An analysis of deflection-based wormhole routing with virtual channels

Network-on-Chip (NoC) is a key element in the total power consumption of a chip multiprocessor. Dynamic Voltage Scaling is a promising method for power saving in NoCs since it contributes to reduction in both static and dynamic power consumptions. In this paper, we propose a novel scheme to reduce on-chip network power consumption when the number of Virtual Channels (VCs) with active allocation requests per cycle is less than the number of total VCs. In our method, we introduce a reconfigurable arbitration logic which can be configured to have multiple latencies and hence, multiple slack times. The increased slack times are then used to reduce the supply voltage of the routers in order to reduce the power consumption. By using this method, we manage to save power by up to 45.7% compared to a baseline architecture without any performance loss.

Wormhole routing makes the communication latency insensitive of the network diameter and reduces the buffer size of each router. It has been widely adopted in many new generation distributed memory multiprocessor systems (DMMPs). The concept of virtual system is invented for deadlock-free routing design. Using a great deal of virtual channels is quite expensive to be implemented in practice and is not limitless. In this paper experimental results are given to evaluate the tradeoff between network congestion delay and throughput. Some results show that the congestion delay does not reduce visibly when the number of virtual channels is greater than that of internal channels by 2. A reasonable number of virtual channels can be derived from the results. A virtual networks system using only four virtual networks, two virtual channels in each direction, is presented. It shows that the minimum number of virtual networks to support a partially adaptive, minimal and deadlock free routing in k-ary n-cubes is just four. It uses only four virtual networks but can get a high degree of adaptiveness and high traffic capacity. Simulation results are presented to verify the performance.

High throughput and low latency are the desirable characteristics of a multi processing system. CLICHE (ChipLevel Integration of Communicating Heterogeneous Elements) is the simplest form of Network-on-Chip architecture in terms of the layout and the local interconnections between resources and switches. The throughput varies with the number of virtual channels. When the number of virtual channels is increased, the throughput is increased. On the other hand, the average message latency increases with the number of virtual channels. To keep the latency low while preserving the throughput, the number of virtual channels is constrained to certain number. A novel multichannel signal acquisition module is proposed to integrate IP blocks. The proposed module achieves the optimal solution for large number of virtual channels. It reduces the memory requirement and power consumption as well. A Field Programmable Gate Arrays-based prototype is developed to verify the proposed module. Different case studies are studied in this research. As a result, the proposed module can schedule given tasks that are not schedulable via round-robin technique. Even in scenarios that both round-robin and proposed module can successfully schedule the given task, proposed module reduces the amount of data being acquired by 59%, which in turn decreases memory requirements and power consumption as well.

The adaptive routing approach has been expected as a promising way to improve network performance by utilizing available network bandwidth. Previous adaptive routing strategies in wormhole-routed multicomputer networks restrict the routing of messages by the routing algorithm to prevent deadlock. This results in low degree of adaptivity and low utilization of physical or virtual channels. In this paper, we examine the possibility of performing restriction-free adaptive routing in wormhole-routed networks as an approach to further improving the performance of these networks. A new flow control policy, called message cutting-in, is proposed, and two adaptive routing strategies are presented. Freedom of communication deadlock is achieved by the proposed flow control policy. The proposed adaptive routing strategies do not restrict routing and maximally utilize the physical and virtual channels. Simulation results show that the restriction-free adaptive routing approach is promising from the fact that it has the lowest latency and highest throughput depending on the number of virtual channels per physical channel and patterns of message traffic.

An experimental study was conducted using a network simulator to investigate the performance of packet communication networks as a function of: the network resource capacities (channels, buffers), the network load (number of virtual channels, virtual channel loads), protocols (flow control, congestion control, routing), and protocol parameters (virtual channel window size, input buffer limits). Performance characteristics are shown and the design of input buffer limits for network congestion control, virtual channel window size, and nodal buffer capacity addressed. Network design strategies for the control of load fluctuations are proposed and discussed.

Networks-on-Chip (NoCs) for embedded multiprocessor architectures require special mechanisms for sharing resources such as links and queues in network routers. A widely used solution is virtual-channel flow control which prevents head-of-line blocking and isolates different data streams. However, the number of virtual channels (VCs) is limited by the available buffer space in the routers. Due to the evergrowing number of applications and mutual exclusive requirements it may not be possible to assign each application a unique VC. In this paper we present a Resource Broker (RB) which is an overarching mechanism that schedules NoC traffic. It allows predictable and efficient sharing of VCs without compromising isolation and Quality-of-Service guarantees. We show that by managing the NoC traffic with the RB, we drastically reduce blocking and improve the overall utilization and hence the system performance.

A new deadlock prevention scheme for nonminimal adaptive wormhole routing

Carrier-sense multiple access (CSMA) algorithms have recently received significant interests in the literature for designing wireless control algorithms. CSMA algorithms incur low complexity and can achieve the optimal capacity under certain assumptions. However, CSMA algorithms suffer the starvation problem and incur large delay that may grow exponentially with the network size. In this paper, our goal is to develop a new algorithm that can provably achieve high throughput utility and low delay with low complexity. Toward this end, we propose a new CSMA-like algorithm, called Virtual-Multi-Channel CSMA (VMC-CSMA), that can dramatically reduce delay. The key idea of VMC-CSMA to avoid the starvation problem is to use multiple virtual channels (which emulate a multichannel system) and compute a good set of feasible schedules simultaneously (without constantly switching/recomputing schedules). Under the protocol interference model and a single-hop utility-maximization setting, VMC-CSMA can approach arbitrarily close-to-optimal system utility with both the number of virtual channels and the computation complexity increasing logarithmically with the network size. Furthermore, once VMC-CSMA converges to the steady state, we can show that under certain assumptions on the utility functions and the topology, both the expected packet delay and the tail distribution of the head-of-line (HOL) waiting time at each link can be bounded independently of the network size. Our simulation results confirm that VMC-CSMA algorithms indeed achieve both high throughput utility and low delay with low-complexity operations.

Current High-Performance Computing (HPC) and data center networks rely on large-radix routers. Hamming graphs (Cartesian products of complete graphs) and dragonflies (two-level direct networks with nodes organized in groups) are some direct topologies proposed for such networks. The original definition of the dragonfly topology is very loose, with several degrees of freedom, such as the inter- and intragroup topology, the specific global connectivity, and the number of parallel links between groups (or trunking level). This work provides a comprehensive analysis of the topological properties of the dragonfly network, providing balancing conditions for network dimensioning, as well as introducing and classifying several alternatives for the global connectivity and trunking level. From a topological study of the network, it is noted that a Hamming graph can be seen as a canonical dragonfly topology with a high level of trunking. Based on this observation and by carefully selecting the global connectivity, the Dimension Order Routing (DOR) mechanism safely used in Hamming graphs is adapted to dragonfly networks with trunking. The resulting routing algorithms approximate the performance of minimal, nonminimal, and adaptive routings typically used in dragonflies but without requiring virtual channels to avoid packet deadlock, thus allowing for lower cost router implementations. This is obtained by properly selecting the link to route between groups based on a graph coloring of network routers. Evaluations show that the proposed mechanisms are competitive with traditional solutions when using the same number of virtual channels and enable for simpler implementations with lower cost. Finally, multilevel dragonflies are discussed, considering how the proposed mechanisms could be adapted to them.

Recently, the growth of applications and services over high-speed Internet increases, ATM networks as wide area back-bone has been a major solution. As the conventional TCP/IP suite is still the standard protocol used to support upper application on current. Internet, the issues regarding whether TCP/IP will operate efficiently on top of an ATM infrastructure and how to control its QoS still remain for studies. TCP uses a window-based protocol for flow control in the transport layer. When TCP uses the UBR service in ATM layer, the control method is only buffer management. If a cell is discarded in ATM layer, one whole packet of TCP will be lost; this fact occur the most TCP performance degradation. Several dropping strategies, such as Tail Drop, EPD, PPD, SPD, FBA, have been proposed to improve the TCP performance over ATM. In this paper, to improve the TCP performance, we propose a packet dropping scheme that is based on comparison with EPD, SPD and FBA. Our proposed scheme is applied to schemes discussed in the previous technology. Our proposed scheme does not need to know each connection's mean packet size. When the buffer exceeds the given threshold, it is based on comparison between the number of dropped packet and the approved packet. Our results are reported and discussed for comparing these discarding schemes under similar conditions. Although the number of virtual channel (VC) is increased, the simulation results showed that the proposed scheme can allocate more fairly each VC than other scheme.

For constructing massively parallel multicomputers with over 5000 processing nodes, the star graph topology is known to be better than the hypercube in terms of the average routing distance, the number of links per node, and the fault diameter. The authors present two deadlock-free algorithms for routing in star graph, assuming the wormhole routing model. Both the algorithms use the concept of virtual channels introduced by W. Dally and C.L. Seitz (1987). The first algorithm is non-optimal in terms of the average routing distance, but uses fewer virtual channels on the whole. The second algorithm is optimal in terms of routing performance, but requires a somewhat larger number of virtual channels per node.

This paper identifies performance degradation in wormhole routed k-ary n-cube networks due to limited number of router-to-processor consumption channels at each node. Many recent research in wormhole routing have advocated the advantages of adaptive routing and virtual channel flow control schemes to deliver better network performance. This paper indicates that the advantages associated with these schemes cannot be realized with limited consumption capacity. To alleviate such performance bottlenecks, a new network interface design using multiple consumption channels is proposed. To match virtual multiplexing on network channels, we also propose each consumption channel to support multiple virtual consumption channels. The impact of message arrival rate at a node on the required number of consumption channels is studied analytically. It is shown that wormhole networks with higher routing adaptivity, dimensionality, degree of hot-spot traffic, and number of virtual lanes have to take advantage of multiple consumption channels to deliver better performance. The interplay between system topology, routing algorithm, number of virtual lanes, messaging overheads, and communication traffic is studied through simulation to derive the effective number of consumption channels required in a system. Using the ongoing technological trend, it is shown that wormhole-routed systems can use up to two-four consumption channels per node to deliver better system performance.

A Comparison of Router Architectures for Virtual Cut-Through and Wormhole Switching in a NOW Environment

The past few years have seen a rise in popularity of massively parallel architectures that use fat-trees as their interconnection networks. In this paper we formalize a parametric family of fat-trees, the k-ary n-trees, built with constant arity switches interconnected in a regular topology. A simple adaptive routing algorithm for k-ary n-trees sends each message to one of the nearest common ancestors of both source and destination, choosing the less loaded physical channels, and then reaches the destination following the unique available path. Through simulation on a 4-ary 4-tree with 256 nodes, we analyze some variants of the adaptive algorithm that utilize wormhole routing with 1, 2 and 4 virtual channels. The experimental results show that the uniform, bit reversal and transpose traffic patterns are very sensitive to the flow control strategy. In all these cases, the saturation points are between 35–40% of the network capacity with 1 virtual channel, 55–60% with 2 virtual channels and around 75% with 4 virtual channels. The complement traffic, a representative of the class of the congestion-free communication patterns, reaches an optimal performance, with a saturation point at 97% of the capacity for all flow control strategies. In this case virtual channels are of little help and the average network latency experiences an increase proportional to the number of virtual channels, due to the multiplexing of several packets onto the same physical links.

On the Benefit of Supporting Virtual Channels in Wormhole Routers