Dynamic forms of resource pricing have recently been introduced by cloud providers that offer Infrastructure as a Service (IaaS) capabilities in order to maximize profits and balance resource supply and demand. The design of a mechanism that efficiently prices perishable cloud resources in line with a provider’s profit maximization goal remains an open research challenge, however. In this article, we propose the Online Extended Consensus Revenue Estimate mechanism in the setting of a recurrent, multiunit and single price auction for IaaS cloud resources. The mechanism is envy-free, has a high probability of being truthful, and generates a near optimal profit for the provider. We combine the proposed auction design with a scheme for dynamically calculating reserve prices based on data center Power Usage Effectiveness (PUE) and electricity costs. Our simulation-based evaluation of the mechanism demonstrates its effectiveness under a broad variety of market conditions. In particular, we show how it improves on the classical uniform price auction, and we investigate the value of prior knowledge on the execution time of virtual machines for maximizing profit. We also developed a system prototype and conducted a small-scale experimental study with a group of 10 users that confirms the truthfulness property of the mechanism in a real test environment.

Infrastructure-as-a-Service cloud providers offer diverse purchasing options and pricing plans, namely on-demand, reservation, and spot market plans. This allows them to efficiently target a variety of customer groups with distinct preferences and to generate more revenue accordingly. An important consequence of this diversification however, is that it introduces a non-trivial optimization problem related to the allocation of the provider's available data center capacity to each pricing plan. The complexity of the problem follows from the different levels of revenue generated per unit of capacity sold, and the different commitments consumers and providers make when resources are allocated under a given plan. In this work, we address a novel problem of maximizing revenue through an optimization of capacity allocation to each pricing plan by means of admission control for reservation contracts, in a setting where aforementioned plans are jointly offered to customers. We devise both an optimal algorithm based on a stochastic dynamic programming formulation and two heuristics that trade-off optimality and computational complexity. Our evaluation, which relies on an adaptation of a large-scale real-world workload trace of Google, shows that our algorithms can significantly increase revenue compared to an allocation without capacity control given that sufficient resource contention is present in the system. In addition, we show that our heuristics effectively allow for online decision making and quantify the revenue loss caused by the assumptions made to render the optimization problem tractable.

Modern data centers use virtualization as a means to increase utilization of increasingly powerful multi-core servers. Applications often require only a fraction of the resources provided by modern hardware. Multiple concurrent workloads are therefore required to achieve adequate utilization levels. Current virtualization solutions allow hardware to be partitioned into Virtual Machines with appropriate isolation on most levels. However, unmanaged consolidation of resource intensive workloads can still lead to unexpected performance variance. Measures are required to avoid or reduce performance interference and provide predictable service levels for all applications. In this paper, we identify and reduce network-related interference effects using performance models based on the runtime characteristics of virtualized workloads. We increase the applicability of existing training data by adding network-related performance metrics and benchmarks. Using the extended set of training data, we predict performance degradation with existing modeling techniques as well as combinations thereof. Application clustering is used to identify several new network-related application types with clearly defined performance profiles. Finally, we validate the added value of the improved models by introducing new scheduling techniques and comparing them to previous efforts. We demonstrate how the inclusion of network-related parameters in performance models can significantly increase the performance of consolidated workloads.

With the increased adoption of cloud computing, new challenges have emerged related to the cost-effective management of cloud resources. The proliferation of resource properties and pricing plans has made the selection, procurement and management of cloud resources a time-consuming and complex task, which stands to benefit from automation. This contribution focuses on the procurement decision of reserved contracts in the context of Infrastructure-as-a-Service (IaaS) providers such as Amazon EC2. Such reserved contracts complement pay-by-the-hour pricing models, and offer a significant reduction in price (up to 70%) for a particular period in return for an upfront payment. Thus, customers can reduce costs by predicting and analyzing their future needs in terms of the number and type of server instances. We present an algorithm that uses load prediction with automated time series forecasting based on a Double-seasonal Holt-Winters model, in order to make cost-efficient purchasing decisions among a wide range of contract types while taking into account an organization's current contract portfolio. We analyze its cost effectiveness through simulation of real-world web traffic traces. Our analysis investigates the impact of different prediction techniques on cost compared to a clairvoyant predictor and compares the algorithm's performance with a stationary contract renewal approach. Our results show that the algorithm is able to significantly reduce IaaS resource costs through automated reserved contract procurement. Moreover, the algorithm's computational cost makes it applicable to large-scale real-world settings.

SUMMARY Computer simulations have become an indispensable tool for the empirical study of large‐scale systems. The timely simulation of these systems, however, is not without its challenges. Simulators have to be able to harness the full computational power of modern multicore architectures through parallel execution and overcome the memory limitations of a single computer. In this paper, we evaluate the performance of a parallel and distributed simulator using several conventional time synchronization protocols executed on modern multicore hardware. In addition, we comprehensively analyze a hybrid approach, combining two traditional protocols, increasing robustness, and enabling improved performance in a wider range of simulation scenarios. Finally, an adaptive algorithm to automatically configure this hybrid protocol is introduced and evaluated, eliminating manual user intervention and further improving robustness with respect to varying simulation conditions. Copyright © 2013 John Wiley & Sons, Ltd.

A discrete-event simulator's ability to distribute the execution of a simulation model allows one to deal with the memory limitations of a single computational resource, and thereby increase the scale or level of detail at which models can be studied. In addition, distribution has the potential to reduce the round trip time of a simulation by incorporating multiple computational cores into the simulation's execution. However, such gains can be voided by the overhead that time synchronization protocols introduce. These protocols are required to prevent the occurrence of causality errors during a parallel execution of a simulation. The overhead depends on the protocol, characteristics of the simulation model, and the architecture of the computational resources used. Recently, infrastructure-as-a-service offerings in cloud computing have introduced flexibility in acquiring computational resources on a pay-as-you-go basis. At present, it is unclear to what extent these offerings are suited for the distributed execution of discrete-event simulations, and how the characteristics of different resource types impact the runtime performance of distributed simulations. In this paper we investigate this issue, and assess the performance of different conservative time synchronization protocols on a range of cloud resource types that are currently available on Amazon EC2.

Cloud computing offerings are maturing steadily, and their use has found acceptance in both industry and research. Cloud servers are used more and more instead of, or in addition to, local compute and storage infrastructure. Deciding which workloads to outsource to what cloud provider in such a setting, however, is far from trivial. This decision should maximize the utilization of the internal infrastructure and minimize the cost of running the outsourced tasks in the cloud, while taking into account the applications' quality of service constraints. Such decisions are generally hard to take by hand, because there are many cost factors, pricing models and cloud provider offerings to consider. In this work, we tackle this problem by proposing a set of heuristics to cost-efficiently schedule deadline-constrained computational applications on both public cloud providers and private infrastructure. Our heuristics take into account both computational and data transfer costs as well as estimated data transfer times. We evaluate to which extent the different cost factors and workload characteristics influence the cost savings realized by the heuristics and analyze the sensitivity of our results to the accuracy of task runtime estimates.

With the recent emergence of public cloud offerings, surge computing -outsourcing tasks from an internal data center to a cloud provider in times of heavy load- has become more accessible to a wide range of consumers. Deciding which workloads to outsource to what cloud provider in such a setting, however, is far from trivial. The objective of this decision is to maximize the utilization of the internal data center and to minimize the cost of running the outsourced tasks in the cloud, while fulfilling the applications' quality of service constraints. We examine this optimization problem in a multi-provider hybrid cloud setting with deadline-constrained and preemptible but non-provider-migratable workloads that are characterized by memory, CPU and data transmission requirements. Linear programming is a general technique to tackle such an optimization problem. At present, it is however unclear whether this technique is suitable for the problem at hand and what the performance implications of its use are. We therefore analyze and propose a binary integer program formulation of the scheduling problem and evaluate the computational costs of this technique with respect to the problem's key parameters. We found out that this approach results in a tractable solution for scheduling applications in the public cloud, but that the same method becomes much less feasible in a hybrid cloud setting due to very high solve time variances.

General purpose compute clusters fulfill a prominent role in a wide range of organizations to deliver the necessary computational power for their processes. In order to manage the shared use of such clusters, scheduling policies are installed to determine if and when the jobs submitted to the cluster are executed. Value-based scheduling policies differ from other policies in that they allow users to communicate the value of their computation to the scheduling mechanism. The design of market mechanisms whereby users are able to bid for resources in a fine-grained manner has proven to be an attractive means to implement such policies. In the clearing phase of the mechanism, supply and demand for resources are matched in pursuit of a value-maximizing job schedule and resource prices are dynamically adjusted to the level of excess demand in the system. Despite their success in simulations and research literature, such fine-grained value-based scheduling policies have been rarely used in practice as they are often considered too fragile, too onerous for end-users to work with, and difficult to implement. A coarse-grained form of value-based scheduling that mitigates aformentioned disadvantages involves the installation of a priority queueing system with fixed costs per queue. At present, it is however unclear to which extent a fine-grained form of value-based scheduling through auctions can outperform such a priority queueing system. Using workload traces of a general purpose cluster, we indicate under which conditions this is the case and quantify the resulting efficiency gains.

In economic resource management systems, the objective is to bring about an efficient allocation of resource while allowing market participants, resource consumers and owners, to act on their individual valuations of jobs and resources. In this paper we investigate the pricing of resources in dynamic grids. Particularly, we investigate grid resource management based on a computational commodity market of CPU resources, where resources are determined by supply-and-demand equilibrium. In particular introduce several categories of CPU's characterized by execution speed. These differ in cost and performance but may be used interchangeably in executing jobs and thus represent so-called substitutable resources. We investigate, using simulation, whether the algorithms for computing the supply-and-demand equilibrium, function in this context.

The efficient scheduling of jobs is an essential part of any grid resource management system. At its core, it involves ending a solution to a problem which is NP-complete by reduction to the knapsack problem. Consequently, this problem is often tackled by using heuristics to derive a more pragmatic solution. Other than the use of heuristics, simplifications and abstractions of the workload model may also be employed to increase the tractability of the scheduling problem. A possible abstraction in this context is the use of divisible load theory (DLT), in which it is assumed that an application consists of an arbitrarily divisible load (ADL). Many applications however, are composed of a number of atomic tasks and are only modularly divisible. In this paper we evaluate the consequences of the ADL assumption on the performance of economic scheduling approaches for grids, in the context of CPU-bound modularly divisible applications with hard deadlines. Our goal is to evaluate to what extent DLT can still serve as a useful workload abstraction for obtaining tractable scheduling algorithms in this setting. The focus of our evaluation is on the recently proposed tsfGrid heuristic for economic scheduling of grid workloads which operates under the assumptions of ADL. We demonstrate the effect of the ADL assumption on the actual instantiation of schedules and on the user value realized by the RMS. In addition we describe how the usage of a DLT heuristic in a high-level admission controller for a mechanism which does take into account the atomicity of individual tasks, can significantly reduce communication and computational overhead.

Abstract Grid computing technology enables the creation of large‐scale IT infrastructures that are shared across organizational boundaries. In such shared infrastructures, conflicts between user requirements are common and originate from the selfish actions that users perform when formulating their service requests. The introduction of economic principles in grid resource management offers a promising way of dealing with these conflicts. We develop and analyze both a centralized and a decentralized algorithm for economic grid resource management in the context of compute bound applications with deadline‐based quality of service requirements and non‐migratable workloads. Through the use of reservations, we co‐allocate resources across multiple providers in order to ensure that applications finish within their deadline. An evaluation of both algorithms is presented and their performance in terms of realized user value is compared with an existing market‐based resource management algorithm. We establish that our algorithms, which operate under a more realistic workload model, can closely approximate the performance of this algorithm. We also quantify the effect of allowing local workload preemption and different scheduling heuristics on the realized user value. Copyright © 2009 John Wiley & Sons, Ltd.

We investigate resource management in computational grids based on an economic perspective. Specifically, we model the grid as a commodity market of CPU resources that are traded at prices determined by the supply-and- demand equilibrium. This approach has been shown to function for small, static grids. We present simulation results to demonstrate it can be extended to grids with a dynamic fabric and large (thousands) number of resources.

The introduction of economic principles in grid resource management provides an interesting avenue for efficiently addressing the problem of conflicting user requirements. In shared computing infrastructures such as grids, such conflicting requirements are prevalent and stem from the selfish actions users follow when formulating their service requests. We develop and analyze both a centralized and a decentralized algorithm for economic resource management in the context of consumer requests for CPU bound applications with deadline-based QoS requirements and non- migratable workloads. A comparison with an algorithm recently proposed in the literature is presented with a focus on performance in terms of realized consumer value. We establish that our algorithms perform well and that they compare favorably to existing approaches.