Host and network optimizations for performance enhancement and energy efficiency in data center networks

Modern data centers host hundreds of thousands of servers to achieve economies of scale. Such a huge number of servers create challenges for the data center network (DCN) to provide proportionally large bandwidth. In addition, the deployment of virtual machines (VMs) in data centers raises the requirements for efficient resource allocation and find-grained resource sharing. Further, the large number of servers and switches in the data center consume significant amounts of energy. Even though servers become more energy efficient with various energy saving techniques, DCN still accounts for 20% to 50% of the energy consumed by the entire data center. The objective of this dissertation is to enhance DCN performance as well as its energy efficiency by conducting optimizations on both host and network sides. First, as the DCN demands huge bisection bandwidth to interconnect all the servers, we propose a parallel packet switch (PPS) architecture that directly processes variable length packets without segmentation-and-reassembly (SAR). The proposed PPS achieves large bandwidth by combining switching capacities of multiple fabrics, and it further improves the switch throughput by avoiding padding bits in SAR. Second, since certain resource demands of the VM are bursty and demonstrate stochastic nature, to satisfy both deterministic and stochastic demands in VM placement, we propose the Max-Min Multidimensional Stochastic Bin Packing (M3SBP) algorithm. M3SBP calculates an equivalent deterministic value for the stochastic demands, and maximizes the minimum resource utilization ratio of each server. Third, to provide necessary traffic isolation for VMs that share the same physical network adapter, we propose the Flow-level Bandwidth Provisioning (FBP) algorithm. By reducing the flow scheduling problem to multiple stages of packet queuing problems, FBP guarantees the provisioned bandwidth and delay performance for each flow. Finally, while DCNs are typically provisioned with full bisection bandwidth, DCN traffic demonstrates fluctuating patterns, we propose a joint host-network optimization scheme to enhance the energy efficiency of DCNs during off-peak traffic hours. The proposed scheme utilizes a unified representation method that converts the VM placement problem to a routing problem and employs depth-first and best-fit search to find efficient paths for flows.

Host and Network Optimizations for Performance Enhancement and Energy Efficiency in Data Center Networks

Modern data centers host hundreds of thousands of servers to achieve economies of scale. Such a huge number of servers create challenges for the data center network (DCN) to provide proportionally large bandwidth. In addition, the deployment of virtual machines (VMs) in data centers raises the requirements for efficient resource allocation and find-grained resource sharing. Further, the large number of servers and switches in the data center consume significant amounts of energy. Even though servers become more energy efficient with various energy saving techniques, DCN still accounts for 20% to 50% of the energy consumed by the entire data center. The objective of this dissertation is to enhance DCN performance as well as its energy efficiency by conducting optimizations on both host and network sides. First, as the DCN demands huge bisection bandwidth to interconnect all the servers, we propose a parallel packet switch (PPS) architecture that directly processes variable length packets without segmentation-and-reassembly (SAR). The proposed PPS achieves large bandwidth by combining switching capacities of multiple fabrics, and it further improves the switch throughput by avoiding padding bits in SAR. Second, since certain resource demands of the VM are bursty and demonstrate stochastic nature, to satisfy both deterministic and stochastic demands in VM placement, we propose the Max-Min Multidimensional Stochastic Bin Packing (M3SBP) algorithm. M3SBP calculates an equivalent deterministic value for the stochastic demands, and maximizes the minimum resource utilization ratio of each server. Third, to provide necessary traffic isolation for VMs that share the same physical network adapter, we propose the Flow-level Bandwidth Provisioning (FBP) algorithm. By reducing the flow scheduling problem to multiple stages of packet queuing problems, FBP guarantees the provisioned bandwidth and delay performance for each flow. Finally, while DCNs are typically provisioned with full bisection bandwidth, DCN traffic demonstrates fluctuating patterns, we propose a joint host-network optimization scheme to enhance the energy efficiency of DCNs during off-peak traffic hours. The proposed scheme utilizes a unified representation method that converts the VM placement problem to a routing problem and employs depth-first and best-fit search to find efficient paths for flows.

Efficient asynchronous scheduling algorithms for cost-effective buffered crossbar switches

Buffered crossbar switches have recently attracted considerable attention as the next generation of high speed interconnects. They are a special type of crossbar switches with an exclusive buffer at each crosspoint of the crossbar. They demonstrate unique advantages over traditional unbuffered crossbar switches, such as high throughput, low latency, and asynchronous packet scheduling. However, since crosspoint buffers are expensive on-chip memories, it is desired that each crosspoint has only a small buffer. This dissertation proposes a series of practical algorithms and techniques for efficient packet scheduling for buffered crossbar switches. To reduce the hardware cost of such switches and make them scalable, we considered partially buffered crossbars, whose crosspoint buffers can be of an arbitrarily small size. Firstly, we introduced a hybrid scheme called Packet-mode Asynchronous Scheduling Algorithm (PASA) to schedule best effort traffic. PASA combines the features of both distributed and centralized scheduling algorithms and can directly handle variable length packets without Segmentation And Reassembly (SAR). We showed by theoretical analysis that it achieves 100% throughput for any admissible traffic in a crossbar with a speedup of two. Moreover, outputs in PASA have a large probability to avoid the more time-consuming centralized scheduling process, and thus make fast scheduling decisions. Secondly, we proposed the Fair Asynchronous Segment Scheduling (FASS) algorithm to handle guaranteed performance traffic with explicit flow rates. FASS reduces the crosspoint buffer size by dividing packets into shorter segments before transmission. It also provides tight constant performance guarantees by emulating the ideal Generalized Processor Sharing (GPS) model. Furthermore, FASS requires no speedup for the crossbar, lowering the hardware cost and improving the switch capacity. Thirdly, we presented a bandwidth allocation scheme called Queue Length Proportional (QLP) to apply FASS to best effort traffic. QLP dynamically obtains a feasible bandwidth allocation matrix based on the queue length information, and thus assists the crossbar switch to be more work-conserving. The feasibility and stability of QLP were proved, no matter whether the traffic distribution is uniform or non-uniform. Hence, based on bandwidth allocation of QLP, FASS can also achieve 100% throughput for best effort traffic in a crossbar without speedup.

A Markov chain location-allocation meta-model for hurricane relief

An emergency is a deviation from a planned course of events that endangers people, properties, or the environment. It can be described as an unexpected event that causes economic damage, destruction, and human suffering. When a disaster happens, Emergency Managers are expected to have a response plan to most likely disaster scenarios. Unlike earthquakes and terrorist attacks, a hurricane response plan can be activated ahead of time, since a hurricane is predicted at least five days before it makes landfall. This research looked into the logistics aspects of the problem, in an attempt to develop a hurricane relief distribution network model. We addressed the problem of how to efficiently and effectively deliver basic relief goods to victims of a hurricane disaster. Specifically, where to preposition State Staging Areas (SSA), which Points of Distributions (PODs) to activate, and the allocation of commodities to each POD. Previous research has addressed several of these issues, but not with the incorporation of the random behavior of the hurricane's intensity and path. This research presents a stochastic meta-model that deals with the location of SSAs and the allocation of commodities. The novelty of the model is that it treats the strength and path of the hurricane as stochastic processes, and models them as Discrete Markov Chains. The demand is also treated as stochastic parameter because it depends on the stochastic behavior of the hurricane. However, for the meta-model, the demand is an input that is determined using Hazards United States (HAZUS), a software developed by the Federal Emergency Management Agency (FEMA) that estimates losses due to hurricanes and floods. A solution heuristic has been developed based on simulated annealing. Since the meta-model is a multi-objective problem, the heuristic is a multi-objective simulated annealing (MOSA), in which the initial solution and the cooling rate were determined via a Design of Experiments. The experiment showed that the initial temperature (T0) is irrelevant, but temperature reduction (δ) must be very gradual. Assessment of the meta-model indicates that the Markov Chains performed as well or better than forecasts made by the National Hurricane Center (NHC). Tests of the MOSA showed that it provides solutions in an efficient manner. Thus, an illustrative example shows that the meta-model is practical.

Monetary cost -efficient query distribution in ISDN -connected heterogeneous databases

Current technology permits connecting local networks via high-bandwidth telephone lines. Central coordinator nodes may use Intelligent Networks to manage data flow over dialed data lines, e.g. ISDN, and to establish connections between LANs. This dissertation focuses on cost minimization and on establishing operational policies for query distribution over heterogeneous, geographically distributed databases. Based on our study of query distribution strategies, public network tariff policies, and database interface standards we propose methods for communication cost estimation, strategies for the reduction of bandwidth allocation, and guidelines for central to node communication protocols. Our conclusion is that dialed data lines offer a cost effective alternative for the implementation of distributed database query systems, and that existing commercial software may be adapted to support query processing in heterogeneous distributed database systems. ^

Interoperable resource brokering with policy-based provisioning and job allocation

The increasing needs for computational power in areas such as weather simulation, genomics or Internet applications have led to sharing of geographically distributed and heterogeneous resources from commercial data centers and scientific institutions. Research in the areas of utility, grid and cloud computing, together with improvements in network and hardware virtualization has resulted in methods to locate and use resources to rapidly provision virtual environments in a flexible manner, while lowering costs for consumers and providers. ^ However, there is still a lack of methodologies to enable efficient and seamless sharing of resources among institutions. In this work, we concentrate in the problem of executing parallel scientific applications across distributed resources belonging to separate organizations. Our approach can be divided in three main points. First, we define and implement an interoperable grid protocol to distribute job workloads among partners with different middleware and execution resources. Second, we research and implement different policies for virtual resource provisioning and job-to-resource allocation, taking advantage of their cooperation to improve execution cost and performance. Third, we explore the consequences of on-demand provisioning and allocation in the problem of site-selection for the execution of parallel workloads, and propose new strategies to reduce job slowdown and overall cost.^

The market value implications of pension asset allocation

Pension funds have been part of the private sector since the 1850's. Defined Benefit pension plans [DB], where a company promises to make regular contributions to investment accounts held for participating employees in order to pay a promised lifelong annuity, are significant capital markets participants, amounting to 2.3 trillion dollars in 2010 (Federal Reserve Board, 2013). In 2006, Statement of Financial Accounting Standards No.158 (SFAS 158), Employers' Accounting for Defined Benefit Pension and Other Postemployment Plans, shifted information concerning funding status and pension asset/liability composition from disclosure in the footnotes to recognition in the financial statements. I add to the literature by being the first to examine the effect of recent pension reform during the financial crisis of 2008-09. This dissertation is comprised of three related essays. In my first essay, I investigate whether investors assign different pricing multiples to the various classes of pension assets when valuing firms. The pricing multiples on all classes of assets are significantly different from each other, but only investments in bonds and equities were value-relevant during the recent financial crisis. Consistent with investors viewing pension liabilities as liabilities of the firm, the pricing multiples on pension liabilities are significantly larger than those on non-pension liabilities. The only pension costs significantly associated with firm value are actual rate of return and interest expense. In my second essay, I investigate the role of accruals in predicting future cash flows, extending the Barth et al. (2001a) model of the accrual process. Using market value of equity as a proxy for cash flows, the results of this study suggest that aggregate accounting amounts mask how the components of earnings affect investors' ability to predict future cash flows. Disaggregating pension earnings components and accruals results in an increase in predictive power. During the 2008-2009 financial crisis, however, investors placed a greater (and negative) weight on the incremental information contained in the individual components of accruals. The inferences are robust to alternative specifications of accruals. Finally, in my third essay I investigate how investors view under-funded plans. On average, investors: view deficits arising from under-funded plans as belonging to the firm; reward firms with fully or over-funded pension plans; and encourage those funds with unfunded pension plans to become funded. Investors also encourage conservative pension asset allocations to mitigate firm risk, and smaller firms are perceived as being better able to handle the risk associated with underfunded plans. During the financial crisis of 2008-2009 underfunded status had a lower negative association with market value. In all three models, there are significant differences in pre- and post- SFAS 158 periods. These results are robust to various scenarios of the timing of the financial crisis and an alternative measure of funding.

Interoperable Resource Brokering with Policy-based Provisioning and Job Allocation

The increasing needs for computational power in areas such as weather simulation, genomics or Internet applications have led to sharing of geographically distributed and heterogeneous resources from commercial data centers and scientific institutions. Research in the areas of utility, grid and cloud computing, together with improvements in network and hardware virtualization has resulted in methods to locate and use resources to rapidly provision virtual environments in a flexible manner, while lowering costs for consumers and providers. However, there is still a lack of methodologies to enable efficient and seamless sharing of resources among institutions. In this work, we concentrate in the problem of executing parallel scientific applications across distributed resources belonging to separate organizations. Our approach can be divided in three main points. First, we define and implement an interoperable grid protocol to distribute job workloads among partners with different middleware and execution resources. Second, we research and implement different policies for virtual resource provisioning and job-to-resource allocation, taking advantage of their cooperation to improve execution cost and performance. Third, we explore the consequences of on-demand provisioning and allocation in the problem of site-selection for the execution of parallel workloads, and propose new strategies to reduce job slowdown and overall cost.