Beyond Fat--tree: Unidirectional Load--Balanced Multistage Interconnection Network

The fat-tree is one of the most widely-used topologies by interconnection network manufacturers. Recently, it has been demonstrated that a deterministic routing algorithm that optimally balances the network traffic can not only achieve almost the same performance than an adaptive routing algorithm but also outperforms it. On the other hand, fat-trees require a high number of switches with a non-negligible wiring complexity. In this paper, we propose replacing the fat--tree by a unidirectional multistage interconnection network (UMIN) that uses a traffic balancing deterministic routing algorithm. As a consequence, switch hardware is almost reduced to the half, decreasing, in this way, the power consumption, the arbitration complexity, the switch size itself, and the network cost. Preliminary evaluation results show that the UMIN with the load balancing scheme obtains lower latency than fat--tree for low and medium traffic loads. Furthermore, in networks with a high number of stages or with high radix switches, it obtains the same, or even higher, throughput than fat-tree.     

OBQA: Smart and cost-efficient queue scheme for Head-of-Line blocking elimination in fat-trees

High-speed interconnection networks are essential elements for different high-performance parallel-computing systems. One of the most common interconnection network topologies is the fat-tree, whose advantages have turned it into the favorite topology of many interconnect designers. One of these advantages is the possibility of using simple but efficient routing algorithms, like the recently proposed deterministic routing algorithm referred to as DET, which offers similar (or better) performance than Adaptive Routing while reducing complexity and guaranteeing in-order packet delivery. However, as other deterministic routing proposals, DET cannot react when packets intensely contend for network resources, leading to the appearance of Head-of-Line (HoL) blocking which spoils network performance. In this paper, we describe and evaluate a simple queue scheme that efficiently reduces HoL-blocking in fat-trees using the DET routing algorithm, without significantly increasing switch complexity and required silicon area. Additionally, we propose an implementation of OBQA in a feasible switch architecture.     

片上网络路由算法分析与评估

本文对片上网络中的确定性XY路由算法和基于拐弯模型的4种自适应路由算法进行分析,并采用Noxim模拟器在6种合成通信模式下对5种路由算法的性能进行评估。实验结果表明,在均匀随机通信模式下,XY路由算法的性能优于自适应路由算法;在置换1和混洗通信模式下,奇偶路由算法的性能优于其他路由算法;在置换2、位反和蝶形通信模式下,负向优先路由算法的性能优于其他路由算法。     

自适应路由算法优于确定性路由算法

在研究并行计算机系统的容错时。自适应路由算法是一个极为重要的研究课题．它是在网络结点出错时，算法通过可选择的路径进行路由．在每个结点具有独立的出错概率的模型下，研究Mesh网络上自适应路由算法和确定性路算法的性能．本文提出的技术使得我们能严格地推导出路由算法的成功的概率，从而能分析和比较算法的性能．研究结果表明自适应路由算法具有明显的优势：一方面确定性路算法需要全局错误信息而变得高效性，另一方面自适应路由算法对于结点出错和网络规模具有更好的健壮性而具有更高的成功概率．     

Static load-balanced routing for slimmed fat-trees

Slimmed fat-trees have recently been proposed and deployed to reduce costs in High Performance Computing (HPC) clusters. While existing static routing schemes such as destination-mod-k (D-mod-k) routing are load-balanced and effective for full bisection bandwidth fat-trees, they incur significant load imbalance in many slimmed fat-trees. In this work, we propose a static load balanced routing scheme, called Round-Robin Routing (RRR$RRR$), for 2$2$- and 3$3$-level extended generalized fat-trees (XGFTs), which represent many fat-tree variations including slimmed fat-trees. RRR$RRR$ achieves near perfect load-balancing for any such XGFT in that links at the same level of a tree carry traffic from almost the same number of source–destination pairs. Our evaluation results indicate that on many slimmed fat-trees, RRR$RRR$ is significantly better than D-mod-k for dense traffic patterns due to its better load-balancing property, but performs worse for sparse patterns. We develop a combined routing scheme that enjoys the strengths of both RRR$RRR$ and D-mod-k by using RRR$RRR$ in conjunction with D-mod-k. The combined routing is a robust load-balanced routing scheme for slimmed fat-trees: it performs similar to D-mod-k for sparse traffic patterns and to RRR$RRR$ for dense patterns.     

计算机网络互连的路由算法的设计及仿真研究

互连网络目前应用最广泛、最流行的一种网络拓扑,广泛应用于多处理器系统、电话网络、分布式计算机系统及路由器交换机等领域。本文主要对直连网络的负载均衡路由算法进行了研究,提出了一种新的负载均衡路由算法。通过对该算法的仿真发现在相同的网络仿真环境下,该算法的性能要优于传统路由算法。     

2D Mesh片上网络分区容错路由算法

为了减小路由表的规模且避免使用较多虚通道(VC),从而降低硬件资源用量,针对虫孔交换的2D Mesh片上网络提出了一种分区容错路由(RFTR)算法。该算法根据故障节点和链路的位置将2D Mesh网络划分为若干个相连的矩形区域,数据包在矩形区域内可使用确定性或自适应路由算法进行路由,而在区域间则按照up^*/down^*算法确定路由路径。此外,利用通道依赖图(CDG)模型,证明了该算法仅需两个虚通道就能避免死锁。在6×6 Mesh网络中,RFTR算法能减少25%的路由表资源用量。仿真结果表明,在队列缓存资源相同的情况下,RFTR算法能实现与up^*/down^*算法和segment算法相当甚至更优的性能。     

Scalable Routing in Cyclic Mobile Networks

The nonexistence of an end-to-end path poses a challenge in adapting traditional routing algorithms to delay-tolerant networks (DTNs). Previous works have covered centralized routing approaches based on deterministic mobility, ferry-based routing with deterministic or semideterministic mobility, flooding-based approaches for networks with general mobility, and probability-based routing for semideterministic mobility models. Unfortunately, none of these methods can guarantee both scalability and delivery. In this paper, we extend the investigation of scalable deterministic routing in DTNs with repetitive mobility based on our previous works. Instead of routing with global contact knowledge, we propose a routing algorithm that routes on contact information compressed by three combined methods. We address the challenge of efficient information aggregation and compression in the time-space domain while maintaining critical information for efficient routing. Then, we extend it to handle a moderate level of uncertainty in contact prediction. Analytical studies and simulation results show that the performance of our proposed routing algorithm, DTN Hierarchical Routing (DHR), is comparable to that of the optimal time-space Dijkstra algorithm in terms of delay and hop count. At the same time, the per-node storage overhead is substantially reduced and becomes scalable.     

Cost‐effective queue schemes for reducing head‐of‐line blocking in fat‐trees

The fat‐tree is one of the most common topologies among the interconnection networks of the systems currently used for high‐performance parallel computing. Among other advantages, fat‐trees allow the use of simple but very efficient routing schemes. One of them is a deterministic routing algorithm that has been recently proposed, offering a similar (or better) performance than adaptive routing while reducing complexity and guaranteeing in‐order packet delivery. However, as other deterministic routing proposals, this deterministic routing algorithm cannot react when high traffic loads or hot‐spot traffic scenarios produce severe contention for the use of network resources, leading to the appearance of Head‐of‐Line (HoL) blocking, which spoils the network performance. In that sense, we describe in this paper two simple, cost‐effective strategies for dealing with the HoL‐blocking problem that may appear in fat‐trees with the aforementioned deterministic routing algorithm. From the results presented in the paper, we conclude that, in the mentioned environment, these proposals considerably reduce HoL‐blocking without significantly increasing switch complexity and the required silicon area. Copyright © 2011 John Wiley & Sons, Ltd.     

在3D-Mesh网络中的两种路由研究

在研究并行计算机系统容错时,路由算法是一个极为重要的研究课题。主要研究的是自适应路由算法和确定性路由算法在3D-Mesh网络上的性能。在每个结点具有独立的出错概率的模型下,提出的方法使得能够严格地推导出路由算法的成功概率,从而能够对算法进行分析和比较。研究结果表明,自适应路由算法具有明显的优势。一方面,自适应路由算法基于局部信息而变得高效;另一方面,自适应路由算法对于结点出错和网络规模具有更好的健壮性,而使其具有更高的成功概率。     

A performance model for analysis of heterogeneous multi-cluster systems

This paper addresses the problem of performance modeling for large-scale heterogeneous distributed systems with emphases on multi-cluster computing systems. Since the overall performance of distributed systems is often depends on the effectiveness of its communication network, the study of the interconnection networks for these systems is very important. Performance modeling is required to avoid poorly chosen components and architectures as well as discovering a serious shortfall during system testing just prior to deployment time. However, the multiplicity of components and associated complexity make performance analysis of distributed computing systems a challenging task. To this end, we present an analytical performance model for the interconnection networks of heterogeneous multi-cluster systems. The analysis is based on a parametric family of fat-trees, the m-port n-tree, and a deterministic routing algorithm, which is proposed in this paper. The model is validated through comprehensive simulation, which demonstrated that the proposed model exhibits a good degree of accuracy for various system organizations and under different working conditions.     

基于PRDT的16节点NoC路由算法

网络结构对于片上网络系统的性能和功耗发挥着重要作用，PRDT(2,1)有着较低的网络直径和平均距离、常数的节点度以及良好的可扩展性，这些特点使其非常适于NoC。为了提高小规模PRDT的路由性能，该文提出了一种binary路由算法，当网络规模不大于16时，该算法无须使用虚拟通道即可实现无死锁路由，通过增加少量虚拟通道，可改进为完全自适应路由算法。对所提出的路由算法与原有的向量路由算法进行仿真比较，结果显示binary算法在硬件成本较低的同时，性能更为优异，完全可以应用于基于PRDT的小规模NoC网络。     

一种QoS平面蚁群路由算法的设计与实现

提高网络服务质量的关键在于寻找出高性能路由,然而传统的路由算法却很难解决此类NP C问题。基于此,本文提出一种基于改进后的自适应蚁群算法的路由解决方案,将路由问题假设为平面路由,并建立相应的网络模型。针对该网络模型,建立特定的平面QoS蚁群路由算法,并在MATLAB上对其进行模拟仿真,从而验证了它的性能。仿真实验结果表明,该路由选择方案在求解实际网络路由问题时具有一定的优越性,能够有效地解决QoS平面网络路由问题。     

片上网络中面向报文有序传输的自适应路由算法研究

与确定性路由算法相比,自适应路由算法可以提高片上网络的通信性能,但是报文可能会无序到达。在目的节点对报文排序将会导致严重的面积和计算开销,甚至可能会抵消采用自适应路由算法带来的性能增益。为此,本文首先提出一种部分自适应路由算法,以满足报文的有序到达。然后,描述了对本文算法提供支持的路由器硬件结构。最后,在二维片上网络下对本文算法及确定性和自适应路由算法进行了性能评估和比较。与XY算法相比,本文算法显著降低了报文延时,提升了饱和点。同时讨论了对路由器面积和功耗影响。虽然路由器的功耗有所上升,但是由于报文交付性能提升,因此每个flit的能耗增长可忽略不计。     

Fuzzy-based Adaptive Routing Algorithm for Networks-on-Chip

In this paper, we propose two adaptive routing algorithms to alleviate congestion in the network. In the first algorithm, the routing decision is assisted by the number of occupied buffer slots at the corresponding input buffer of the next router and the congestion level of that router. Although this algorithm performs better than the conventional method, DyXY, in some cases the proposed algorithm leads to non-optimal decisions. Fuzzy controllers compensate for ambiguities in the data by giving a level of confidence rather than declaring the data simply true or false. To make a better routing decision, we propose an adaptive routing algorithm based on fuzzy logic for Networks-on-chip where the routing path is determined based on the current condition of the network. The proposed algorithm avoids congestion by distributing traffic over the routers that are less congested or have a spare capacity. The output of the fuzzy controller is the congestion level, so that at each router, the neighboring router with the lowest congestion value is chosen for routing a packet. To evaluate the proposed routing method, we use two multimedia applications and two synthetic traffic profiles. The experimental results show that the fuzzy-based routing scheme improves the performance over the DyXY routing algorithm by up to 25% with a negligible hardware overhead.     

Efficient network isolation and load balancing in multi-tenant HPC clusters

Multi-tenancy promises high utilization of available system resources and helps maintaining cost-effective operations for service providers. However, multi-tenant high-performance computing (HPC) infrastructures, like dynamic HPC clouds, bring unique challenges, both associated with providing performance isolation to the tenants, and achieving efficient load-balancing across the network fabric. Each tenant should experience predictable network performance, unaffected by the workload of other tenants. At the same time, it is equally important that the network links are balanced, avoiding network saturation. The network saturation can lead to unpredictable application performance, and a potential loss of profit for the cloud service providers.In this paper, we present two significant extensions to our previously proposed partition-aware fat-tree routing algorithm, pFTree, for InfiniBand-based HPC systems. First, we extend pFTree to incorporate provider defined partition-wise policies that govern how the nodes in different partitions are allowed to share network resources with each other. Second, we present a weighted version of the pFTree routing algorithm, that besides partitions, also takes node traffic characteristics into account to balance load across the network links more evenly. A comprehensive evaluation comprising both real-world experiments and simulations confirms the correctness and feasibility of the proposed extensions.     

2-Dilated flattened butterfly: A nonblocking switching topology for high-radix networks

High-performance computing is highly dependent on the communication network connecting the nodes. In this paper, we propose a 2-Dilated flattened butterfly (2DFB) network which provides non-blocking performance for relatively low cost overhead. We study the topological properties of the proposed 2DFB network and compare it with different nonblocking switching topologies. We observe that a dilation factor of two is sufficient to obtain nonblocking property for a flattened butterfly structure irrespective of its size or dimension. Dilating each link in a flattened butterfly causes an increase in cost. Therefore, we modeled the implementation cost of a 2DFB network and compared it with other popular nonblocking networks. We observe that the cost of a 2DFB is less than other nonblocking networks, while at the same time providing reduced latency because of its reduced diameter and hop count. We also propose a procedure to develop a conflict-free static routing schedule as well as an adaptive load balanced routing scheme (ALDFB) for 2DFB networks. Finally, we also describe the hardware implementation of a 2DFB network using the NetFPGA as the switching element and verify the nonblocking behavior of a 2DFB. We also show that the 2DFB topology can be used to build high speed switching systems with reduced cost.     

P2P计算网格路由和负载均衡算法

路由和负载均衡是P2P计算网格的两个技术难题,由于P2P网络的分布性和动态性,以及缺乏统一的中心控制,使得传统的路由和负载均衡算法不能应用于P2P网络。提出了一种源自蚁群智能的混合路由和负载均衡算法,通过移动代理,即人工蚂蚁在节点间移动时所释放的信息素来作为路由和任务调度的依据。仿真结果表明该算法是有效的,且适用于具有分散和自组织特性的P2P网络。     

Explanation of Performance Degradation in Turn Model

The Turn model routing algorithms for mesh interconnection network achieve partial adaptivity without any virtual channels. However, the routing performance measured by simulations is worse than with the simple deterministic routing algorithm. Authors have explained these results simply by uneven dynamic load through the network. However, this phenomenon has not been studied further. This paper investigates performance degradation with Turn model and drawbacks of partially adaptive routing in comparison with the deterministic routing, and it introduces some new concepts. Our simulations deal with individual channels and results are presented by 3D graphs, rather than by commonly used averages. An additional parameter—channel occupation, which is consistent with queuing theory commonly used in many proposed analytical models, is introduced. We also propose a new structure, the Channel Directions Dependency Graph (CDDG). It provides a new approach in analysis, helps in understanding of dynamic routing behaviour, and it can be generalized in other routing algorithms.     

Probabilistic odd–even: an adaptive wormhole routing algorithm for 2D mesh network-on-chip

Wormhole routing is a popular routing technique used in network-on-chip. It is efficient but susceptible to deadlock, while deadlock will significantly degrade the network performance of NoC. Most existing adaptive wormhole routings avoid deadlock by reducing the degree of adaptiveness and thus sacrificing network performance. In this paper, we address both deadlock and network performance issues jointly, and propose a probabilistic odd–even (POE) routing algorithm that achieves the minimum packet delivery delay. The proposed POE dynamically adjusts the probabilities of constrained turns that may lead to deadlocks according to the current network conditions, and uses an efficient deadlock detection and recovery scheme when a deadlock happens. By adopting constrained turns adaptively to the network status, it not only reduces the frequency of deadlock and allows the network to be swiftly recovered when it occurs, but also greatly improves the degree of adaptiveness to obtain high network performance. Experimental results show that our method achieves a significant performance improvement both in terms of network throughput and average packet latency compared with the existing methods such as XY, odd–even, abacus turn model and fully adaptive routing algorithm while it only has moderate energy consumption.