Maotong Xu

PODCA: A passive optical data center architecture

Optical interconnects for data centers can offer reduced power consumption, low latency, and high scalability, compared to electrical interconnects. However, active optical components such as tunable wavelength converters and Micro-Electro-Mechanical Systems (MEMS) switches suffer from high cost or slow reconfiguration times. In this paper, we propose three different Passive Optical Data Center Architectures (PODCAs), depending on the size of the network. Our key device is the Arrayed Waveguide Grating Router (AWGR), a passive device that can achieve contention resolution in the wavelength domain [1]. In our architectures, optical signals are transmitted from fast tunable transmitters and pass through couplers, AWGR, demultiplexers, and are received by wide-band receivers. Our architecture can easily accommodate over 2 million servers. Simulation results show that packet latency is below 9μs, and 100% throughput is achievable. We compare the power consumption and capital expenditure (CapEx) cost of PODCA with other recent optical data center network architectures such as DOS, Proteus, and Petabit. Results show that our architectures can save up to 90% on power consumption and 88% on CapEx.

Optimizing Speculative Execution of Deadline-Sensitive Jobs in Cloud

In this paper, we bring various speculative scheduling strategies together under a unifying optimization framework, which defines a new metric, Probability of Completion before Deadlines (PoCD), to measure the probability that MapReduce jobs meet their desired deadlines. We propose an optimization problem to jointly optimize PoCD and execution cost in different strategies. Three strategies are prototyped on Hadoop MapReduce and evaluated against two baseline strategies using experiments. A 78% net utility increase with up to 94% PoCD and 12% cost improvement is achieved.

LASER: A Deep Learning Approach for Speculative Execution and Replication of Deadline-Critical Jobs in Cloud

Meeting desired application deadlines is crucial as the nature of cloud applications is becoming increasingly mission-critical and deadline-sensitive. Empirical studies on large-scale clusters reveal that a few slow tasks, known as stragglers, could significantly stretch job execution times. A number of strategies are proposed to mitigate stragglers by launching speculative or clone (task) attempts. These strategies often rely on a model-based approach to optimize key operating parameters and are prone to inaccuracy/incompleteness in the underlying models. In this paper, we present LASER, a deep learning approach for speculative execution and replication of deadline-critical jobs. Machine learning has been successfully used to solve a large variety of classification and prediction problems. In particular, the deep neural network (DNN), consisting of multiple hidden layers of units between input and output layers, can provide more accurate regression (prediction) than traditional machine learning algorithms. We compare LASER with SRQuant, a speculative- resume strategy that is based on quantitative analysis. Both these scheduling algorithms aim to improve Probability of Completion before Deadlines (PoCD), i.e., the probability that MapReduce jobs meet their desired deadlines, and reduce the cost of speculative execution, measured by the total (virtual) machine time. We evaluate and compare the two strategies through testbed experiments. The results show that our two strategies outperform Hadoop without speculation (Hadoop-NS) and Hadoop with speculation (Hadoop-S) by up to 89% in PoCD and 13% in cost.

Joint banding-node placement and resource allocation for multigranular elastic optical networks

The fine-grained grid of elastic optical networks (EONs) facilitates flexible bandwidth allocation and increased spectrum utilization efficiency and is seen as a promising solution to handle ever-increasing traffic demands. Despite this, fiber capacity exhaustion is imminent, and multifiber links are expected to be prevalent in future optical networks. A challenge this brings about is the high port count of optical cross-connects (OXCs). Conventional OXCs built using flex-grid wavelength selective switches do not scale well. To achieve scalability of OXCs, a flexible wavebanding OXC architecture (FLEX) has been proposed recently. FLEX reduces the complexity and cost of OXCs while sacrificing some performance in terms of limited switching flexibility. Taking the reduced switching capability into consideration, a cost-function pluggable auxiliary layered-graph framework has been proposed in our previous work to solve the routing, fiber, waveband, and spectrum assignment (RFBSA) problem in multifiber-based EONs with flexible wavebanding nodes. In this paper, we address the following problem. Given the total number of available WSSs for the network as a budget, we determine how many FLEX nodes to deploy and where to deploy them, and solve the RFBSA problem jointly to optimize the network performance. An integer linear programming formulation is proposed for a set of traffic requests. We also propose a heuristic algorithm to solve this joint problem efficiently. The results show that our algorithm achieves good network performance, which is indicated by the average maximum spectrum usage as well as considerably reducing hardware costs. We also evaluate our algorithm for dynamically arriving traffic requests in terms of demand blocking ratio.

Routing, fiber, band, and spectrum assignment (RFBSA) for multi-granular elastic optical networks

The dramatic growth of Internet traffic brings challenges for optical network designers. There have been a number of advances recently in increasing fiber bandwidth and spectrum utilization efficiency through elastic optical networking (EON). In EON, a flexible and more fine-grained grid than conventional approaches is employed, and this allows allocated fiber bandwidth to better match traffic demands. Despite these advances, imminent fiber capacity exhaustion means that multiple fibers per link will be inevitable. In an effort to reduce the complexity of optical crossconnects, a flexible wavebanding crossconnect has been proposed recently. Elastic networking and flexible wavebanding introduce a new problem, namely, the routing, fiber, waveband, and spectrum assignment (RFBSA) problem. In this work, we propose new cost functions that are pluggable into an auxiliary layered-graph framework to solve the RFBSA problem with different objectives. We focus on minimizing the maximum spectrum usage for a set of traffic demands, and show that our approach outperforms traditional approaches.

Joint Banding-Node Placement and Resource Allocation for Multi-Granular Elastic Optical Networks

The fine-grained grid of elastic optical networks (EONs) facilitates flexible bandwidth allocation and increased spectrum utilization efficiency and is seen as a promising solution to handle ever-increasing traffic demands. Despite this, fiber capacity exhaustion is imminent, and multifiber links are expected to be prevalent in future optical networks. A challenge this brings about is the high port count of optical cross-connects (OXCs). Conventional OXCs built using flex-grid wavelength selective switches do not scale well. To achieve scalability of OXCs, a flexible wavebanding OXC architecture (FLEX) has been proposed recently. FLEX reduces the complexity and cost of OXCs while sacrificing some performance in terms of limited switching flexibility. Taking the reduced switching capability into consideration, a cost-function pluggable auxiliary layered-graph framework has been proposed in our previous work to solve the routing, fiber, waveband, and spectrum assignment (RFBSA) problem in multifiber-based EONs with flexible wavebanding nodes. In this paper, we address the following problem. Given the total number of available WSSs for the network as a budget, we determine how many FLEX nodes to deploy and where to deploy them, and solve the RFBSA problem jointly to optimize the network performance. An integer linear programming formulation is proposed for a set of traffic requests. We also propose a heuristic algorithm to solve this joint problem efficiently. The results show that our algorithm achieves good network performance, which is indicated by the average maximum spectrum usage as well as considerably reducing hardware costs. We also evaluate our algorithm for dynamically arriving traffic requests in terms of demand blocking ratio.

CRED: Cloud Right-Sizing with Execution Deadlines and Data Locality

As demands for cloud-based data processing continue to grow, cloud providers seek effective techniques that deliver value to the businesses without violating Service Level Agreements (SLAs). Cloud right-sizing has emerged as a very promising technique for making cloud services more cost-effective. In this paper, we present CRED, a novel framework for cloud right-sizing with execution deadlines and data locality constraints. CRED jointly optimizes data placement and task scheduling in data centers with the aim of minimizing the number of nodes needed while meeting users’ SLA requirements. We formulate CRED as an integer optimization problem and present a heuristic algorithm with provable performance guarantees to solve the problem. Competitive ratios of the proposed algorithm are quantified in closed form for arbitrary task parameters and cloud configurations. We also extend our work to obtain a resilient solution, which allows successful recovery at run time from any single node failure and is guaranteed to meet both deadline and locality constraints. Simulation results using Google trace show that our proposed algorithm significantly outperforms existing heuristics such as first-fit by reducing the number of required active servers by up to 47 percent, and achieves near-optimal performance. We also show that our algorithm can significantly improve utilization of both computational resources and storage space by up to 28 and 15 percent, respectively.

Evaluation and performance modeling of two OXC architectures

This paper presents an evaluation of two Optical Cross-Connect (OXC) node architectures with multiple fibers per link — one, a conventional architecture, and the second, a hierarchical architecture that has lower complexity than the first architecture. Analytical models for computing the blocking probability of connection requests are proposed and validated. Heuristics for resource assignment are then proposed, and the performance of the two architectures are compared. The two architectures are then compared in terms of the cost of the OXC node and the power consumption. Our results show that the hierarchical architecture exhibits a good balance between performance, cost, and power consumption.

A Reconfigurable High-Performance Optical Data Center Architecture

Abstract-Optical data center network architectures are be- coming attractive because of their low energy consumption, large bandwidth, and low cabling complexity. In [1], an AWGR-based passive optical data center architecture (PODCA) is presented. Compared with other optical data center architectures, e.g., DOS [2], Proteus [3], and Petabit [4], PODCA can save up to 90% on power consumption and 88% in cost. Also, average latency can be low as 9 μs at close to 100% throughput. However, PODCA is not reconfigurable and cannot optimize the network topology to dynamic traffic. In this paper, we present a novel, scalable and flexible recon- figurable architecture called RODCA. RODCA is built on and augments PODCA with a flexible localized intra-cluster optical network. With the reconfigurable intra-cluster network, racks with mutually large traffic can be located within the same cluster, and share the large bandwidth of the intra-cluster network. We present an algorithm for DCN topology reconfiguration, and present simulation results to demonstrate the effectiveness of reconfiguration.