Advance reservation, co-allocation and pricing of network and computational resources in grids

https://doi.org/10.1016/j.future.2014.07.004Get rights and content

Highlights

  • A novel economic RMS supporting co-allocation and advance reservation of computational and network resources.

  • Estimation of network and compute prices to normalize budgets for and enable prioritization of requests.

  • Extended Minimal Cut List (EMCL) algorithm determines the network’s “limiting links”.

  • Usage of EMCLs to distribute path prices over individual links based on their importance.

  • Simulation results show that ENARA realizes significantly more value than an online system.

Abstract

Through the introduction of economic principles, a Resource Management System (RMS) can incorporate user valuations and externalities such as the usage cost of resources, into its scheduling logic. Consequently, an economic RMS can extract more value from the available infrastructure compared to traditional RMSs. In order to use the available infrastructure efficiently, a grid RMS must take an application’s data requirements into account. Although RMSs exist that support the co-allocation and advance reservation of both network and computational resources, they do not incorporate economic principles. In this paper we present ENARA, an economic RMS with advance reservation and co-allocation support for both network and computational resources. The RMS both allocates and prices resources in line with the demand and supply conditions in the network. Through the use of a novel algorithm that determines the limiting links within a network, the RMS is able to estimate prices of individual network links at the start of its planning phase. This resolves an interdependency issue between resource pricing and allocation. We demonstrate that ENARA is able to align resource allocations with user valuations, significantly increasing the user value generated compared to an online approach, while attaining high utilization levels of the infrastructure.

Introduction

In shared computing environments such as grid systems, Resource Management Systems (RMSs) have to deal with conflicting requirements due to the fact that users of such infrastructures act out of self-interest when formulating their requests for workload execution. A problem for traditional RMSs in this regard, is that they cannot incorporate externalities such as the usage cost of resources into the planning and scheduling logic. Additionally they cannot take individual user valuations into account in order to prioritize requests. Introducing economic principles in grid resource management allows one to deal with these issues effectively.

Traditional RMSs focus primarily on the efficient scheduling of applications from the point of view of the system itself. In short, the RMS attempts to schedule applications in such a way that either the utilization of the available resources or the throughput of the system is maximized. Consequently, while the RMS may perform optimally from a system-oriented view, this may not be the case from the user perspective  [1], [2].

Another problem related to current generation grid RMSs is that they are based on a system of static multilateral sharing agreements between the stakeholders of the shared infrastructure. The details of these agreements are difficult to establish and as such they have an innate resistance to change. Additionally, they are often based on reciprocality. Organizations that do not contribute infrastructure can therefore find it difficult to participate. Conversely, organizations that mainly provide infrastructure and do not make as much use of the grid as logically permitted by their contribution will most likely not benefit from their participation. This limits the flexibility and openness of the grid system and ultimately this leads to mutually exclusive grid systems or ‘islands’ which fail to realize the intrinsic potential of their combined infrastructures.

A final issue with current generation grids is that their operation is not self-sustainable. They are typically funded by government grants and subject to bargaining and quid pro quos of different universities and research institutions. There is little oversight over the efficient use of these systems and a long term vision for the evolution towards self-sustainability is often lacking.

We believe that, contrary to traditional RMSs and scheduling approaches in grid systems, the use of economic principles enables the creation of more open and sustainable grid systems that are oriented towards value maximization. These charge the users of the system according to their requirements and resulting allocations while taking into account those of other users as well. The introduction of accounting systems and the notion of prices allows for the internalization of the negative externalities  [3] caused by a user through its service request. This approach results in a number of benefits over current-generation RMSs.

Firstly, the incorporation of payments enables a decoupling between infrastructure providers and consumers which removes the need to negotiate complex sharing agreements. Charging for resource usage also creates incentives for resource providers to join the grid infrastructure and as such increases the long-term viability of the grid system. Secondly, consumer valuations and resource prices can potentially be used to prioritize conflicting requirements in periods of congestion to maximize the value realized by the system. Finally, since there is a tangible cost associated with formulating excessively strict QoS requirements, users are stimulated to formulate these conscientiously and in line with their true requirements. Since prices are formed dynamically based on the aggregate demand of all users, users are indirectly forced to take into account the requirements of others when optimizing their own utility.

In this article, we describe and evaluate ENARA, an economic network- and resource-aware RMS that prices resource usage on co-allocated and reserved computational resources and network paths in an offline manner. We extend the system model used in our previous work  [4] with data dependencies and network resources. Our workload model is centered on bag-of-task applications with input data requirements of which the execution is constrained by a hard deadline requirement. Our network model, used to simulate the delays associated with network communication and data transfer, is based on Lambda Grids. Our contributions can be summarized as follows:

  • We present and evaluate an economic RMS that allows for the co-allocation, reservation and pricing of both network and computational resources. The RMS prioritizes the use of the grid infrastructure in such a way that the execution of an application without data dependencies has a minimal impact on the probability of scheduling data dependent applications within their deadline, and that the overall user value generated by the resource allocations is maximized.

  • We present algorithms to determine the set of limiting links in the network to allow for discriminatory pricing on links and to use the available network infrastructure effectively. To define the set of limiting links, we introduce the concepts of minimal cut lists, extended minimal cuts and extended minimal cut lists (EMCL). The structure of the EMCL allows for a distribution of network path prices over the individual links.

  • We present a method to determine how the budget of an application should be split into a budget for computation and communication. To do so, we introduce a method for estimating the expected congestion and type of resource congestion to subsequently derive price estimations for computational and network resources. This allows us to prioritize applications based on a normalized budget that takes into account the relative cost of computation to communication and the total load generated by the application in these two dimensions. We also address a potential catch-22 situation where the overpricing of a given resource leads to the avoidance of that resource in the planning phase, which in turn leads to underutilization and lower efficiency.

  • As an alternative to the discretization of time in fixed-size slots, we rely on continuous time intervals in our planning and scheduling phase, removing the issue of internal fragmentation in resource schedules.

We do not require users to specify a fixed parallelization degree for their applications in contrast to several other existing approaches  [5], [6], [7]. Instead, the ENARA RMS is free to choose this degree when scheduling data transfers and computational workloads, as long as it can guarantee that the input files are transferred from storage to the execution site before the computation starts, and that the application finishes by the given deadline. We explicitly take into account the atomicity of jobs, the limited parallelization degree of an application, and do not require job preemption and migration in the construction of job schedules. We believe this is the first system to date that can co-allocate and reserve both network and CPU resources economically under these modeling assumptions.

This article is organized as follows. After reviewing related work in the next section, we present an overview of the ENARA system model in Section  3 and describe the algorithms for planning and pricing in Section  4. We conclude the paper with an evaluation of our approach, comparing its performance in terms of realized user value to an online network-aware scheduling policy.

Section snippets

Related work

There is to the best of our knowledge no other work that combines both economic aware network and CPU resource co-allocation and advance reservation. There is a sizeable body of work on the co-allocation of both network and computational resources on the one hand, and on the economic co-allocation of network resources on the other.

Another aspect that differentiates our work from existing approaches for advance reservation and co-allocation of network resources is that our approach foregoes

System model

This section presents the elements of our system model. First we introduce the GESNET network model that supports the simulation of data transfers and network reservations. Subsequently we describe the use of Planning Windows, our model of Jobs, Workflows and Data, and the models for Requests, Consumers and Providers. We conclude this section with a description of the ENARA broker.

Algorithms

This section describes the details of the most important algorithms used by the ENARA broker. First, we discuss the search for links in the network that limit the transfer capacity between a set of sources S and a set of sinks C in a network. We then explain how we estimate resource prices and distribute path prices over individual links. Finally, we explain how we create advance reservations for network transfers.

Experiments

Through a set of experiments, we demonstrate in this section that the different elements of our overall approach deliver correct and expected results. In what follows, we use a number of parameters to describe the settings of each simulated scenario of which we provide a summary in Table 1 for ease of reference. We note that the variance var of a value v is always given as a number between 0 and 1 and that for each member of an entity group, a uniformly distributed random value is chosen in the

Conclusion

Managing network and computational resources effectively in the context of deadline-constrained applications with data dependencies, requires advance reservation and co-allocation of these resources. In order to maximize the user value generated by the infrastructure, the execution of applications must be prioritized according to the valuations users attribute to their applications. This in turn requires the resource management system to derive prices for resource usage in order to incentivize

Acknowledgment

This work has been supported by the Research Foundation Flanders under grant 1.1.546.09N.

Wim Depoorter is a Ph.D. Fellow of the Research Foundation Flanders (FWO) supported by grant 1.1.546.09N.

His research interests include economically inspired resource management systems in grid environments and co-allocation and advance reservation of resources.

In 2007 he received his M.S. in Computer Science from the University of Antwerp, Magna Cum Laude.

References (21)

  • P. Kokkinos et al.

    Efficient data consolidation in grid networks and performance analysis

    Future Gener. Comput. Syst.

    (2011)
  • T. Stevens et al.

    Multi-cost job routing and scheduling in grid networks

    Future Gener. Comput. Syst.

    (2009)
  • M. Dramitinos et al.

    An auction mechanism for allocating the bandwidth of networks to their users

    Int. J. Comput. Telecomm. Netw.

    (2007)
  • K. Krauter et al.

    A taxonomy and survey of grid resource management systems for distributed computing

    Softw. - Pract. Exp.

    (2002)
  • C.B. Lee et al.

    On the user-scheduler dialogue: studies of user-provided runtime estimates and utility functions

    Int. J. High Perform. Comput. Appl.

    (2006)
  • A.C. Pigou

    The Economics of Welfare, Vol. 38

    (1960)
  • K. Vanmechelen et al.

    Market-based grid resource co-allocation and reservation for applications with hard deadlines

    Concurrency Comput. Pract. Exp.

    (2009)
  • B.N. Chun et al.

    Mirage: a microeconomic resource allocation system for sensornet testbeds

  • B. Schnizler

    Resource allocation in the grid—a market engineering approach

    (2007)
  • J. Stößer, D. Neumann, GreedEx—a scalable clearing mechanism for utility computing, in: Proceedings of the Networking...
There are more references available in the full text version of this article.

Cited by (10)

  • Resource allocation using Dynamic Pricing Auction Mechanism for supporting emergency demands in Cloud Computing

    2021, Journal of Parallel and Distributed Computing
    Citation Excerpt :

    Pricing of resources attains a critical manner in allotment of resources. Researchers describe merchandise retail which recommends the present vendors arranged the worth as long as commodity market with the purchasers settle cashes via gain it [10]. The cost fixed through the vendor does not get modified after some period established in the quantity - requirement relationship.

  • A fault-tolerant hybrid resource allocation model for dynamic computational grid

    2021, Journal of Computational Science
    Citation Excerpt :

    Initially, the gridsim simulator utilizes in the advance reservation which was previously originated after [27]. In [28], the author introduced a common technique for both tasks having reservation requests and scheduling a batch of computational tasks. Moreover, some research divulged on the reservation of resources using advance reservation negotiation strategy to ensure the quality of service [29].

  • Joint affinity aware grouping and virtual machine placement

    2017, Microprocessors and Microsystems
    Citation Excerpt :

    Jay Smitha et al. investigated a robust static resource allocation problem for distributed computing systems operating under imposed Quality of Service (QoS) constraints and used a unique application of both path relinking and local search within a Genetic Algorithm [31]. Recently, Depoorter et al. provided a resource management system to advance reservation, co-allocation and pricing of network and computational resources in grids [7]. Xiaoling Li et al. study a resource allocation with multi-factor node ranking in data center networks [19].

  • Design and evaluation of a dual dynamic adaptive reservation approach in media production networks

    2017, Journal of Network and Computer Applications
    Citation Excerpt :

    Section 5 provides simulation results and Section 6 concludes the article. Advance network resource reservation has applications for both wide-area and grid networks and has been studied frequently in recent years (Depoorter et al., 2014; Bochenina et al., 2016; Bai et al., 2016; Guok, 2005; Gibbard et al., 2006; Chunlin et al., 2009; Gu et al., 2011). Current research mostly focuses on optical networks in combination with wavelength division multiplexing (Charbonneau and Vokkarane, 2012).

  • Distributed meta-scheduling in lambda grids by means of Ant Colony Optimization

    2016, Future Generation Computer Systems
    Citation Excerpt :

    Different techniques have been proposed in the literature for dealing with advance network reservation and co-allocation of processing and networking resources in lambda grids. For instance, in [23], the authors proposed a model incorporating economic principles to address this problem, where the demand and supply conditions of the network are taken into account on the allocation and the pricing of the resources. Network reservations are made by iterating over the information of all previous reservations at the network links in the specified time window and by relying on a shortest path algorithm.

View all citing articles on Scopus

Wim Depoorter is a Ph.D. Fellow of the Research Foundation Flanders (FWO) supported by grant 1.1.546.09N.

His research interests include economically inspired resource management systems in grid environments and co-allocation and advance reservation of resources.

In 2007 he received his M.S. in Computer Science from the University of Antwerp, Magna Cum Laude.

Kurt Vanmechelen is a post-doctoral fellow in the Department of Mathematics and Computer Science at the University of Antwerp (UA), Belgium. His research interests include resource management in (smart) grid and cloud environments in general, and the adoption of market mechanisms in such systems in particular. In 2009 he received his Ph.D. in Computer Science, at the University of Antwerp (UA), Belgium.

Jan Broeckhove is a professor in the Department of Mathematics and Computer Science at the University of Antwerp (UA), Belgium.

His research interests include computational science and distributed computing. He received his Ph.D. in Physics, in 1982 at the Free University of Brussels (VUB), Belgium.

View full text