Research And Implementation Of The Key Technology Of Flow Control For Network-on-Chip

Multi-core and Many-core technology are widely used in high performance processors. With the increase of cores in a chip, the processor from the computation-intensive transfers to communication-intensive. The increasing communication pressure becomes a key limiting factor to improve processor performance. Network-on-Chip(No C) with the high scalability and high band width, has become the effective solution for communication bottleneck of multi-core processors. However, limited by power and area consumption,the hardware resource in No C is precious. This requests more efficient flow control to manage and allocate hardware resource. For exploring high performance flow control mechanisms, we carry out the following researches.(1)A comprehensive comparison between virtual cut-through and wormhole flow control.Virtual cut-through(VCT) is an efficient flow control mechanism and is widely used in the off-chip network. It consumes a larger number of buffers and is not suitable for No C. However, with the transistor feature size decreases, the buffers become cheaper and VCT maybe a promising flow control in the future. This paper performs a comprehensive comparison between the VCT and wormhole switching. Based on detailed RTL-level implementations, we evaluate the hardware costs with both deterministic and adaptive routing. Compared with the wormhole switching, the VCT one shortens the critical path by up to 27%, and induces less area and power overheads. Furthermore, the allocator in VCT routers exhibits a better scalability in area overheads. These factors make it possible for VCT to apply on high-dimensional, large-scale No Cs.(2)A various packet size virtual cut-through flow mechanism.The traditional VCT flow control requests all packets have the same number of flits,but there are a lot of short packets. These packets have to contain more or less invalid flits to keep up with the same number of flits in large packets. But, the invalid flits not only consume larger number of buffers, but also occupy channel resource and block other packets. In order to solve this problem, we have proposed a various packet size virtual cutthrough flow mechanism. By improving the arbiter to keep grant single in various clock circles, we can guarantee that the transmission of packet cannot be interrupted. Adding the packet size information to credits, we can compute available buffers in downstream routers. In this way, we can guarantee that downstream routers have enough buffers to store coming packets. The result shows that our designs have on influence on the critical path. With little area and power increase, the performance of the network is improved.(2) Adaptive remaining hop count flow control.The packets with different remaining hop counts transfer in the network, compete with each other and affect the throughput and globe fairness of No Cs. In general, packets with small hop counts(PSR) need less time to transfer in the network and will occupy smaller buffer resource. But, due to the competition between packets, PSR may stay longer in the network and continuously occupy the buffers. If we prefer to transfer packets with smaller remaining hop counts(PPSR), the total flying packets can be reduced and network performance is improved. Yet, the global fairness is negatively affected. Packets with large hop counts need more hop counts to transfer in the network and consume large channel resource. But, due to the competition between PSR and PLR, PLR may stay longer in the current router and cannot use the rest of channels in the routing path. If we prefer to transfer packets with larger remaining hop counts, we can achieve higher globe fairness, but the network performance is lower. So, we proposed adaptive remaining hop count flow control. If the network load is low, we use PPLR to improve globe fairness and resource utilization. If the network load is high, we sue PPSR to enhance network performance. In this way, we can improve the network performance and globe fairness.In conclusion, we focus on flow control mechanisms: exploring the VCT design space, improving traditional VCT, and proposed a novel adaptive remaining hop count flow control. This paper not only has high engineering practical value, but also is meaningful for theoretical study.