Xiangming Dai

SDN-based multi-class QoS-guaranteed inter-data center traffic management

When allocating network bandwidth to multiple classes of applications in inter-data center communication, coordination always yields a better utilization of the backbone network. Yet, it often comes at a prohibitively heavy computational and communication cost, making it thus far not a practically viable approach. SDN helped in bridging the communication cost gap by enabling centralized control, and SDN has been recently applied in such inter-DC traffic management. However, the computational cost is still an issue as the efficient and fast response of the centralized traffic engineering algorithm has become crucial to the practicality of such SDN-based approach. In this paper, we present MCTEQ, a utility-optimization-based joint-bandwidth allocation for inter-DC communication with multiple traffic classes, that handles priorities between traffic classes in a soft manner and explicitly considers the delay requirement of Interactive flows. MCTEQ being NP-hard, we apply approximation techniques to lean on the mature and efficient LP solver and obtain fast and accurate approximations. We demonstrate via experiments with Googles inter-DC backbone topology that MCTEQ achieves about 160 Gbps higher network utilization than the existing SWAN solution, yet runs 2.5 times faster. In particular, MCTEQ guarantees that the allocated bandwidth for Interactive flows strictly meets their end-to-end delay requirements.

Energy-efficient planning of QoS-constrained virtual-cluster embedding in data centres

It is crucial for cloud service providers to fulfil their tenants targeted quality of service (QoS) requirements, yet, providing such QoS in an energy efficient manner is of paramount importance to optimizing operational costs. In this paper, we study the energy efficiency of virtual cluster embedding in data centers. More specifically, we focus on designing fast algorithms to place in an energy efficient manner virtual clusters, including virtual machines, virtual switches and virtual topology with bandwidth requirements. We first model the problem as an integer programming problem, and prove its NP hardness, then based on a relaxed version of this model, we propose an approximate algorithm to solve it efficiently.

Iterative greedy algorithms for energy efficient LTE small cell networks

The increasing demand for high bandwidth in cellular wireless networks led to a dramatic increase in the number of deployed LTE femtocell base stations (FBS) in recent years. While the energy consumed by a single FBS is relatively negligible, the collective energy footprint of all FBSs deployed by an operator turns out to be huge. Energy-efficient protocols are needed thus to balance the trade-off between energy saving and bandwidth utilization. The key component of the desired protocol to prevent both energy and bandwidth waste is interference mitigation, which, most previous work has failed to properly consider. To this end, in this paper, we manipulate user equipment (UE) association and OFDMA scheduling with a combination of interference mitigation to enable an energy efficient QoS-constrained medium access for femtocell networks. Recognizing the NP-hardness of the problem, we propose two algorithms with guaranteed convergence. Extensive simulations show that our algorithms outperform alternatives in multiple metrics such as utility, power consumption, and convergence speed.

Scheduling for response time in Hadoop MapReduce

In this paper, we propose a novel task scheduling algorithm for Hadoop MapReduce called dynamic priority multiqueue scheduler (DPMQS). DPMQS i) increases the data locality of jobs, and, ii) dynamically increases the priority of jobs that are near to completing their Map phase, to bridge the time gap between the start of the reduce tasks and the execution of the reduce function for these jobs. We discuss the details of DPMQS and its practical implementation, then assess its performance in a small physical cluster and large-scale simulated clusters and compare it to the other schedulers available in Hadoop. Both real experiments and simulation results show that DPMQS decreases significantly the response time, and demonstrate that DPMQS is insensitive to changes in the cluster geometry.

Content peering in content centric networks

The settlement-free peering relationships play a vital role in todays Internet, notably in helping ISPs cope with the dramatic increase in traffic load caused by the recent surge in the demand for videos and user generated content. Because of the added caching capability of CCN, peering in CCN can be expanded to encompass not only content that is permanently stored in the ISPs network, but also content that is temporarily cached at CCN routers in the ISPs network. Intuitively, this content-level peering is likely to benefit the networks involved in the peering relationship; however, this comes at a non-trivial cost. In this paper, we try to answer the question of whether the additional overhead and complexity to extend content-level peering is justifiable in view of the benefits it brings. To this end, we formulate the content-level peering problem as an optimization problem to study its maximum potential benefit. We conduct extensive numerical experiments to evaluate the potential peering benefit under realistic AS-level peering graphs, using realistic video traffic. The experimental results show that the interconnectivity of the peering graph significantly affects the maximum benefit of content-level peering. Compared to local greedy caching, cooperative caching can bring higher peering benefits; yet it is sensitive to parameters like the peering link bandwidth and the AS-level cache size.

Energy Efficient Virtual Cluster Embedding in Public Data Centers

To improve their profits, it is crucial for cloud service providers to fulfil their tenants target QoS requirements in an energy efficient way. The services provided in many modern public clouds have graduated to include more sophisticated and flexible virtual clusters, encompassing not only the resource requirement of the virtual machines (VMs) but also the virtual topology between VMs. In this paper, we study the energy efficiency of virtual cluster embedding in data centers. We carefully develop a mathematical model, aiming to minimize the energy consumption of a data center, and given the NP- hardness of the problem, we propose an approximate algorithm MinE-VCE to solve the problem efficiently. We show via numerical experiments how MinE-VCE consistently outperforms other alternative algorithms.

A novel decentralized asynchronous scheduler for Hadoop

In cloud computing systems, such as Hadoop, system performance is a significant target for improvement. In classic master node-central schedulers, decision is made in the heartbeat time scale, and idle slots during a heartbeat, remain idle until allocated a task by the master node. In this paper, we propose a novel scheduler named multiple queues scheduler (MQS) that improves the throughput of the system by increasing data locality rate of map tasks, reducing thereby the average completion time of jobs. To achieve this, we associate slave nodes with individual queues, and distribute the tasks of a job at arrival to those nodes that contain the associated input data, based on data locality. To reduce the load on overloaded slave nodes, task migration is performed asynchronously between nodes within a rack, without the intervention of the master node. Our results demonstrate the effectiveness of the proposed algorithm. The benefits of MQS are three-fold: first, it decreases the probability of allocating map tasks to non data-local nodes; second, it decreases the time wasted between heartbeats; these two aspects immediately improve the system performance; and third, it mitigates the stress on the master node by assigning part of the schedulers functions to slave nodes.

Potential Game based Energy Efficient Resource Allocation in HeNB Networks

Powering an individual LTE femtocell base station, or Home eNodeB (HeNB) in the LTE jargon, requires relatively very little energy. It is only when HeNBs are deployed massively, as has happened in the past few years, that energy efficiency becomes an important issue. With large numbers of co-located HeNBs and the increased inter-cell interference resource utilization becomes highly inefficient resulting in a high unnecessary energy consumption. To tackle this problem, coordination techniques could be invoked by the network operator to consolidate the offered workload to the network on as few HeNBs as possible and power down idle ones. Recognizing, however, that such techniques usually impair user-perceived quality of service (QoS), especially with bursty traffic, more sophisticated methods need to be investigated to also consider QoS. Despite the volume of prior work, the key issue -- viz. simultaneously reducing energy waste, increasing bandwidth utilization while guaranteeing user-perceived QoS -- has not been properly considered so far. In this paper, we model the trade-off between energy efficiency and QoS preservation by manipulating user equipment (UE) association and OFDMA scheduling in controlled networks of HeNBs. The problem being NP-hard, we propose two distributed learning algorithms, within a potential game-based framework, to obtain good and fast solutions to the problem. We demonstrate via numerical results the effectiveness of the proposed algorithms in achieving better performance in terms of utility, power, energy efficiency, convergence speed, and complexity compared to other alternatives.