Design and Analysis of Stochastic Processing and Matching Networks

We construct a generic, simple, and efficient scheduling policy for stochastic processing networks, and provide a general framework to establish its stability. Our policy is randomized and prioritized: with high probability it prioritizes jobs that have been least routed through the network. We show that the network is globally stable under this policy if there exists an appropriate quadratic local Lyapunov function that provides a negative drift with respect to nominal loads at servers. Applying this generic framework, we obtain stability results for our policy in many important examples of stochastic processing networks: open multiclass queueing networks, parallel server networks, networks of input-queued switches, and a variety of wireless network models with interference constraints. Our main novelty is the construction of an appropriate global Lyapunov function from quadratic local Lyapunov functions, which we believe to be of broader interest.

We consider a class of stochastic processing networks. Assume that the networks satisfy a complete resource pooling condition. We prove that each maximum pressure policy asymptotically minimizes the workload process in a stochastic processing network in heavy traffic. We also show that, under each quadratic holding cost structure, there is a maximum pressure policy that asymptotically minimizes the holding cost. A key to the optimality proofs is to prove a state space collapse result and a heavy traffic limit theorem for the network processes under a maximum pressure policy. We extend a framework of Bramson [Queueing Systems Theory Appl. 30 (1998) 89--148] and Williams [Queueing Systems Theory Appl. 30 (1998b) 5--25] from the multiclass queueing network setting to the stochastic processing network setting to prove the state space collapse result and the heavy traffic limit theorem. The extension can be adapted to other studies of stochastic processing networks.

Complex systems like semiconductor wafer fabrication facilities (fabs), networks of data switches, and large-scale call centers all demand efficient resource allocation. Deterministic models like linear programs (LP) have been used for capacity planning at both the design and expansion stages of such systems. LP-based planning is critical in setting a medium range or long-term goal for many systems, but it does not translate into a day-to-day operational policy that must deal with discreteness of jobs and the randomness of the processing environment. A stochastic processing network, advanced by J. Michael Harrison (2000, 2002, 2003), is a system that takes inputs of materials of various kinds and uses various processing resources to produce outputs of materials of various kinds. Such a network provides a powerful abstraction of a wide range of real-world systems. It provides high-fidelity stochastic models in diverse economic sectors including manufacturing, service, and information technology. We propose a family of maximum pressure service policies for dynamically allocating service capacities in a stochastic processing network. Under a mild assumption on network structure, we prove that a network operating under a maximum pressure policy achieves maximum throughput predicted by LPs. These policies are semilocal in the sense that each server makes its decision based on the buffer content in its serviceable buffers and their immediately downstream buffers. In particular, their implementation does not use arrival rate information, which is difficult to collect in many applications. We also identify a class of networks for which the nonpreemptive, non-processor-splitting version of a maximum pressure policy is still throughput optimal. Applications to queueing networks with alternate routes and networks of data switches are presented.

An algorithm for network and data-aware placement of multi-tier applications in cloud data centers

Stochastic Processing Networks (SPNs) model manufacturing, communication, and service systems. In such a network, service activities require parts and resources to produce other parts. Because service activities compete for resources, a scheduling problem arises. This paper proposes a deficit maximum weight (DMW) algorithm to achieve throughput optimality and maximize the net utility of the production. It overcomes the instability problem of Maximum-Weight Scheduling in SPNs.

We propose a general framework, dubbed Stochastic Processing under Imperfect Information (SPII), to study the impact of information constraints and memories on dynamic resource allocation. The framework involves a Stochastic Processing Network (SPN) scheduling problem in which the decision maker may access the system state only through a noisy channel, and resource allocation decisions must be carried out through the interaction between an encoding policy (who observes the state) and allocation policy (who chooses the allocation). Applications in the management of large-scale data centers and human-in-the-loop service systems are among our chief motivations. We quantify the degree to which information constraints reduce the size of the capacity region in general SPNs, and how such reduction depends on the amount of memories available. Using a novel metric, capacity factor, our main theorem characterizes the reduction in capacity region (under "optimal" policies) for all nondegenerate channels, and across almost all combinations of memory sizes. Notably, the theorem demonstrates, in substantial generality, that (1) the presence of a noisy channel always reduces capacity, (2) more memories for the allocation policy always improve capacity, and (3) more memories for the encoding policy have little to no effect on capacity. Finally, all of our positive (achievability) results are established through constructive, implementable policies. Our proof program involves the development of a host of new techniques by combining ideas from information theory, learning and queueing theory. We create a simple yet powerful generalization of the Max-Weight policy, in which individual Markov chains are selected dynamically, in a manner analogous to how schedules are used in a conventional Max-Weight policy.

With the fast development of information technology (IT) industry, the demand and market volume for off-line power supplies keep increasing, especially those for telecommunication, computer servers and data centers. As the capital expenditure was measured by the square footage occupied rather than power consumption, the development of high power density converter system is of greater interesting. The rising energy prices have resulted in the cost of power and cooling exceeding the purchase cost in less than two years. Therefore, highly efficient power conversion is required for the power converter system. Generally, the power supply unit (PSU) for power distribution system (PDS) in data center and telecom are the standard two-stage approach which normally consists of power factor correction (PFC) circuit and isolated DC-DC converter. The two-stage power conversion has demonstrated excellent performance and high reliability, since the design can be optimized for each stage. However, limitations to prevent the existing two-stage PSU to fulfill future requirements for the PDS in data center and telecom applications are revealed, and a very promising and fundamentally different approach with the single-stage isolated converter is proposed in this dissertation. The development of single-stage converters with the option of placing the energy storage outside of the PSU creates new degrees of freedom regarding e.g. simplified rectifier racks in telecom and data center. This provides tangible benefits in the form of space saving, better airflow for power unit in rectifier racks and improved lifespan. The three-phase isolated buck matrix-type rectifier, capable of achieving high power density and high efficiency, is identified as an excellent candidate for the medium power level (5 kW~10 kW) single-stage power supply design. Nevertheless, the matrix-type rectifiers are known for their relatively complex modulation and commutation techniques, and lack of ride-through capability such as the stringent case of one phase loss operation. This dissertation work provides comprehensive study on the commutation method and modulation scheme design for the isolated buck matrix-type rectifier. It aims to analyze the operation principle of the rectifier and propose viable modulation and commutation schemes for this rectifier under both three-phase and single-phase operation. The method is verified by the hardware experiments of the PSUs with high efficiency (> 98%) and high power density (> 70 W/in3 ) for 54 V and 380 VDC applications. The prototypes demonstrated in the experiments show the effectiveness of the proposed modulation and commutation schemes for industry.

In data center, applications which are typically multi-server, high-parallel, long-running like web-related applications are really common. They always compute over large-scale dataset and consume plenty of energy. Up until now, most researches focuses on trading a loss of performance for energy saving. However, managing the power is more important compared with reducing it. In this paper, we add energy consumption to the list of managed resources and help managers to control power profile of applications in data center. Tenants put forward the power budget and corresponding response time target of their own applications before they rent servers. We designed strategies to make every application in data center running under a global power cap and their own power budget. We first propose a Global Feedback Power Allocation Policy to periodically allocate the global power cap among the applications. We also devise a Local Efficient Power Policy to determine application-level power cap and allocate it among servers running the application. Extra budget during each period can be used in the rest of the tenancy to increase the application-level power cap to minimize the response time. We use Shell, WWW, DNS and Mail workloads to evaluate our policy.

Large data centers consume large amounts of electricity. Estimating the energy consumption in a data center can be of great importance to data centers administrators in order to know the energy-consuming tasks and take actions for reducing the total energy consumption. Smart workflow mechanisms can be built to reduce the energy consumption of data centers significantly. In this paper, we are investigating the factors that affect the energy consumption of scientific applications in data centers. We also use eight machine learning methods to estimate the energy consumption of multi-threaded scientific applications. Extensive computational results on a computer with 20 cores show that the CPU usage is the most important parameter in the power consumed by an application. However, better results can be obtained when the CPU utilization is combined with other parameters. We generate various regression models that predict the energy consumption of an application with an average accuracy of 99%. Simpler models with one and two parameters can achieve comparable accuracy with more complex models. We also compare various machine learning methods for their ability to obtain accurate predictions using as few parameters as possible.

With the emerging diverse applications in data centers, the demands on quality of service in data centers also become diverse, such as high throughput of elephant flows and low latency of deadline-sensitive flows. However, traditional TCPs are ill-suited to such situations and always result in the inefficiency (e.g. missing the flow deadline, inevitable throughput collapse) of data transfers. This further degrades the user-perceived quality of service (QoS) in data centers. To reduce the flow completion time of mice and deadline-sensitive flows along with promoting the throughput of elephant flows, an efficient and deadline-aware priority-driven congestion control (PCC) protocol, which grants mice and deadline-sensitive flows the highest priority, is proposed in this paper. Specifically, PCC computes the priority of different flows according to the size of transmitted data, the remaining data volume, and the flows' deadline. Then PCC adjusts the congestion window according to the flow priority and the degree of network congestion. Furthermore, switches in data centers control the input/output of packets based on the flow priority and the queue length. Different from existing TCPs, to speed up the data transfers of mice and deadline-sensitive flows, PCC provides an effective method to compute and encode the flow priority explicitly. According to the flow priority, switches can manage packets efficiently and ensure the data transfers of high priority flows through a weighted priority scheduling with minor modification. The experimental results prove that PCC can improve the data transfer performance of mice and deadline-sensitive flows while guaranting the throughput of elephant flows.

Virtualization technologies are extensively utilized in data centers, particularly cloud computing. This facilitates data center management and diminishes the number of physical machines (servers) and, subsequently, their cooling requirements, leading to cost, space, and power consumption reductions. When applications in data centers are executing independent parallel transactions, but with similar performance requirements, the appropriate level of virtual machine consolidation on a server poses a fundamental challenge for capacity planning. This article introduces a method to evaluate the performance speedup achieved through virtualization on any server and the effects of virtualization and consolidation overheads on physical or virtual machine scalability. This research formalizes the speedup and overheads, using classical computer architecture statements. but at the same time proposes a new method to analyze these overhead amounts and types, showing the scalability and efficiency of different consolidations in the same server and its comparison against no consolidation. This work also proposes a new way to determine the optimal number of physical servers and the optimal number of consolidated virtual machines for a given transaction workload. The real experimentation was performed with different workload sizes, types of virtualizations and different servers. The method presented also facilitates the representation of linear scalability against the real degree of parallelism of either physical machines or consolidated virtual machines for a given transaction workload, as well as striking the right balance between speedup and energy in virtual server consolidation.

Cloud computing has been a cost-efficient paradigm for deploying various applications in datacenters in recent years. Therefore, efficient provisioning for virtual data center (VDC) requests from different service providers (SPs) over physical data centers plays a vital role in improving the quality of service (QoS) and reducing the operational cost of SPs. Therefore, a significant attention has been paid for the VDC provisioning problem. However, few approaches have been proposed for the problem of reliable VDC embedding across multiple data centers, as most of them only consider the problem of VDC mapping within a single data center. In this paper, we study the problem of QoS-aware VDC provisioning across multiple data centers, such that the total bandwidth consumption in the inter-data center backbone network is minimized while satisfying the reliability requirement of each VDC request. We formulate this problem as a mathematical optimization problem by using integer linear programming (ILP) and propose an efficient heuristic algorithm called reliable VDC embedding (RVDCE) algorithm to solve this NP-hard problem. The simulation results show that the proposed algorithm performs better in terms of blocking ratio, CPU resource consumption, and bandwidth consumption of backbone network than the existing solution. In addition, this paper has also incorporated integrated security to minimize security vulnerabilities seen in other similar approaches. Apart from demonstrating how to resolve security challenges in our VDC proposal, cost calculations have been implemented to demonstrate the robustness, resiliency, validity, and effectiveness of the VDC provisioning solution for cloud computing.

Data centers are a key building block in the rapidly growing area of internet technology. A typical data center has tens of thousands of servers, and communication between them must be flexible and robust. Vortex light beams have orbital angular momentum and can provide a useful and flexible method for optical wireless communication in data centers. Vortex beams can be generated with orbital angular momentum but independent of polarization, and used in a multiplexed system. We propose a multiplexing vortex system to increase the communication capacity using optical wireless communication for data center networks. We then evaluate performance. This paper is intended for use as an engineering guideline for design of vortex multiplexing in data center applications.

There are striking similarities between network of dynamic systems and results developed in the area of machine learning for Bayesian Networks. Given these similarities, a fundamental goal is to formulate the problem of reconstructing a network of stochastic processes on the ground of general theoretical foundations. To this aim, we will rely on the mathematical framework provided by the theory of “semi-graphoids” developed by Judea Pearl. We will translate familiar concepts of Control Theory, such as Wiener filtering, in the language of semi-graphoids. This endevour will bring several positive outcomes: the results, mainly on acyclic networks, already obtained for semi-graphoids in the artificial intelligence community will seamlessly extend to the case of networks of stochastic processes; the control theory community will be exposed to new concepts that can be readily applied to stochastic systems; finally, it will motivate research towards new significant directions, such as the development of network identification/reconstruction techniques where feedback loops are present.

Large scale applications in data centers are composed of computers connected with a network. Traditional network switches do not perform routing based on packet contents. Thus, packets cannot be transmitted to the optimal computer for the request which is written in the packet payload. On the other hand, Deeply Programmable Network (DPN) switches can completely analyze packet payloads and perform routing based on their contents. In this paper, we focus on packet routing based on Deep Packet Inspection (DPI), i.e. analyzing packet payload, using DPN switches, and discuss performance of applications on a network using application aware DPN switches. We introduce application aware network and propose a method for improving application performance with DPN switch support. We evaluate our method with the several applications and show that our method can increase application performance.