Adaptive and Deadlock-Free Tree-Based Multicast Routing for Networks-on-Chip

Abstract

Abstract-This paper presents the first synthesizable network-on-chip (NoC) based on a mesh topology, which supports adaptive and deadlock-free tree-based multicast routing without virtual channels. The deadlock-free routing algorithms for unicast and multicast packets are the same. Therefore, the routing function gate-level implementation is very efficient. Multicast packets are injected to the network by sending multiple packet headers beforehand. The packet headers contain destination addresses to set up multicast trees connecting a source with multiple destination nodes. An additional locally uniform identification (ID) field is packetized together with flits belonging to the same packet. Therefore, flits of different unicast or multicast packets can be interleaved in the same queue because of the local ID-tags, which are updated and mapped dynamically to support bandwidth scalability of interconnection links. Deadlocks in tree-based multicast routing are handled using a flit-by-flit round arbitration and a fair hold-release tagging mechanism. The effectiveness of the novel mechanism has been experimented under multiple multicast conflicts scenarios, where the experimental results show that all traffic is accepted in-order and lossless in their destination nodes even if adaptive routing functions are used and the sizes of the multicast messages are very long.Index Terms-Network-on-chip (NoC), tree-based multicast routing, synchronous parallel pipeline, wormhole packet switching.

New Theory for Deadlock-Free Multicast Routing in Wormhole-Switched Virtual-Channelless Networks-on-Chip

“…The size of ID-tags in the RRT can be set less than the number of node entries in the NoC to run specific data distribution scenarios. This case has been experimented in [25] and [26] by setting 16 ID-tags per link in the 80 cores (mesh 8 Â 10) NoC.…”

“…The "wormhole cut-through switching" is implementable by attaching a local ID-tag on each flit of packets, where flits belonging to the same packets will have the same local ID-tag on every local communication channel [24], [25], [26]. Furthermore, an arbitration unit distributed at every output port m must rotate (circulate) its selection among requests at instant time t s .…”

“…This paper focuses on the proof of the new theory for deadlock-free multicast routing in a wormhole-switched virtual channelless NoC. The verification through RTLsimulation as well as the early VLSI architecture implementations using CMOS standard-cell technology can be found in [24] in 2D 4 Â 4 Mesh NoC with static XY routing algorithm, in [25] in 2D 8 Â 10 Mesh NoC with static and partial adaptive routing algorithms, and in [26] in 2D 8 Â 10 Mesh-Planar NoC with 2D planar adaptive routing algorithm.…”

A Multicast Tree Router for Multichip Neuromorphic Systems

“…6 Deadlock has been dealt with by interleaving words from up to n packets on the same link (for n nodes) [30], extending the approach used in the unicast case, where two packets are interleaved using virtual channels on the same physical link [31]. In both cases, the destination must either provide enough buffering to locally reconstruct entire packets or employ non-blocking multiport memory structures to do so (see Discussion in [30]). These options are not viable for our application.…”

Design and implementation of a NoC router supporting multicast

“…However, it has to maintain the big virtual circuit tree table and needs to send a setup packet to build a tree before data transmission, which involves large storage resources overhead and multicast latency. Combined with wormhole switching, there are many other tree-based multicast schemes [5,6]. But as far as we know, none of them had completely solved the complex multicast problem coming with this mechanism.…”