EDXY – A low cost congestion-aware routing algorithm for network-on-chips

Abstract

In this paper, an adaptive routing algorithm for two-dimensional mesh network-on-chips (NoCs) is presented. The algorithm, which is based on Dynamic XY (DyXY), is called Enhanced Dynamic XY (EDXY). It is congestion-aware and more link failure tolerant compared to the DyXY algorithm. On contrary to the DyXY algorithm, it can avoid the congestion when routing from the current switch to the destination whose X position (Y position) is exactly one unit apart from the switch X position (Y position). This is achieved by adding two congestion wires (one in each direction) between each two cores which indicate the existence of congestion in a row (column). The same wires may be used to alarm a link failure in a row (column). These signals enable the routing algorithm to avoid these paths when there are other paths between the source and destination pair. To assess the latency of the proposed algorithm, uniform, transpose, hotspot, and realistic traffic profiles for packet injection are used. The simulation results reveal that EDXY can achieve lower latency compared to those of other adaptive routing algorithms across all workloads examined, with a 20% average and 30% maximum latency reduction on SPLASH-2 benchmarks running on a 49-core CMP. The area of the technique is about the same as those of the other routing algorithms.

Introduction

Recently on-chip transistor density has increased enabling the integration of dozens of intellectual property cores on a single die to form system-on-chips (SoCs). One byproducts of the greater integration is that for communication in these systems, shared buses should be replaced by interconnection networks. The network-on-chips (NoCs) has been proposed as a new paradigm for realizing complex SoCs [1], [2]. NoCs scale better than traditional forms of on-chip interconnections and have better performance and fault tolerant characteristics [2]. Among different possible topologies, the two-dimensional mesh is one of the most common topologies [3], [4].

In NoCs, routing algorithms are used to determine the path of a packet from the source to the destination. These algorithms are classified as deterministic and adaptive. The implementations of deterministic routing algorithms are simple but they are not able to balance the load across the links in non-uniform or bursty traffic [5], [6]. Adaptive routing algorithms are proposed to address these limitations. By better distributing load across links, adaptive algorithms improve network performance and also provide tolerance if link or router failure occurs. In adaptive routing algorithms, the path of a packet from the source to the destination is determined by the network condition. An adaptive routing algorithm decreases the probability of passing a packet from a congested or mal-function link. Despite its implementation complexity, adaptive routing is attractive for large NoCs especially when these NoCs facing with non-uniform or bursty traffic.

There are a number of routing algorithms which we briefly review those related to the algorithm proposed in this work. In [7], a static routing algorithm for two-dimensional meshes which is called XY is introduced. In this routing algorithm, each packet first travels along the X and then the Y direction to reach the destination. For this method, deadlock never occurs but no adaptivity exists in this algorithm. An adaptive routing algorithm named turn-model is introduced in [8] based on which another adaptive routing algorithm called Odd–Even turn is proposed in [9]. To avoid deadlock, Odd–Even method restricts the position that turns are allowed in the mesh topology. Another algorithm called DyAD is introduced in [10]. This algorithm is a combination of a static routing algorithm called oe-fix, and an adaptive routing algorithm based on the Odd–Even turn algorithm. Depending on the congestion condition of the network, one of the routing algorithms is invoked. Another adaptive routing is hot potato or deflection routing [11], [12] which is based on the idea of delivering a packet to an output channel at each cycle. If all the channels belonging to minimal paths are occupied, then the packet is misrouted. When contention occurs and the desired channel is not available, the packet, instead of waiting, will pick any alternative available channels (minimal or non-minimal) to continue moving to the next router; therefore the router does not need buffers. In hot potato routing, if the number of input channels is equal to the number of output channels at every router node, packets can always find an exit channel and they are deadlock free. However, livelock is a potential problem in this routing. Also, hot potato increases message latency even in the absence of congestion and bandwidth consumption. Accordingly, performance of hot potato routing is not as good as other wormhole routing methods [13]. Also, there are adaptive routings for increasing fault tolerance of the on-chip network. Stochastic communication method has been proposed to deal with permanent and transient faults of network links and nodes [14]. This method has the advantage of simplicity and low overhead. The selection of links and of the number of redundant copies to be sent on the links is stochastically done at runtime by the network routers. As a result, the transmission latency is unpredictable and, hence, it cannot be guaranteed. Also, stochastic communication is not efficient in terms of power dissipation and latency.

An adaptive deadlock free routing algorithm called Dynamic XY (DyXY) has been proposed in [15]. In this algorithm, which is based on the static XY algorithm, a packet is sent either to the X or Y direction depending on the congestion condition. It uses local information which is the current queue length of the corresponding input port in the neighboring routers to decide on the next hop. It is assumed that the collection of these local decisions should lead to a near-optimal path from the source to the destination. The main weakness of DyXY is that the use of the local information in making routing decision could forward the packet in a path which has congestion in the routers farther than the current neighbors. This situation could happen when the routing unit is one unit apart from the destination in X or Y dimension. Such non-optimal routing decisions will cause NoC to face with increasing in its network latency. The technique described in [16], may overcome this problem. It uses global information in making a routing decision. The technique requires a mechanism to mix local and global congestion information. This has been obtained at the cost of higher hardware overhead.

In this paper, we propose a technique for solving the problem of the DyXY routing algorithm with little area overhead. In addition, the proposed technique increases tolerance against single link failure compared to the DyXY technique. The rest of this paper is organized as follows: Section 2 describes the basic structure of the XY and DyXY routers. Section 3 describes the proposed routing algorithm and its architecture. Single link failure tolerances of the routing algorithm are compared in Section 4 while experimental results are discussed in Section 5. Finally, the conclusion of the paper is given in Section 6.