Data Center Topology Research Articles

Intelligent load balancing in data center software‐defined networks

AbstractIn response to the increasing demand for efficient resource utilization in data center networks (DCNs), the development of intelligent load‐balancing algorithms has become crucial. This article introduces the dual double deep Q network (DDQN) algorithm, designed for software‐defined networking (SDN) environments within data centers. By leveraging deep reinforcement learning, DDQN addresses the challenges posed by dynamic traffic patterns, diverse flow requirements, and the coexistence of elephant and mice flows. Our algorithm adopts a comprehensive SDN approach, evaluating the network's status by analyzing switch load and bandwidth utilization. Using convolutional neural networks for elephant and mice flows in DCN, our algorithm enables adaptive learning and training tailored to the specific demands of elephant flows. Employing a double deep Q network architecture (DDQN), DDQN optimizes paths for both elephant and mice flows independently. Real‐time adaptation mechanisms make routing decisions based on the robust learning capabilities of DDQN, enhancing network utilization and reducing packet loss by generating optimal forwarding paths according to the current network state and traffic patterns. Simulations conducted in a Mininet environment with RYU as the controller, utilizing a fat‐tree data center topology, validate the efficacy of DDQN. The results demonstrate its effectiveness in achieving higher throughput, lower latency, and superior load balancing compared to traditional algorithms like equal‐cost multipath and Hedera.

PDLB: Path Diversity-aware Load Balancing with adaptive granularity in data center networks

Modern data center topologies often take the form of a multi-rooted tree with rich parallel paths to provide high bandwidth. However, various path diversities caused by traffic dynamics, link failures and heterogeneous switching equipment widely exist in the production datacenter network. Therefore, the multi-path load balancer in data center should be robust to these diversities. Although prior fine-grained schemes such as RPS and Presto make the best use of available paths, However, they are prone to experiencing packet reordering problem under the asymmetric topology. The coarse-grained solutions such as ECMP and LetFlow effectively avoid packet reordering, but easily lead to under-utilization of multiple paths. To cope with these inefficiencies, we propose a load balancing mechanism called PDLB, which adaptively adjusts flowcell granularity according to path diversity. PDLB increases flowcell granularity to alleviate packet reordering under large degrees of topology asymmetry, while reducing flowcell granularity to obtain high link utilization under small degrees of topology asymmetry. PDLB is only deployed on the sender without any modification on switch. We evaluate PDLB through large-scale NS2 simulations. The experimental results show that PDLB reduces the average flow completion time by up to ∼\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\sim$$\\end{document}8-53% over the state-of-the-art load balancing schemes.

Path Diversity and Survivability for the HyperX Datacenter Topology

Network survivability has been recognized as an issue of a major importance in terms of security, stability and prosperity. This paper studies fundamental properties of HyperX, an emerging topology for connecting supercomputing and datacenter networks. We focus on the establishment of paths with guaranteed survivability, allowing path availability even upon a restricted number of link failures. We first examine the availability of disjoint paths connecting a pair of input nodes. Disjoint paths guarantee path existence even upon a bounded number of link failures. We explore the inherent tradeoff between allowing slightly longer paths and the ability to extend available sets of mutually disjoint paths. Second, we study the availability of paths in a HyperX topology that already observed link failures. Such failures can increase the length of available paths or even eliminate connectivity between network parts. We also compare basic properties of shortest paths in HyperX to other datacenter topologies. Last, we provide an evaluation to illustrate path availability along with the potential impact of failures.

ABLA: Application-Based Load-Balanced Approach for Adaptive Mapping of Datacenter Networks

Cloud-based services are growing more rapidly than ever, and so does the management challenge on their providers’ side. Cloud-based datacenter networks are built with nodes of huge processing power, connected by high bandwidth capacities to carry their interior traffic requirements. However, such cloud networks still have limits that are imposed not necessarily by their physical components, but by the schemes of resource management being deployed. Traditionally, for an institute to provide services, it needs to have its own datacenter facility that interconnects its servers through a topology that matches its desired administrative policies and scaling objectives. With the theme of cloud-based IaaS, such datacenter topologies can be created virtually over the cloud. Nowadays, a significant part of those institutes who provide us with our daily services have their infrastructures hosted over cloud ones. Therefore, resources of such cloud networks need to be efficiently utilized, in order to keep their performance and hosting prices competitive. A typical datacenter network mainly consists of server nodes and network links. Besides the resources of the server nodes, the network bandwidth resources are considered a crucial key determinant for the whole datacenter performance. Indeed, a server without sufficient bandwidth capacities is almost useless. Proposals in the literature present schemes for resource utilization on either side of the problem at a time: the nodes or the links. Working in isolation can never deliver efficient mapping solutions. ABLA is an Application-Based, and Load balancing Approach for adaptive mapping proposal. ABLA’s methodology tackles both sides of the datacenter, the nodes and links. It starts by (1) breaking down the node’s resource requirement for the requested applications to be hosted over the virtual server machines besides (2) reading the topological connectivity and bandwidth requirements for each virtual node to all other nodes in the virtual datacenter topology. Compared to other models in the literature, the simulation results show that our proposed ABLA model provides for complete mapping services via load-balanced hosting networks. This allows for competitive hosting prices, with higher performance and service satisfaction rates.

Formal Algebraic Model of an Edge Data Center with a Redundant Ring Topology

Data center organization and optimization presents the opportunity to try and design systems with specific characteristics. In this sense, the combination of artificial intelligence methodology and sustainability may lead to achieve optimal topologies with enhanced feature, whilst taking care of the environment by lowering carbon emissions. In this paper, a model for a field monitoring system has been proposed, where an edge data center topology in the form of a redundant ring has been designed for redundancy purposes to join together nodes spread apart. Additionally, a formal algebraic model of such a design has been exposed and verified.

Large-Scale Machine Learning Cluster Scheduling via Multi-Agent Graph Reinforcement Learning

Efficient scheduling of distributed deep learning (DL) jobs in large GPU clusters is crucial for resource efficiency and job performance. While server sharing among jobs improves resource utilization, interference among co-located DL jobs occurs due to resource contention. Interference-aware job placement has been studied, with white-box approaches based on explicit interference modeling and black-box schedulers with reinforcement learning. In today’s clusters containing thousands of GPU servers, running a single scheduler to manage all arrival jobs in a timely and effective manner is challenging, due to the large workload scale. We adopt multiple schedulers in a large-scale cluster/data center, and propose a multi-agent reinforcement learning (MARL) scheduling framework to cooperatively learn fine-grained job placement policies, towards the objective of minimizing job completion time (JCT). To achieve topology-aware placements, our proposed framework uses hierarchical graph neural networks to encode the data center topology and server architecture. In view of a common lack of precise reward samples corresponding to different placements, a job interference model is further devised to predict interference levels in face of various co-locations, for training of the MARL schedulers. Testbed and trace-driven evaluations show that our scheduler framework outperforms representative scheduling schemes by more than 20% in terms of average JCT, and is adaptive to various machine learning cluster topologies.

Fast and scalable routing protocols for data center networks

Data center networks may comprise tens or hundreds of thousands of nodes, and, naturally, suffer from frequent software and hardware failures as well as link congestions. Packets are routed along the shortest paths with sufficient resources to facilitate efficient network utilization and minimize delays. In such dynamic networks, links frequently fail or get congested, making the recalculation of the shortest paths a computationally intensive problem. Various routing protocols were proposed to overcome this problem by focusing on network utilization rather than speed. Surprisingly, the design of fast shortest-path algorithms for data centers was largely neglected, though they are universal components of routing protocols. Moreover, parallelization techniques were mostly deployed for random network topologies, and not for regular topologies that are often found in data centers. The aim of this paper is to improve scalability and reduce the time required for the shortest-path calculation in data center networks by parallelization on general-purpose hardware. We propose a novel algorithm that parallelizes edge relaxations as a faster and more scalable solution for popular data center topologies.

SiC Mass Commercialization: Present Status and Barriers to Overcome

In an increasingly electrified technology driven world, power electronics is central to the entire clean energy manufacturing economy. Silicon (Si) power devices have dominated power electronics due to their low cost volume production, excellent starting material quality, ease of fabrication, and proven reliability. Although Si power devices continue to improve, they are approaching their operational limits primarily due to their relatively low bandgap, critical electric field, and thermal conductivity that result in high conduction and switching losses, and poor high temperature performance. Silicon Carbide’s (SiC) compelling efficiency and system benefits have led to significant development efforts over the last two decades and today planar and trench MOSFETs, and JFETs are commercially available from several vendors as discrete components or in high power modules in the of 650 V to 1700 V voltage range. High impact application opportunities, where SiC devices are displacing their incumbent Si counterparts, have emerged and include automotive and rail power electronics with reduced losses and reduced cooling requirements; novel data center topologies with reduced cooling loads and higher efficiencies; variable frequency drives for efficient high power electric motors at reduced overall system cost; more efficient, flexible, and reliable grid applications with reduced system footprint; and “more electric aerospace” with weight, volume, and cooling system reductions contributing to energy savings. In particular, SiC insertion in electric vehicles brings major competitive advantages and is a volume application opportunity that can spur manufacturing economies of scale and lower system costs. As SiC continues to grow, the industry is lifting the last barriers to mass commercialization that include higher than Si device cost, relative lack of wafer planarity, the presence of basal plane dislocations, reliability and ruggedness concerns, and the need for a workforce skilled in SiC power technology to keep up with the rising demand. It should be noted that in many applications, insertion of SiC reduces overall system cost compared to Si even though SiC devices can cost 2-3 more than their Si counterparts. This is due to the passive component and cooling system simplifications enabled by the efficient high frequency SiC operation. In this paper, we will review key aspects of SiC technology and discuss overcoming barriers to mass commercialization.

Improved Scalability of Demand-Aware Datacenter Topologies With Minimal Route Lengths and Congestion

The performance of more and more cloud-based applications critically depends on the performance of the interconnecting datacenter network. Emerging reconfigurable datacenter networks have the potential to provide an unprecedented throughput by dynamically reconfiguring their topology in a demand-aware manner. This paper studies the algorithmic problem of how to design low-degree and hence scalable datacenter networks optimized toward the current traffic they serve. Our main contribution is a novel network design which provides asymptotically minimal route lengths and congestion. In comparison to prior work, we reduce the degree requirements by a factor of four for sparse demand matrices. We further show the problem to be already NP-hard for tree-shaped demands, but permits a 2-approximation on the route lengths and a 6-approximation for congestion. We further report on a small empirical study on Facebook traces.

WildMinnie: compression of software-defined networking (SDN) rules with wildcard patterns.

Software-defined networking (SDN) enables fast service innovations through network programmability. In SDN, a logically centralized controller compiles a set of policies into the network-level rules. These rules are inserted in the TCAM memory of SDN-enabled switches enabling high-speed matching and forwarding of packets. Unfortunately, TCAMs are available in limited capacities and fall short of accommodating all intended rules, especially in networks with large distinct flows like datacenters. Rule compression is a technique that reduces the number of rules by aggregating them with some similarity factors. This paper introduces WildMinnie, a new rule compression method that aggregates rules based on their common address non-prefix wildcards derived from a group of rules with the same output port number. We explore rule conflict issues and provide solutions to resolve them. We demonstrate the capability of WildMinnie in various datacenter topologies with traffics having different diversity of source-destination addresses and show that WildMinnie outperforms the best-known compression method by 20%, on average.

Improved scalability of demand-aware datacenter topologies with minimal route lengths and congestion

The performance of more and more cloud-based applications critically depends on the performance of the interconnecting datacenter network. Emerging reconfigurable datacenter networks have the potential to provide an unprecedented throughput by dynamically reconfiguring their topology in a demand-aware manner. This paper studies the algorithmic problem of how to design low-degree and hence scalable datacenter networks that are optimized toward the current traffic they serve. Our main contribution is a novel network design which provides asymptotically minimal route lengths and congestion. In comparison to prior work, our design reduces the degree requirements by a factor of four for sparse demand matrices. We further show that the problem is already NP-hard for tree-shaped demands, but permits a 2-approximation on the route lengths and a 6-approximation for congestion. We further report on a small empirical study on Facebook traces.

DeepRLB: A deep reinforcement learning‐based load balancing in data center networks

SummaryData center networks (DCNs) are facing challenging control problems such as flow scheduling, congestion control, load balancing, and bandwidth allocation when dynamically handling heterogeneous mice and elephant flows. This paper focuses on the load balancing in DCNs as an online decision‐making problem with high complexity (NP‐hard problem). Traditional load balancing approaches are usually the heuristic algorithms relying on the operator's viewpoint and a short‐term knowledge of the traffic conditions and network environment. This paper proposes DeepRLB, a deep reinforcement learning (DRL)‐based load balancing approach for software‐defined networking‐based DCNs which exploits the deep deterministic policy gradient (DDPG) algorithm to adaptively learn the link‐weight values by observing the traffic flow characteristics. The actor and critic neural networks in the DDPG model have been designed based on two learning models: fully connected (DeepRLB‐Full) and convolutional (DeepRLB‐Conv). The output of the model is further used by the controller to determine the forwarding paths for traffic flows accordingly. We evaluated the performance of DeepRLB by comparing it with ECMP as well as a per‐packet load balancing algorithm in three different data center topologies under 30 heterogeneous traffic demand matrices. The obtained results showed that the DeepRLB algorithm outperforms the ECMP with respect to studied load balancing metrics considerably (DeepRLB‐Conv achieved the best performance). As a sample, DeepRLB‐Full and DeepRLB‐Conv were able to reach 19.14% and 24.46% reduction in the total number of over‐utilized and under‐utilized network links compared to ECMP, respectively.

SRFabric: A Semi-Reconfigurable Rack Scale Topology

Rack scale design is a promising trend towards customized hardware design, where high density clusters of SoCs are integrated in the rack. One of the biggest challenges for rack scale computing is the interconnection network. Traditional data center topologies require too many ToR switches to support hundreds of SoCs, while distributed fabrics deliver a considerably high end-to-end latency and network oversubscription. Since no one topology fits all kinds of workloads, a flexible in-rack topology requires a careful redesign to dynamically adapt to diverse data center traffic within tight cost and space constraints in the rack. SRFabric is a semi-reconfigurable rack scale network topology that exploits the high path diversity, the cost-effectiveness of distributed fabrics, and the dynamic reconfigurability of circuit switches. This is accomplished by enabling multiple static ports and dynamic ports for each SoC. Leveraging the partial link reconfigurability, SRFabric is able to optimize its topology to dynamically adapt to various workload patterns. We further propose the design of SRFabric to decide the nearly optimal number of dynamic ports and static ports for expected communication density and performance. Extensive evaluations demonstrate that SRFabric can deliver lower average path length, i.e., 2.21 hops on average, and higher bisection bandwidth, i.e., up to 77% nonblocking bandwidth, and provide comparable performance with state-of-the-art strategy XFabric at a lower cost, i.e., XFabric costs up to 3 times more than that of SRFabric.

A community detection based approach for Service Function Chain online placement in data center network

With the emerging paradigm of Network Function Virtualization (NFV), the Internet Service Provider (ISP) can outsource their service functions to the cloud data center (DC) to reduce Operating Expenditures (OPEX) and Capital Expenditures (CAPEX). In this paper, we study the virtual network function (VNF) online placement and migration problem in DC considering user’s Service Function Chain (SFC). In order to solve the practical problems, we take the data center topology, Basic Resource Consumption, multi-tenancy, flow characteristics, and VNF relationship into consideration. Firstly, we formulate this problem into a dynamic programming model with the aim to minimize the average operational cost in a long term. To reduce the complexity of the online decision, an online two-stage heuristic (OTSH) algorithm is designed to optimally place SFCs. The OTSH consists of a community detection based differentiated greedy algorithm for SFC mapping and an offline iterative migration algorithm for VNF migrating. At last, the joint online heuristic algorithm is proven to make intelligent predictions based on the historical traffic and provide good performance guarantees by simulation.

Topology Independent Multipath Routing for Data Center Networks

Data center communication networks are characterized by fast and intensive traffic changes, to the extent that a statistical approach to routing has significant drawbacks. This paper proposes the first routing solution that takes into account every data center flow. In order to achieve the required performance, the proposed algorithm operates in the data plane, i.e. in the packet processing pipeline of the programmable packet forwarding chips. Route updates are triggered by the arrival of new flows, and are immediately transferred to the upstream switches in order for them to direct newly arriving flows along the best path. This enables the proposed algorithm to handle fast and intensive traffic changes, such as large-scale flow incast communication patterns which historically presented challenges to the data center routing algorithms. With the explicit knowledge of the current path state, the proposed algorithm can improve routing performance in the data center topologies used today. Additionally, being able to operate with high performance in arbitrary topology, the proposed solution can enable the introduction of new topologies into the data center networks, instead of topologies designed with the goal of supporting equal-cost multipath routing which are dominant in modern data centers.

Clustered Multicast Source Routing for Large-Scale Cloud Data Centers

The multi-tenancy concept in cloud data center (DC) networks paves the way towards advancements and innovation in the underlying infrastructure such as network virtualization. Multicast routing is essential in leveraging multi-tenancy to its full potential. However, traditional IP multicast routing is not suitable for DC networks due to the need to support a massive amount of multicast groups and hosts. State-of-the-art DC multicast routing approaches aim to overcome these scalability issues by, for instance, taking advantage of the symmetry of DC topologies and the programmability of DC switches to compactly encode multicast group information inside packets, thereby reducing the overhead resulting from the need to store the states of flows at the network switches. Although these approaches scale well with the number of multicast groups, they do not perform well with group sizes and, as a result, yield substantial traffic control overhead and network congestion. In this article, we present Bert, a scalable source-initiated DC multicast routing approach that scales well with both the number and size of multicast groups through the clustering of multicast group members where each cluster employs its own forwarding rules. Compared to the state-of-the-art approach, Bert yields much less traffic control overhead by significantly reducing packet header sizes and eliminating switch memory usage across the switches.

LaScaDa: A Novel Scalable Topology for Data Center Network

The growth of cloud-based services is mainly supported by the core networking infrastructures of large-scale data centers, while the scalability of these services is influenced by the performance and dependability characteristics of data centers. Hence, the data center network must be agile and reconfigurable in order to respond quickly to the ever-changing application demands and service requirements. The network must also be able to interconnect the big number of nodes, and provide an efficient and fault-tolerant routing service to upper-layer applications. In response to these challenges, the research community began exploring novel interconnect topologies, namely: Flecube, DCell, Ficonn, HyperFlaNet and BCube. However, these topologies either scale too fast (grows exponentially in size), or too slow, and therefore suffer from performance bottlenecks. In this paper, we propose a novel data center topology called LaScaDa (Layered Scalable Data Center) as a new solution for building scalable and cost-effective data center networking infrastructures. The proposed topology organizes nodes in clusters of similar structure, then interconnect these clusters in a well-crafted pattern and system of coordinates for nodes to reduce the number of redundant connections between clusters, while maximizing connectivity. LaScaDa forwards packets between nodes using a new hierarchical row-based routing algorithm. The algorithm constructs the route to the source based on the modular difference between the source and destination coordinates. Furthermore, the proposed topology interconnects a large number of nodes using a small node degree. This strategy increases the number of directly connected clusters and avoids redundant connections. As a result, we get a good quality of nodes in terms of average path length (APL), bisection bandwidth, and aggregated bottleneck throughput. Experimental results show that LaScaDa has better performance than DCell, BCube, and HyperBcube in terms of scalability, while providing a good quality of service.

Avoiding bottlenecks in networks by short paths

Data center networks are typically characterized by high density communication components that process and exchange large amount of information using shortest paths. Mostly, data centers network topologies contain multi-rooted tree with multiple equal cost shortest paths between pairs of hosts. Usually, data center networks operation is based on the result of shortest path algorithms and per-flow static hashing which may cause poor network utilization rates with some links becoming congested while some parts of the network are underused. This work presents a flow scheduling algorithm that exploits the path diversity in data center topologies and dynamically reroutes large flows through less congested shortest and non-shortest paths based on the current state of the network without causing packets reordering. The algorithm aims to minimize flows latency while maximizing the network utilization rates. Results show that the algorithm proposed in this work reduces flows completion time by 13–24% over Equal-Cost Multi-Path routing, while improving the average network utilization by up to 10%.

Elmo: Source Routed Multicast for Public Clouds

We present Elmo, a system that addresses the multicast scalability problem in multi-tenant datacenters. Modern cloud applications frequently exhibit one-to-many communication patterns and, at the same time, require sub-millisecond latencies and high throughput. IP multicast can achieve these requirements but has control- and data-plane scalability limitations that make it challenging to offer it as a service for hundreds of thousands of tenants , typical of cloud environments. Tenants, therefore, must rely on unicast-based approaches ( e.g. , application-layer or overlay-based) to support multicast in their applications, imposing bandwidth and end-host CPU overheads, with higher and unpredictable latencies. Elmo scales network multicast by taking advantage of emerging programmable switches and the unique characteristics of data-center networks; specifically, the hypervisor switches, symmetric topology, and short paths in a datacenter. Elmo encodes multicast group information inside packets themselves, reducing the need to store the same information in network switches. In a three-tier data-center topology with 27,000 hosts, Elmo supports a million multicast groups using an average packet-header size of 114 bytes (max. 325 bytes), requiring as few as 1,100 multicast group-table entries on average in leaf switches, and having a traffic overhead as low as 5% over ideal multicast.

On the Complexity of Traffic Traces and Implications

This paper presents a systematic approach to identify and quantify the types of structures featured by packet traces in communication networks. Our approach leverages an information-theoretic methodology, based on iterative randomization and compression of the packet trace, which allows us to systematically remove and measure dimensions of structure in the trace. In particular, we introduce the notion of trace complexity which approximates the entropy rate of a packet trace. Considering several real-world traces, we show that trace complexity can provide unique insights into the characteristics of various applications. Based on our approach, we also propose a traffic generator model able to produce a synthetic trace that matches the complexity levels of its corresponding realworld trace. Using a case study in the context of datacenters, we show that insights into the structure of packet traces can lead to improved demand-aware network designs: datacenter topologies that are optimized for specific traffic patterns.