Applying backfilling over a non-dedicated cluster

Space-sharing is regarded as the proper resource management scheme for many-core OSes. For today's many-core chips and parallel programming models providing no explicit resource requirements, an important research problem is to provide a proper resource allocation to the running applications while considering not only the architectural features but also the characteristics of the parallel applications. In this paper, we introduce a space-shared scheduling strategy for shared-memory parallel programs. To properly assign the disjoint set of cores to simultaneously running parallel applications, the proposed scheme considers the performance characteristics of the executing (parallel) code section of all running applications. The information about the performance is used to compute a proper core allocation in accordance to the goal of the scheduling policy given by the system manager. We have first implemented a user-level scheduling framework that runs on Linux-based multi-core chips. A simple performance model based solely on online profile data is used to characterize the performance scal-ability of applications. The framework is evaluated for two scheduling policies, balancing and maximizing QoS, and on two different many-core platforms, a 64-core AMD Opteron platform and a 36-core Tile-Gx36 processor. Experimental results of various OpenMP benchmarks show that in general our space-shared scheduling outperforms the standard Linux scheduler and meets the goal of the active scheduling policy.

Proceedings of the 17th …, 2005

Scientific investigations have to deal with rapidly growing amounts of data from simulations and experiments. During data analysis, scientists typically want to extract subsets of the data and perform computations on them. In order to speed up the analysis, computations are performed on distributed systems such as computer clusters, or Grid systems. A well-known difficult problem is to build systems that execute the computations and data movement in a coordinated fashion. In this paper, we describe an architecture for executing co-scheduled tasks of computation and data movement on a computer cluster that takes advantage of two technologies currently being used in distributed Grid systems. The first is Condor, that manages the scheduling and execution of distributed computation, and the second is Storage Resource Managers (SRMs) that manage the space usage and content of storage systems. This is achieved by including the information about the availability of files on the nodes provided by SRMs into the advertised information that Condor uses for the purpose of matchmaking. The system is capable of dynamically load balancing by replicating popular files on idle nodes. To confirm the feasibility of our approach, a prototype system was built on a computer cluster. Several experiments based on real work logs were performed. We observed that without replication compute nodes are underutilized and job wait times in the scheduler's queue are longer. This architecture can be used in wide-area Grid systems since the basic components are already used for the Grid.

Lecture Notes in Computer Science, 2003

Our research is focussed on keeping both local and parallel jobs together in a time-sharing NOW. In such systems, new scheduling approaches are needed to allocate the underloaded resources among parallel jobs. In this paper, a distributed system, named Cooperating CoScheduling, is presented to assign CPU and memory resources efficiently by applying resource balancing and dynamic coscheduling between parallel jobs. This is shown experimentally in a PVM-Linux NOW.

The ability to fully utilise the computational resources of a cluster of workstations (COW) through the initial placement and movement of workload is desirable not only by the user of the system, but also from an efficiency point of view since the modern COWs are an expensive resource. The ability to have an even spread of load over an entire COW can be achieved through the employment of Global Scheduling. This paper introduces the concept of global scheduling exploiting static allocation and dynamic load balancing along with the implementation of such a facility and support services, remote process creation and process migration, respectively, on the RHODOS distributed operating system. The suitability of the RHODOS implementation is demonstrated by results obtained from initial test runs.

Proceedings International Conference on Parallel Computing in Electrical Engineering. PARELEC 2000, 2000

This paper presents an algorithm for scheduling communication-intensive parallel applications in workstation clusters environment. The proposed scheduling policy combines the best attributes of both space-sharing and time-sharing principles and coexists with local schedulers (e.g., the Windows NT scheduler), which both provides coordinated scheduling and can generalise to provide a wide range of resource abstractions.

2009

Multiprocessor systems are the wave of the future rightly said because they offer tremendous potential to solve inherently parallel and complex computation intensive applications. In order to exploit the full potential of such computing systems, job scheduling or processor allocation (both are considered synonyms here) decisions plays a great role. Such scheduling decisions involves determining number of jobs to execute simultaneously as well as the number of processors to be allocated to each running application in a manner so as to minimize job's execution time and/or maximizing throughput. The growth of such multiprocessor systems has in turn paves the way for creation of efficient processor allocation policies in order to reduce job response time and make efficient utilization of system's processors. When we submit jobs or applications to multiprocessor system which in turn relies on job scheduling policies to allocate processors to such incoming jobs, we are really interested to know how well such policies are performing. Performance evaluation methodologies like actual experimental setup i.e. multiprocessor or parallel system, Theoretical/Analytical modelling and Simulation can be used to evaluate the performance of scheduling policies. Actual experimentation on multiprocessor or parallel system is still a costly and complex approach and moreover these systems are still out of reach to young researchers even doing research in higher education institutes like universities or technical colleges in developing countries India, Pakistan, Malaysia and Bangladesh etc. There may be several reasons for the nonavailability of these systems. One can very well evade out theoretical/analytical modelling to be used for the same purpose due to their inaccuracy. All these drawbacks have motivated us to switch towards virtualization or simulation of multiprocessor environment for the performance measurement of processor allocation policies. Simulation provides the powerful way to measure performance before the system under study has not actually been implemented. Such simulation can capture the dynamic interaction between applications and parallel architectures. Also it offers flexibility as one can make modifications to the simulation model and check their effect easily.

2008

Co-scheduling of parallel jobs in the chips is well-known to produce benefits in both system and individual job efficiency. The existing works have shown that job co-scheduling can effectively improve the contention, yet the question on the determination of optimal co-schedules still remains unanswered. The need for co-scheduling has been typically associated with communication bandwidth and the memory. In our work we have proposed a novel scheduling algorithm for optimal co-scheduling of parallel jobs. This algorithm facilitates the scheduling of parallel jobs using bandwidth and memory concepts. The coscheduling of the processes in the chips using the proposed algorithm shows satisfactory improvement in performance of running the parallel jobs.

IEEE Transactions on Parallel and Distributed Systems, 2000

Scheduling of processes onto processors of a parallel machine has always been an important and challenging area of research. The issue becomes even more crucial and difficult as we gradually progress to the use of off-the-shelf workstations, operating systems, and high bandwidth networks to build cost-effective clusters for demanding applications. Clusters are gaining acceptance not just in scientific applications that need supercomputing power, but also in domains such as databases, web service and multimedia, which place diverse Qualityof-Service (QoS) demands on the underlying system. Further, these applications have diverse characteristics in terms of their computation, communication and I/O requirements, making conventional parallel scheduling solutions, such as space sharing or coscheduling, unattractive. At the same time, leaving it to the native operating system of each node to make decisions independently can lead to ineffective use of system resources whenever there is communication. Instead, an emerging class of dynamic coscheduling mechanisms, that attempt to take remedial actions to guide the system towards coscheduled execution without requiring explicit synchronization, offers a lot of promise for cluster scheduling.

2004

In this study, a cluster-computing environment is employed as a computational platform. In order to increase the efficiency of the system, a dynamic task scheduling algorithm is proposed, which balances the load among the nodes of the cluster. The technique is dynamic, nonpreemptive, adaptive, and it uses a mixed centralised and decentralised policies. Based on the divide and conquer principle, the algorithm models the cluster as hyper-grids and then balances the load among them. Recursively, the hyper-grids of dimension k are divided into grids of dimensions k - 1, until the dimension is 1. Then, all the nodes of the cluster are almost equally loaded. The optimum dimension of the hyper-grid is chosen in order to achieve the best performance. The simulation results show the effective use of the algorithm. In addition, we determined the critical points (lower bounds) in which the algorithm can to be triggered.