Parallel Jobs Research Articles

Developing Graph-Based Co-Scheduling Algorithms on Multicore Computers

It is common that multiple cores reside on the same chip and share the on-chip cache. As a result, resource sharing can cause performance degradation of co-running jobs. Job co-scheduling is a technique that can effectively alleviate this contention and many co-schedulers have been reported in related literature. Most solutions however do not aim to find the optimal co-scheduling solution. Being able to determine the optimal solution is critical for evaluating co-scheduling systems. Moreover, most co-schedulers only consider serial jobs, and there often exist both parallel and serial jobs in real-world systems. In this paper a graph-based method is developed to find the optimal co-scheduling solution for serial jobs; the method is then extended to incorporate parallel jobs, including multi-process, and multi-threaded parallel jobs. A number of optimization measures are also developed to accelerate the solving process. Moreover, a flexible approximation technique is proposed to strike a balance between the solving speed and the solution quality. Extensive experiments are conducted to evaluate the effectiveness of the proposed co-scheduling algorithms. The results show that the proposed algorithms can find the optimal co-scheduling solution for both serial and parallel jobs. The proposed approximation technique is also shown to be flexible in the sense that we can control the solving speed by setting the requirement for the solution quality.

TU-AB-BRC-07: Efficiency of An IAEA Phase-Space Source for a Low Energy X-Ray Tube Using Egs++

Purpose: To extend the capability of the EGSnrc C++ class library (egs++) to write and read IAEA phase-space files as a particle source, and to assess the relative efficiency gain in dose calculation using an IAEA phase-space source for modelling a miniature low energy x-ray source. Methods: We created a new ausgab object to score particles exiting a user-defined geometry and write them to an IAEA phase-space file. A new particle source was created to read from IAEA phase-space data. With these tools, a phase-space file was generated for particles exiting a miniature 50 kVp x-ray tube (The INTRABEAM System, Carl Zeiss). The phase-space source was validated by comparing calculated PDDs with a full electron source simulation of the INTRABEAM. The dose calculation efficiency gain of the phase-space source was determined relative to the full simulation. The efficiency gain as a function of i) depth in water, and ii) job parallelization was investigated. Results: The phase-space and electron source PDDs were found to agree to 0.5% RMS, comparable to statistical uncertainties. The use of a phase-space source for the INTRABEAM led to a relative efficiency gain of greater than 20 over the full electron source simulation, with an increase of up to a factor of 196. The efficiency gain was found to decrease with depth in water, due to the influence of scattering. Job parallelization (across 2 to 256 cores) was not found to have any detrimental effect on efficiency gain. Conclusion: A set of tools has been developed for writing and reading IAEA phase-space files, which can be used with any egs++ user code. For simulation of a low energy x-ray tube, the use of a phase-space source was found to increase the relative dose calculation efficiency by factor of up to 196. The authors acknowledge partial support by the CREATE Medical Physics Research Training Network grant of the Natural Sciences and Engineering Research Council (Grant No. 432290).

TU‐AB‐BRC‐12: Optimized Parallel MonteCarlo Dose Calculations for Secondary MU Checks

Purpose:Secondary MU checks are an important tool used during a physics review of a treatment plan. Commercial software packages offer varying degrees of theoretical dose calculation accuracy, depending on the modality involved. Dose calculations of VMAT plans are especially prone to error due to the large approximations involved. Monte Carlo (MC) methods are not commonly used due to their long run times. We investigated two methods to increase the computational efficiency of MC dose simulations with the BEAMnrc code. Distributed computing resources, along with optimized code compilation, will allow for accurate and efficient VMAT dose calculations.Methods:The BEAMnrc package was installed on a high performance computing cluster accessible to our clinic. MATLAB and PYTHON scripts were developed to convert a clinical VMAT DICOM plan into BEAMnrc input files. The BEAMnrc installation was optimized by running the VMAT simulations through profiling tools which indicated the behavior of the constituent routines in the code, e.g. the bremsstrahlung splitting routine, and the specified random number generator. This information aided in determining the most efficient compiling parallel configuration for the specific CPU's available on our cluster, resulting in the fastest VMAT simulation times. Our method was evaluated with calculations involving 108 – 109 particle histories which are sufficient to verify patient dose using VMAT.Results:Parallelization allowed the calculation of patient dose on the order of 10 – 15 hours with 100 parallel jobs. Due to the compiler optimization process, further speed increases of 23% were achieved when compared with the open‐source compiler BEAMnrc packages.Conclusion:Analysis of the BEAMnrc code allowed us to optimize the compiler configuration for VMAT dose calculations. In future work, the optimized MC code, in conjunction with the parallel processing capabilities of BEAMnrc, will be applied to provide accurate and efficient secondary MU checks.

A New Approach of Dynamic Load Balancing Scheduling Algorithm for Homogeneous Multiprocessor System

Load balancing of parallel tasks in homogeneous as well as heterogeneous systems are major trouble for researches of both industry and academia. Load balancing were broadly classified into two categories namely static load balancing and dynamic load balancing. Nowadays, numerous of researches focus towards dynamic load balancing schemes for multiprocessor system. In this paper, the authors propose a dynamic load balancing strategy for homogeneous multiprocessor system and apply on cube based interconnection network named as Folded Crossed Cube network. The performance of folded crossed cube network gives better result in terms of diameter. Experimental results show that lesser load imbalance factor has been achieved along with execution time. By this algorithm, parallel jobs are solved with largest number of tasks. The merit of this algorithm is that when the number of tasks increases, the execution time decreases with lesser load imbalance factor.

A highly scalable parallel algorithm for maximally informative k-itemset mining

The discovery of informative itemsets is a fundamental building block in data analytics and information retrieval. While the problem has been widely studied, only few solutions scale. This is particularly the case when (1) the data set is massive, calling for large-scale distribution, and/or (2) the length k of the informative itemset to be discovered is high. In this paper, we address the problem of parallel mining of maximally informative k-itemsets (miki) based on joint entropy. We propose PHIKS (Parallel Highly Informative $$\underline{K}$$ -ItemSet), a highly scalable, parallel miki mining algorithm. PHIKS renders the mining process of large-scale databases (up to terabytes of data) succinct and effective. Its mining process is made up of only two efficient parallel jobs. With PHIKS, we provide a set of significant optimizations for calculating the joint entropies of miki having different sizes, which drastically reduces the execution time, the communication cost and the energy consumption, in a distributed computational platform. PHIKS has been extensively evaluated using massive real-world data sets. Our experimental results confirm the effectiveness of our proposal by the significant scale-up obtained with high itemsets length and over very large databases.

A Generalization of Multiple Choice Balls-into-Bins: Tight Bounds

This paper investigates a general version of the multiple choice model called the (k, d)-choice process in which n balls are assigned to n bins. In the process, $$kn$$m>n balls are placed into n bins is also presented for the case $$d \ge 2k$$d?2k. Potential applications are discussed such as distributed storage as well as parallel job scheduling in a cluster.

Opportunistic Spinlocks

With increasing demand for big-data processing and faster in-memory databases, cloud providers are moving towards large virtualized instances besides focusing on the horizontal scalability. However, our experiments reveal that such instances in popular cloud services (e.g., 32 vCPUs with 208 GB supported by Google Compute Engine) do not achieve the desired scalability with increasing core count even with a simple, embarrassingly parallel job (e.g., Linux kernel compile). On a serious note, the internal synchronization scheme (e.g., paravirtualized ticket spinlock) of the virtualized instance on a machine with higher core count (e.g., 80-core) dramatically degrades its overall performance. Our finding is different from the previously well-known scalability problem (i.e., lock contention problem) and occurs because of the sophisticated optimization techniques implemented in the hypervisor---what we call sleepy spinlock anomaly. To solve this problem, we design and implement OTICKET, a variant of paravirtualized ticket spinlock that effectively scales the virtualized instances in both undersubscribed and oversubscribed environments.

Work stealing for interactive services to meet target latency

Interactive web services increasingly drive critical business workloads such as search, advertising, games, shopping, and finance. Whereas optimizing parallel programs and distributed server systems have historically focused on average latency and throughput, the primary metric for interactive applications is instead consistent responsiveness, i.e., minimizing the number of requests that miss a target latency. This paper is the first to show how to generalize work-stealing, which is traditionally used to minimize the makespan of a single parallel job, to optimize for a target latency in interactive services with multiple parallel requests. We design a new adaptive work stealing policy, called tail-control , that reduces the number of requests that miss a target latency. It uses instantaneous request progress, system load, and a target latency to choose when to parallelize requests with stealing, when to admit new requests, and when to limit parallelism of large requests. We implement this approach in the Intel Thread Building Block (TBB) library and evaluate it on real-world workloads and synthetic workloads. The tail-control policy substantially reduces the number of requests exceeding the desired target latency and delivers up to 58% relative improvement over various baseline policies. This generalization of work stealing for multiple requests effectively optimizes the number of requests that complete within a target latency, a key metric for interactive services.

ARLS: A MapReduce-based output analysis tool for large-scale simulations

As simulations are becoming popular in the analysis of the complex behavior of large-scale systems with immense inputs and outputs, there is an increasing demand to efficiently store, manage, and analyze massive simulation outputs. Hadoop and MapReduce have been used in various applications to speed up the process of analyzing large amounts of datasets. In this paper, we present ARLS (After-action Reviewer for Large-scale Simulations), a MapReduce-based output analysis tool for simulation outputs. ARLS clusters distributed storages using Hadoop and automatically composes Map and Reduce functions to process the simulation outputs. ARLS has been applied to our SAM (Surface-to-Air Missile) simulator. The SAM simulator has been developed to analyze the dynamics of a missile in designing air-defense systems. ARLS takes a large amount of unstructured simulation outputs from SAM simulator, automatically generates Map and Reduce functions to analyze the missile and the aircraft component of SAM simulator, and executes Map and Reduce jobs in parallel. The results of our experiments show that ARLS can efficiently analyze a large amount of unstructured simulation datasets by distributing datasets and computations over the cluster of commodity machines.

An adaptive algorithm for scheduling parallel jobs in meteorological Cloud

From traditional clusters to cloud systems, job scheduling is one of the most critical factors for achieving high performance in any distributed environment. In this paper, we propose an adaptive algorithm for scheduling modular non-linear parallel jobs in meteorological Cloud, which has a unique parallelism that can only be configured at the very beginning of the execution. Different from existing work, our algorithm takes into account four characteristics of the jobs at the same time, including the average execution time, the deadlines of jobs, the number of assigned resources, and the overall system loads. We demonstrate the effectiveness and efficiency of our scheduling algorithm through simulations using WRF (Weather Research and Forecasting model) that which is widely used in scientific computing. Our evaluation results show that the proposed algorithm has multiple advantages compared with previous methods, including more than 10% reduction in terms of execution time, a higher completion ratio in terms of meeting soft deadlines, and a much smaller standard deviation of the average weighted execution time. Moreover, we show that the proposed algorithm can tolerate inaccuracy in system load estimation.

Boosting performance of a Statistical Machine Translation system using dynamic parallelism

In this work we introduce a new Statistical Machine Translation (SMT) system whose main objective is to reduce the translation times exploiting efficiently the computing power of the current processors and servers. Our system processes each individual job in parallel using different number of cores in such a way that the level of parallelism for each job changes dynamically according to the load of the translation server. In addition, the system is able to adapt to the particularities of any hardware platform used as server thanks to an autotuning module. An exhaustive performance evaluation considering different scenarios and hardware configurations demonstrates the benefits and flexibility of our proposal.

Comparative Study of Parallel Scheduling Algorithm for Parallel Job

Job scheduling is a technique which is applied on parallel computing systems, whose main focus is to measure the parameters of a system. For job scheduling menu algorithms are used in technical line such as priority based FCFS reservation, backfilling, improved round robin scheduling all are used to measure the parameters of a parallel computing. Since they all have some limitations and advantages to use all processors equally .This paper describes the various job scheduling algorithm such as priority based FCFS, PFCFS1, PFCFS2, PFCFS3 of static job scheduling algorithm. General Terms Parallel Computing, Scheduling algorithm

Communication-Load Impact on the Performance of Processor Allocation Strategies in 2-D Mesh Multicomputer Systems

A number of processor allocation strategies have been proposed in literature. A key performance factor that can highlight the difference between these strategies is the amount of communication conducted between the parallel jobs to be allocated. This paper aims to identifying how the density and pattern of communication can affect the performance of these strategies. Compared to the work already presented in literature, we examined a wider range of communication patterns. Other works consider only two types of communication patterns; those are the one-to-all and all-to-all patterns. This paper used the C language to implement the different allocation strategies, and then combined them with the ProcSimity simulation tool. The processor-allocation strategies are examined under the First-Come-First-Serve scheduling strategy. Results show that communication pattern and load are effective factors that have a significant impact on the performance of the processor allocation strategy used.

Optimisation of the usage of LHC and local computing resources in a multidisciplinary physics department hosting a WLCG Tier-2 centre

We present the approach of the University of Milan Physics Department and the local unit of INFN to allow and encourage the sharing among different research areas of computing, storage and networking resources (the largest ones being those composing the Milan WLCG Tier-2 centre and tailored to the needs of the ATLAS experiment). Computing resources are organised as independent HTCondor pools, with a global master in charge of monitoring them and optimising their usage. The configuration has to provide satisfactory throughput for both serial and parallel (multicore, MPI) jobs. A combination of local, remote and cloud storage options are available. The experience of users from different research areas operating on this shared infrastructure is discussed. The promising direction of improving scientific computing throughput by federating access to distributed computing and storage also seems to fit very well with the objectives listed in the European Horizon 2020 framework for research and development.

Design, Results, Evolution and Status of the ATLAS Simulation at Point1 Project

During the LHC Long Shutdown 1 (LSI) period, that started in 2013, the Simulation at Point1 (Sim@P1) project takes advantage, in an opportunistic way, of the TDAQ (Trigger and Data Acquisition) HLT (High-Level Trigger) farm of the ATLAS experiment. This farm provides more than 1300 compute nodes, which are particularly suited for running event generation and Monte Carlo production jobs that are mostly CPU and not I/O bound. It is capable of running up to 2700 Virtual Machines (VMs) each with 8 CPU cores, for a total of up to 22000 parallel jobs. This contribution gives a review of the design, the results, and the evolution of the Sim@P1 project, operating a large scale OpenStack based virtualized platform deployed on top of the ATLAS TDAQ HLT farm computing resources. During LS1, Sim@P1 was one of the most productive ATLAS sites: it delivered more than 33 million CPU-hours and it generated more than 1.1 billion Monte Carlo events. The design aspects are presented: the virtualization platform exploited by Sim@P1 avoids interferences with TDAQ operations and it guarantees the security and the usability of the ATLAS private network. The cloud mechanism allows the separation of the needed support on both infrastructural (hardware, virtualization layer) and logical (Grid site support) levels. This paper focuses on the operational aspects of such a large system during the upcoming LHC Run 2 period: simple, reliable, and efficient tools are needed to quickly switch from Sim@P1 to TDAQ mode and back, to exploit the resources when they are not used for the data acquisition, even for short periods. The evolution of the central OpenStack infrastructure is described, as it was upgraded from Folsom to the Icehouse release, including the scalability issues addressed.

Local search to improve coordinate-based task mapping

Local search algorithm that improves on task mapping algs for stencil patterns.Algorithm shown to reduce total running time and running time variability.Improvement shown to depend on the allocation algorithm used.Number of swaps made shown to be reasonable in practice. We present a local search strategy to improve the coordinate-based mapping of a parallel job's tasks to the MPI ranks of its parallel allocation in order to reduce network congestion and the job's communication time. The goal is to reduce the number of network hops between communicating pairs of ranks. Our target is applications with a nearest-neighbor stencil communication pattern running on mesh systems with non-contiguous processor allocation, such as Cray XE and XK Systems. Using the miniGhost mini-app, which models the shock physics application CTH, we demonstrate that our strategy reduces application running time while also reducing the runtime variability. We further show that mapping quality can vary based on the selected allocation algorithm, even between allocation algorithms of similar apparent quality.

Graph colouring-based algorithm to parallel jobs scheduling on parallel factories

This paper studies the multi-factory scheduling problem where a number of different individual factories join together to form a production network. In this problem, it is assumed each job has an original factory that initially ordered it, jobs can transport between the different factories in overloaded situations and the speed of machines in each factory can be different. Here, it is also assumed that the jobs may use more than one machine at the same time which is known as parallel jobs scheduling. This research extends a two-phase algorithm. In the first phase, the jobs according to need of machines decompose into the several job sets via semidefinite programming and a rounding algorithm in the graph. Then, in the second phase for scheduling the job sets obtained from first phase, we propose a discrete particle swarm optimisation (PSO). Finally, an extensive numerical study is conducted to assess the performance of proposed algorithm.

Improved Algorithms for Online Scheduling of Malleable Parallel Jobs on Two Identical Machines

This paper addresses online scheduling of malleable parallel jobs to minimize the maximum completion time, i.e., makespan. It is assumed that the execution time of a job Jj with processing time pj is pj/k + (k-1)c if the job is assigned to k machines, where c > 0 is a constant setup time. We consider online algorithms for the scheduling problem on two identical machines. Namely, the job Jj can be processed on one machine with execution time pj or alternatively two machines in parallel with execution time pj/2+c. For the asymptotical competitive ratio, we provide an improved online algorithm with makespan no more than (3/2)C* +c/2, where C* is the optimal makespan. For the strict competitive ratio, we propose an online algorithm with competitive ratio of 1.54, which is close to the lower bound of 1.5.

Combining performance and priority for scheduling resizable parallel applications

We illustrate and evaluate the potential impact of dynamic resizability on parallel job scheduling. Our ReSHAPE framework includes a job scheduler that supports dynamic resizing of malleable parallel applications. We propose and evaluate new scheduling policies and strategies enabled by the ReSHAPE framework. These strategies use both application performance and user-assigned priorities to inform decisions about expanding or contracting the set of processors assigned to a particular job. Experimental results show that the scheduling policies significantly improve individual job turn around time as well as overall cluster utilization.

Distributed CTL model checking using MapReduce: theory and practice

SummaryThe recent extensive availability of ‘cloud’ computing platforms is very appealing for the formal verification community. In fact, these platforms represent a great opportunity to run massively parallel jobs and analyze ‘big data’ problems, although classical formal verification tools and techniques must undergo a deep technological transformation to take advantage of the available powerful architectures. A distributed approach to verification of computation tree logic formulas on very large state spaces is described. The approach exploits and integrates our parametric state–space builder, designed to ease the adoption of ‘big data’ platforms. The whole framework adopts a MAPREDUCEapproach as the core computational model and can be tailored to different modeling formalisms. This paper includes proofs of correctness, a short theoretical discussion about complexity, and reports a practical experience with some benchmarking Petri net models. The outcomes of several tests are presented, thus showing the convenience of the proposed approach. Copyright © 2015 John Wiley & Sons, Ltd.