MRFI Stream Arbitration: An Efficient Arbitration Scheme for NoC with Emerging Interconnect Technology

📝 Abstract

An improved version of stream arbitration based on multiband RF interconnect (MRFI) is proposed. Thanks to the simultaneous multiple channel transmitting/receiving feature of MRFI, dynamic bandwidth allocation is achieved in the proposed arbitration algorithm. With dynamic bandwidth allocation, MRFI based arbitration can guarantee 100% channel bandwidth utilization, which is a significant improvement compared with original RF-I based stream arbitration whose channel bandwidth utilization is only around 30%~50%.

💡 Analysis

An improved version of stream arbitration based on multiband RF interconnect (MRFI) is proposed. Thanks to the simultaneous multiple channel transmitting/receiving feature of MRFI, dynamic bandwidth allocation is achieved in the proposed arbitration algorithm. With dynamic bandwidth allocation, MRFI based arbitration can guarantee 100% channel bandwidth utilization, which is a significant improvement compared with original RF-I based stream arbitration whose channel bandwidth utilization is only around 30%~50%.

📄 Content

MRFI Stream Arbitration: An Efficient Arbitration Scheme for NoC with Emerging Interconnect Technology Rex Lee, Yilei Li fortelee2016@gmail.com Abstract An improved version of stream arbitration based on multiband RF interconnect (MRFI) is proposed. Thanks to the simultaneous multiple channel transmitting/receiving feature of MRFI, dynamic bandwidth allocation is achieved in the proposed arbitration algorithm. With dynamic bandwidth allocation, MRFI based arbitration can guarantee 100% channel bandwidth utilization, which is a significant improvement compared with original RF-I based stream arbitration whose channel bandwidth utilization is only around 30%~50%.

1. Introduction
   CMOS circuits are the mainstream in current integrated circuits, and its performance keeps evolving with feature size scaling [1-7]. Analog and RF CMOS circuits are usually used for communication [8-10], navigation [11-12], human-machine interface [13-15], and so on; however, RF-based circuits can also be used in processor/memory interface, which is RF-Interconnect (RFI) [16-21]. RFI can provide excellent energy efficiency, and also its RF nature provides potential for flexible arbitration scheme in network-on-chip applications. In [22], stream arbitration based on RF-I was proposed. Compared with other arbitration schemes in emerging interconnect technology, RF-I based stream arbitration has excellent flit transmission latency even with temporal and spatial communication heterogeneity. The authors of [22] also suggest that the efficiency of stream arbitration could be further improved by using dynamic bandwidth allocation. However, with RF-I node which can be tuned to only one frequency channel at one time, it would be very hard to implement dynamic bandwidth allocation.
   On the other hand, the newly proposed multi-band RF interconnect (MRFI) [17-21] node can receive/transmit with multiple frequency channels simultaneously, which enables dynamic bandwidth allocation with minimal overhead. What’s more, MRFI has better scalability with technology than original RF-I. Consequently, it is very promising to implement an improved version of stream arbitration with MRFI to achieve superior overall NoC interconnect performance. This paper is organized as follows. After introduction in section 1, detailed analysis of RF-I based stream arbitration is described in section 2. The key feature of feature of MRFI that enables dynamic bandwidth allocation is present in section 3. Based on this feature, MRFI stream arbitrationis proposed in section 4. Finally, future works are discussed in section 5.
2. Limit of RF-I based stream arbitration In RF-I based stream arbitration, each node sends out sub-stream vector and sub-stream vectors form full stream by appending to each other in time domain (trip1). Then each node obtains arbitration result and frequency channel ID from full stream (trip2). Each source-destination pair can only use one frequency channel at a time. The primary limit of RF-I is that one node can only transmit with one frequency channel at one time. This is due to carrier generation part of RF-I node, which can only generate one frequency for TX/RX at one time.
   This implementation limit greatly degrades the resource (channel bandwidth) utilization of RF-I NoC, especially when the traffic pattern is not uniform. One source-destination pair can transfer data with only one frequency channel regardless of whether other channels are idle or not. In many cases other channels are actually idle, and consequently a great portion of channel bandwidth is wasted. This waste is significant when number of active nodes is low. Consider the following case when node0 tries to transmit four flits to node1. Suppose there are four frequency channels for data transfer (each channel can transfer one flit in one cycle), and node0 is the only active node in the network. If node0 could use full bandwidth of the channel (four channels) it could finish data transfer of four flits within one cycle. However, with original RF-I based stream arbitration, node0 can only transfer one flit through one channel in one cycle (while the bandwidth of other three frequency channels are wasted). As a result, it takes four cycles for node0 to finish data transfer. Things become worse when multiple nodes are trying to transfer large amount of data with one node at the same time. Most likely nodes with lower priority have to wait for a long time before they can get channel while the nodes with higher priority are transferring data with low resource utilization (the bandwidth of other idle channels is wasted). Fig. 1 shows the simulated channel utilization in [1]. The average bandwidth utilization of RF-I arbitration is far from efficient with common benchmarks (32% for Blackscholes, 37.8% for EKF-SLAM, 60% for Deblur and 59.5% for Denoise). Unfortunately due to the limit of RF-I, dynamic bandwidth arbitration is not

This content is AI-processed based on ArXiv data.