Wireless link scheduling for data center networks

To solve the bandwidth and latency issues in current hierarchical data center network (DCN) architectures based on electrical switches, we propose a novel hybrid DCN architecture based on distributed flow-controlled fast optical switches (FOS) and modified top-of-the-rack (TOR) switches (HiFOST). The intra-cluster interconnection of HiFOST is built by FOS with wavelength switching in nanoseconds’ time for an efficient statistical multiplexing operation, while the inter-cluster interconnection is connected by the TOR interfaces directly. Due to the lack of practical optical buffers, optical flow control is implemented to retransmit packets in case of contention. We investigate the performance of HiFOST DCN with different TOR buffer sizes, optical link capacities, elastic allocation of transceivers, and network scales under realistic data center (DC) traffic. The results show an average server-to-server latency of less than 2.8 μs, a packet loss <5.6×10−6 at load of 0.5 for a DC size of 94,080 servers with limited 50 KB TOR buffer. In addition, for scaling out the servers’ number and scaling up the data rate of connected servers, the cost and power consumption of the HiFOST DCN have been investigated and compared with the electrical Fat-Tree and Leaf-Spine DCN architectures, as well as with the optical H-LION and OPSquare DCN architectures. Results indicate that, for 94,080 servers operating at 10 Gb/s, HiFOST has a 48.2% and 34.1% savings of the cost and 46.3% and 32.5% savings of the power consumption with respect to the Fat-Tree and Leaf-Spine, respectively. For a HiFOST DCN supporting a 10880 server, scaling up the operating data rate of the server to 100 Gb/s, the HiFOST solution has a cost savings of 35.6% and 34.1% and power consumption of 56.5% and 59.2% as compared to the Fat-Tree and Leaf-Spine, respectively.

We study the problem of flow scheduling in data center networks. Using speed scaling, our aim is to find an online scheduling algorithm that minimizes the total energy consumption of the network by determining both the transmission order and rates of the arriving flows while providing a strict flow deadline guarantee. Observing the superlinear property of link power consumption, the key challenge is in constantly determining the minimum transmission rate for “delay-tolerable” flows without any priori knowledge. To leverage the flow arrival pattern, we propose a probability-based flow prediction model to capture the uncertainty of the network flows. Based on the prediction model, we propose a tunable online flow scheduling algorithm to solve the online flow scheduling problem effectively. By introducing a scaling factor on bandwidth allocation, this algorithm allows us to conduct arbitrary trade-offs between the conservative and aggressive behaviors in terms of energy conser- vation. The effectiveness of the proposed algorithm is validated through rigorous theoretical analysis and further confirmed by extensive numerical simulations.

Numerous scheduling approaches have been proposed to improve user experiences in a data center network (DCN) by reducing flow completion time (FCT). Mimicking the shortest job first (SJF) has been proved to be the prominent way to improve FCT. To do so, some approaches require flow size or completion time information in advance, which is not possible in scenarios like HTTP chunk transfer or database query response. Some information-agnostic schemes require involving end-hosts for counting the number of bytes sent. We present Longer Stay Less Priority (LSLP), an information-agnostic flow scheduling scheme, like Multi-Level Feedback Queue (MLFQ) scheduler in operating systems, that aims to mimic SJF using P4 switches in a DCN. LSLP considers all the flows as short flows initially and assigns them to the highest priority queue, and flows get demoted to the lower priority queues over time. LSLP estimates the active time of a flow by leveraging the state-of-the-art P4 switch’s programmable nature. LSLP estimates the active time of a group of new flows that arrive during a time interval and assigns their packets to the highest priority. At the beginning of the next time interval, arriving packets of old flows are placed one priority lower except for those already in the lowest priority queue. Therefore, short flows can be completed in the few higher priority queues while long flows are demoted to lower priority queues. We have evaluated LSLP via a series of tests and shown that its performance is comparable to the existing scheduling schemes.

With the development of cloud computing, data center network (DCN) architectures as the core of the cloud platform received a surge of interesting from both the industry and academia. However, assessments of those new DCN architectures are mainly concentrated in load balancing, improvement of architectures and as well as some research of performance analysis in visualized environment. Moreover, none of them focus on the security in DCN architectures. In this paper, we propose contribution matrix method based on information entropy theory from the point of view of node importance which is the basic of the topology vulnerability. In addition, we conduct an experimental evaluation of the state-of-the-art Multi-rooted Tree and FiConn architectures, each respectively as a representative of the switch-centric network architecture and server-centric network architecture. Firstly, we use an undirected graph theory to describe the architecture. Secondly, we obtain the associated matrix of betweenness and degree, and calculate their weights by information entropy theory. And then, according to the definition 3 (see Section III-B); we obtain the contribution matrix for each one. Last but not least, we get the value of node importance for each node of normalized. Compared with other methods, our proposed method is effective and has a much higher accuracy. Furthermore, our method is generic and can be widely used for the new DCN architectures.

The advent and continued growth of large data centers has led to much interest in switch architectures that can economically meet the high capacities needed for interconnecting the thousands of servers in these data centers. Various multilayer architectures employing thousands of switches have been proposed in the literature. We make use of the observation that the traffic in a data center is a mixture of relatively static and rapidly fluctuating components, and develop a combined scheduler for both these components using a generalization of the load-balanced scheduler. The presence of the known static component introduces asymmetries in the ingress-egress capacities, which preclude the use of a load-balanced scheduler as is. We generalize the load-balanced scheduler and also incorporate an opportunistic scheduler that sends traffic on a direct path when feasible to enhance the overall switch throughput. Our evaluations show that this scheduler works very well despite avoiding the use of a central scheduler for making packet-by-packet scheduling decisions.

Traffic in today’s data center networks grows rapidly. Traditional routing and forwarding protocols in data center usually rely on static hashing, and may cause avoidable bottlenecks. We present a genetic algorithm to make better use of the resources. Our goal is to minimize the maximum single link utilization of the network. And results show that our algorithm performs better than a recent greedy algorithm.

Online social networks (OSN) have become a prevalent communication tool with a user scale of billions. Users send messages and share photos and videos to their friends every day. As a consequence, a large amount of flows are produced by file storage and backup in the back end. Most flows are deadline intensive, and the scale of flow account is large; these characteristics create new challenges for flow scheduling in data center networks. In this study, we conclude the flow distribution of OSN based on previous studies. OSN flow can be classified into a large amount of bursty flow and a limited amount of large flow. The bursty mice flows are mostly critical real-time requirements, which support instant messages and routing information; the large flows seldom require real-time transmission, because they mainly support file transferring and updating. A mixed flow causes bottlenecks in rack switches easily, thereby resulting in deadline loss for real-time applications and lower throughput. The study divides flow based on different metrics and provides priority-based flow scheduling for OSN datacenter (PFO). PFO allocates rate for different flows on the basis of flow size and deadline information. Our objective is to decrease the average completion time for bursty flows and ensure a high throughput. The analysis and evaluation shows that the proposed scheduling increases the goodput more than DCTCP and D2TCP, and decreases the fraction of missed deadlines.

The Visible Light Communication (VLC) has the potential to provide dense and fast connectivity at low cost. In this paper, we propose WSWDC, a novel VLC enabled wireless small-world data center. It employs VLC links to achieve a fully wireless data center network (DCN) across racks for the first time. The using of VLC links eliminates hierarchical switches and inter-rack cables, and thus reducing hardware investment, as well as maintenance cost. More precisely, to simplify the configuration and control operations, we propose three DCN design rationales: (1) fully-wireless, all inter-rack links are wireless; (2) easy-deployable, it is not necessary to change the existing infrastructure inside data center; (3) plug-and-play, no extra centralized control operations are required. Previous proposals, however, cannot achieve the three rationales simultaneously. To this end, we first use regular VLC links to interconnect racks as a regular grid DCN. To further exploiting the benefits of VLC links, a few random VLC links are carefully introduced to update the wireless grid DCN as a wireless small-world DCN. To avoid the potential interference among VLC links, we deploy VLC transceivers at different height on the top of each rack. In this way, VLC links would not interfere with others at each height level. Moreover, we design a greedy but efficient routing method for any pair of racks using their identifiers as inputs. Comprehensive evaluation results indicate that our WSWDC exhibits good network performance.

In recent years, data center network (DCN) architectures (e.g., DCell [5], FiConn [6], BCube [4], FatTree [1], and VL2 [2]) received a surge of interest from both the industry and academia. However, none of existing efforts provide an in-depth understanding of the impact of these architectures on application performance in practical multi-tier systems under realistic workload. Moreover, it is also unclear how these architectures are affected in virtualized environments. In this paper, we fill this void by conducting an experimental evaluation of FiConn and FatTree, each respectively as a representative of hierarchical and flat architectures, in a three-tier transaction system using virtual machine (VM) based implementation. We observe several fundamental characteristics that are embedded in both classes of network topologies and cast a new light on the implication of virtualization in DCN architectures. Issues observed in this paper are generic and should be properly addressed by any DCN architectures before being considered for actual deployment, especially in mission-critical real-time transaction systems.

Understanding data center network architectures in virtualized environments: A view from multi-tier applications

As the volume of data keeps growing rapidly, more and more storage devices, servers and network devices are continuously added into data centers to store, manage and analyze the data. The industry experience indicates that, instead of a huge number of servers added at a time, the data center network also expands gradually by adding a small number of servers from time to time. As a result, how to achieve an incremental scalability is becoming a very important challenge in designing modern data center network architectures in order to maintain the topological properties unchanged when the size of data centers grows. In this paper, we propose a new type of data center network architecture named HSDC (High Scalability Data Center Network Architecture) based on the hypercube network. The HSDC is constructed by using $m$m-port switches and 2-port servers. The fault-tolerant routing algorithm designed in this paper for HSDC can be executed on any vertex and is able to construct a path between any pair of vertices. In order to achieve an incremental scalability, we further propose three types of incomplete HSDC structures that allow gradually adding servers into the structures, while maintaining all the topological properties. The simulation experiments and performance results demonstrate that the throughput of HSDC is comparable to that of Fat-Tree, BCube and DCell. Furthermore, the analysis results indicate that HSDC strikes a good balance among diameter, bisection width, incremental scalability, cost and energy consumption in contrast to the state-of-the-art data center network architectures.

We investigate the performance of a novel flat data center network (DCN) architecture based on distributed buffer-less optical switches and fast flow control. Scalability of the novel DCN is assessed by numerical simulation of medium and large size data centers reporting an end-to-end latency of less than 8μs (including retransmission), a packet loss 9Tb/s average network throughput under load of 0.4.

Many data center applications perform rich and complex tasks (e.g., executing a search query or generating a user's news-feed). From a network perspective, these tasks typically comprise multiple flows, which traverse different parts of the network at potentially different times. Most network resource allocation schemes, however, treat all these flows in isolation -- rather than as part of a task -- and therefore only optimize flow-level metrics. In this paper, we show that task-aware network scheduling, which groups flows of a task and schedules them together, can reduce both the average as well as tail completion time for typical data center applications. To achieve these benefits in practice, we design and implement Baraat, a decentralized task-aware scheduling system. Baraat schedules tasks in a FIFO order but avoids head-of-line blocking by dynamically changing the level of multiplexing in the network. Through experiments with Memcached on a small testbed and large-scale simulations, we show that Baraat outperforms state-of-the-art decentralized schemes (e.g., pFabric) as well as centralized schedulers (e.g., Orchestra) for a wide range of workloads (e.g., search, analytics, etc).

The use of wireless links can provide flexibility and reconfigurability to data center networks. Recent investigations show that 60GHz millimeter-wave technology can be used in wireless data center networks. The steerable millimeter-wave transceiver works as a multi-port switch where point-to-point links are established using beam-forming. Considering an interference model, specific to the wireless data center, we propose a new algorithm for k-pod fat-tree data center topology mapping. Further, we propose to group all links of the network into sets of mutually concurrent links. By assigning a time slot for all concurrent links in a set, we devise a TDMA scheduling algorithm for the wireless network. We show that the concurrent wireless transmissions could result in O(k2) number of slots that is k times better than graph based coloring approach. By increasing the concurrency, we argue that the use of 60GHz technology is more viable in building completely wireless data centers.

All optical or hybrid electrical/optical switching data center network (DCN) architectures are capable of adapting the optical-layer topology to various network traffic demands based on the flexibility of optical switching technology. The topology reconfiguration strategies in previous research simply attempt to reconfigure optical circuit switching (OCS) links to offload heavier rack-to-rack traffic. However, due to the complex multi-layer interconnections and the various network traffic patterns in DCNs, the OCS links will not only offload the network traffic between its source and destination racks, but also forward the network traffic that belongs to the adjacent racks of the source and destinations racks. Previous strategies ignore the forwarding traffic and result in sub-optimal solutions. In order to overcome the shortcomings of previous methods and make full use of the OCS link resources in DCNs, an effective topology reconfiguration strategy named optical circuit bridge (OCBridge) has been proposed in our recent work. The main idea of our OCBridge is to take into account the traffic demands between the source and destination regions of each OCS link. In this paper, we first review our OCBridge, and then in simulation, we evaluate the performance of our OCBridge by testing it in two typical OCS-based DCNs, including OpenScale and Helios. Under our OCBridge, we analyze the outstanding network throughput performance of the DCNs and investigate the communication distance of traffic demands. Finally, in experiment, we deploy our OCBridge to a small prototype of OpenScale and verify that our OCBridge is capable of accelerating the distributed computing applications compared with previous topology reconfiguration strategies.