Managing data transfers in computer clusters with orchestra

Job-Deadline-Guarantee-Based Joint Flow Scheduling and Routing Scheme in Data Center Networks

Many emerging Internet of Things (IoT) applications deployed on cloud platforms have strict latency requirements or deadline constraints, and thus meeting the deadlines is crucial to ensure the quality of service for users and the revenue for service providers in these delay-stringent IoT applications. Efficient flow scheduling in data center networks (DCNs) plays a major role in reducing the execution time of jobs and has garnered significant attention in recent years. However, only few studies have attempted to combine job-level flow scheduling and routing to guarantee meeting the deadlines of multi-stage jobs. In this paper, an efficient heuristic joint flow scheduling and routing (JFSR) scheme is proposed. First, targeting maximizing the number of jobs for which the deadlines have been met, we formulate the joint flow scheduling and routing optimization problem for multiple multi-stage jobs. Second, due to its mathematical intractability, this problem is decomposed into two sub-problems: inter-coflow scheduling and intra-coflow scheduling. In the first sub-problem, coflows from different jobs are scheduled according to their relative remaining times; in the second sub-problem, an iterative coflow scheduling and routing (ICSR) algorithm is designed to alternately optimize the routing path and bandwidth allocation for each scheduled coflow. Finally, simulation results demonstrate that the proposed JFSR scheme can significantly increase the number of jobs for which the deadlines have been met in DCNs.

Solid-state 360° optical beamforming for reconfigurable multicast optical wireless communications

Optical wireless communication is an attractive technique for data center interconnects due to its low latency line-of-sight connectivity. Multicast, on the other hand, is an important data center network function that can improve traffic throughput, reduce latency, and make efficient use of network resources. To enable reconfigurable multicast in data center optical wireless networks, we propose a novel 360° optical beamforming scheme based on the principle of superposition of orbital angular momentum modes, emitting beams from the source rack pointing towards any combination of other racks so that connections are established between the source and multiple destination racks. We experimentally demonstrate the scheme using solid state devices for a scenario where racks are arranged in a hexagonal formation in which a source rack can connect with any number of adjacent racks simultaneously, with each link transmitting 70 Gb/s on-off-keying modulations at bit error rates of <10^-6 at 1.5-m and 2.0-m link distances.

BTP: automatic identification and prediction of tasks in data center networks

Modern data centers have widely deployed lots of cluster computing applications such as MapReduce and Spark. Since the coflow/task abstraction can exactly express the requirements of cluster computing applications, various task-based solutions have been proposed to improve application-level performance. However, most of solutions require modification of the applications to obtain task information, making them impractical in many scenarios. In this paper, we propose a Bayesian decision-based Task Prediction mechanism named BTP to identify task and predict the task-size category. First, we design an automatic identification mechanism to identify tasks without manually modifying the applications. Then we leverage bayesian decision to predict the task-size category. Through a series of large-scale NS2 simulations, we demonstrate that BTP can accurately identify task and predict the task-size category. More specifically, BTP achieves 96% precision and 92% recall while obtaining accuracy by up to 98%.

Coded Parallel Transmission for Half-Duplex Distributed Computing

This work studies a general distributed coded computing system based on the MapReduce-type framework, where distributed computing nodes within a half-duplex network wish to compute multiple output functions. We first introduce a definition of communication delay to characterize the time cost during the date shuffle phase, and then propose a novel coding strategy that enables parallel transmission among the computation nodes by delicately designing the data placement, message symbols encoding, data shuffling, and decoding. Compared to the coded distributed computing (CDC) scheme proposed by Li et al., the proposed scheme significantly reduces the communication delay, in particular when the computation load is relatively smaller than the number of computing nodes K. Moreover, the communication delay of CDC is a monotonically increasing function of K, while the communication delay of our scheme decreases as K increases, indicating that the proposed scheme can make better use of the computing resources.

Link Load Correlation-Based Blocking Performance Analysis for Tree-Type Data Center Networks

With the explosive growth of cloud computing applications, the east-west traffic among servers has come to occupy the dominant proportion of the traffic in data center networks (DCNs). Cloud computing tasks need to be executed in a distributed manner on multiple servers, which exchange large amounts of intermediate data between the adjacent stages of each multi-stage task. Therefore, the congestion in DCNs can reduce the processing performance when conducting multi-stage tasks. To address this, the relationship between the blocking performance and the traffic load can be adopted as a theoretical basis for network planning and traffic engineering. In this paper, the traffic load correlation between edge links and aggregation links is considered, and an iterative blocking performance analysis method is proposed for two-layer tree-type DCNs. The simulation results show the good accuracy of the proposed method with respect to the theoretical results especially in the blocking rate range below 4% and with over-subscription ratio 1.5.

Optical multicast system for data center networks

We present the design and experimental evaluation of an Optical Multicast System for Data Center Networks, a hardware-software system architecture that uniquely integrates passive optical splitters in a hybrid network architecture for faster and simpler delivery of multicast traffic flows. An application-driven control plane manages the integrated optical and electronic switched traffic routing in the data plane layer. The control plane includes a resource allocation algorithm to optimally assign optical splitters to the flows. The hardware architecture is built on a hybrid network with both Electronic Packet Switching (EPS) and Optical Circuit Switching (OCS) networks to aggregate Top-of-Rack switches. The OCS is also the connectivity substrate of splitters to the optical network. The optical multicast system implementation requires only commodity optical components. We built a prototype and developed a simulation environment to evaluate the performance of the system for bulk multicasting. Experimental and numerical results show simultaneous delivery of multicast flows to all receivers with steady throughput. Compared to IP multicast that is the electronic counterpart, optical multicast performs with less protocol complexity and reduced energy consumption. Compared to peer-to-peer multicast methods, it achieves at minimum an order of magnitude higher throughput for flows under 250 MB with significantly less connection overheads. Furthermore, for delivering 20 TB of data containing only 15% multicast flows, it reduces the total delivery energy consumption by 50% and improves latency by 55% compared to a data center with a sole non-blocking EPS network.

Rethinking the physical layer of data center networks of the next decade

Not only do big data applications impose heavy bandwidth demands, they also have diverse communication patterns denoted as *-cast) that mix together unicast, multicast, incast, and all-to-all-cast. Effectively supporting such traffic demands remains an open problem in data center networking. We propose an unconventional approach that leverages physical layer photonic technologies to build custom communication devices for accelerating each *-cast pattern, and integrates such devices into an application-driven, dynamically configurable photonics accelerated data center network. We present preliminary results from a multicast case study to highlight the potential benefits of this approach.