Catálogo de publicaciones - libros

The popularity of research on the scheduling of parallel jobs demands a periodic review of the status of the field. Indeed, several surveys have been written on this topic in the context of parallel supercomputers [17, 20]. The purpose of the present paper is to update that material, and to extend it to include work concerning clusters and the grid.

The well-known TOP500 list ranks the 500 most powerful high-performance computers. However, the list lacks details about the job management and scheduling on these machines. As this statistic is interesting for researchers and system designers, this paper gives an overview and survey on scheduling relevant information for the first 50 entries in the TOP500 list.

In order to evaluate different scheduling strategies for parallel computers, simulations are often executed. As the scheduling quality highly depends on the workload that is served on the parallel machine, a representative workload model is required. Common approaches such as using a probability distribution model can capture the static feature of real workloads, but they do not consider the temporal relation in the traces. In this paper, a workload model is presented which uses Markov chains for modeling job parameters. In order to consider the interdependence of individual parameters without requiring large scale Markov chains, a novel method for transforming the states in different Markov chains is presented. The results show that the model yields closer results to the real workloads than other common approaches.

The self-tuning dynP scheduler for modern cluster resource management systems switches between different basic scheduling policies dynamically during run time. This allows to react on changing characteristics of the waiting jobs. In this paper we present enhancements to the decision process of the self-tuning dynP scheduler and evaluate their impact on the performance: (i) While doing a self-tuning step a performance metric is needed for ranking the schedules generated by the different basic scheduling policies. This allows different objectives for the self-tuning process, e.g. more user centric by improving the response time, or more owner centric by improving the makespan. (ii) Furthermore, a self-tuning process can be called at different times of the scheduling process: only at times when the characteristics of waiting jobs change (half self-tuning), i.e. new jobs are submitted; or always when the schedule changes (full self-tuning), i.e. when jobs are submitted or running jobs terminate.

We use discrete event simulations to evaluate the achieved performance. As job input for driving the simulations we use original traces from real supercomputer installations. The evaluation of the two enhancements to the decision process of the self-tuning dynP scheduler shows that a good performance is achieved, if the self-tuning metric is the same as the metric used measuring the overall performance at the end of the simulation. Additionally, calling the self-tuning process only when new jobs are submitted, is sufficient in most scenarios and the performance difference to full self-tuning is small.

Using a single traditional gang scheduling algorithm cannot provide the best performance for all workloads and parallel architectures. A solution for this problem is an algorithm that is capable of dynamically changing its form (configuration) into a more appropriate one, according to environment variations and user requirements. In this paper, we propose, implement and analyze the performance of a Reconfigurable Gang Scheduling Algorithm (RGSA) using simulation. A RGSA uses combinations of independent features that are often implemented in GSAs such as: packing and re-packing schemes (alternative scheduling etc.), multiprogramming levels etc. Ideally, the algorithm may assume infinite configurations and it reconfigures itself according to entry parameters such as: performance metrics (mean utilization, mean response time of jobs etc.) and workload characteristics (mean execution time of jobs, mean parallelism degree of jobs etc.). Also ideally, a reconfiguration causes the algorithm to output the best configuration for a particular situation considering the system’s state at a given moment. The main contributions of this paper are: the definition, proposal, implementation and performance analysis of RGSA.

This article is written in the context of running a suite of time-critical operational numerical weather prediction batch jobs, along with a substantial number of research batch jobs on a large IBM Cluster 1600 system. The batch subsystem used is IBM’s LoadLeveler incorporating a little known feature called Resource Reservation.

The article describes how the mixture of operational and research parallel batch jobs are scheduled to run on the 117 nodes provided, and how Resource Reservation for operational jobs is performed without reference to job class. Where research parallel batch jobs are jobs requesting more than 1 CPU and must run consistently to ensure resources are released predictably. Note – information is given explaining how consistent runtimes are achieved.

We are developing a distributed computing environment based on virtual machines featuring application monitoring, network monitoring, and an adaptive virtual network. In this paper, we describe our initial results in monitoring the communication traffic of parallel applications, and inferring its spatial communication properties. The ultimate goal is to be able to exploit such knowledge to maximize the parallel efficiency of the running parallel application by using VM migration, virtual overlay network configuration and network reservation techniques, which are a part of the distributed computing environment. Specifically, we demonstrate that: (1) we can monitor the parallel application network traffic in our layer 2 virtual network system with very low overhead, (2) we can aggregate the monitoring information captured on each host machine to form a global picture of the parallel application’s traffic load matrix, and (3) we can infer from the traffic load matrix the application topology. In earlier work, we have demonstrated that we can capture the time dynamics of the applications. We begin here by considering offline traffic monitoring and inference as a proof of concept, testing it with a variety of synthetic and actual workloads. Next, we describe the design and implementation of our online system, the Virtual Topology and Traffic Inference Framework (VTTIF), and evaluate it using a NAS benchmark.

Three-dimensional torus is a common topology of network interconnects of multicomputers due to its simplicity and high scalability. A parallel job submitted to a three-dimensional toroidal machine typically requires an isolated, contiguous, rectangular partition connected as a mesh or a torus. Such partitioning leads to fragmentation and thus reduces resource utilization of the machines. In particular, toroidal partitions often require allocation of additional communication links to close the torus. If the links are treated as dedicated resources (due to the partition isolation requirement) this may prevent allocation of other partitions that could, otherwise, use those links. Overall, on toroidal machines, the likelihood of successful allocation of a new partition decreases as the number of toroidal partitions increases.

This paper presents a novel ”multi-toroidal” interconnect topology that is able to accommodate multiple adjacent meshed and toroidal partitions at the same time. We prove that this topology allows connecting every free partition of the machine as a torus without affecting existing partitions. We also show that for toroidal jobs this interconnect topology increases machine utilization by a factor of 2 to 4 (depending on the workload) compared with three-dimensional toroidal machines. This effect exists for different scheduling policies. The BlueGene/L supercomputer being developed by IBM Research is an example of a multi-toroidal interconnect architecture.

We present an analysis of the costs and benefits of load sharing of parallel jobs in the computational grid. We begin with a workload generation model that captures the essential properties of parallel jobs and use it as input to a grid simulation model. Our experiments are performed for both homogeneous and heterogeneous grids. We measured average job slowdown with respect to both local and remote jobs and we show that, with some reasonable assumptions concerning the migration policy, load sharing proves to be beneficial when the grid is homogeneous, and that load sharing can adversely affect job slowdown for lightly-loaded machines in a heterogeneous grid. With respect to the number of sites in a grid, we find that the benefits obtained by load sharing do not scale well. Small to modest-size grids can employ load sharing as effectively as large-scale grids. We also present and evaluate an effective scheduling heuristic for migrating a job within the grid.

This paper presents a comprehensive characterization of a multi-cluster supercomputer workload using twelve-month scientific research traces. Metrics that we characterize include system utilization, job arrival rate and interarrival time, job cancellation rate, job size (degree of parallelism), job runtime, memory usage, and user/group behavior. Correlations between metrics (job runtime and memory usage, requested and actual runtime, etc) are identified and extensively studied. Differences with previously reported workloads are recognized and statistical distributions are fitted for generating synthetic workloads with the same characteristics. This study provides a realistic basis for experiments in resource management and evaluations of different scheduling strategies in a multi-cluster research environment.