latency.
In the packet switching technique, each message is partitioned into fixed
length packets and the packets are transmitted without reserving the entire
path. Packet-switched networks can further be classified as store-and-forward
(SAF), virtual cut-through (VCT), and wormhole. In SAF switching, a packet is
completely buffered at each immediate node before it is forwarded to the
next node. Therefore, it needs huge silicon area. The latency in communication depends on the size of the packet. CLICHÉ is an example of SAF switching (Kumar et al. 2002). In VCT switching, a packet is forwarded to the next
router as soon as there is enough space to store the packet. VCT switching
overcomes the latency penalty of SAF switching but also requires huge silicon area to store the entire packet. In both SAF and VCT switching, message
flow control is performed at packet level. In Proteo network, VCT switching
technique has been adopted (Tortosa et al. 2004).
In wormhole switching, packets are divided into flow control units (flits)
such as header flit, payload flit, and tailer flit. The header flit contains information about source and destination addresses. The payload flit consists
of data, whereas the tailer flit contains the end of packet information. The
buffers are expected to store only a few flits. As a result, the buffer space
requirement in the switches can be small, compared to SAF and VCT switching. Header flit decoding enables the switches to establish the path, whereas
payload and tailer flits simply follow this path in a pipelined fashion. If a