Hasil Pencarian  ::  Simpan CSV :: Kembali

Abstrak :
The proceedings of the 5th International Workshop on Parallel Tools for High Performance Computing provide an overview on supportive software tools and environments in the fields of system management, parallel debugging and performance analysis. In the pursuit to maintain exponential growth for the performance of high performance computers the HPC community is currently targeting exascale systems. The initial planning for exascale already started when the first petaflop system was delivered. Many challenges need to be addressed to reach the necessary performance. Scalability, energy efficiency and fault-tolerance need to be increased by orders of magnitude. The goal can only be achieved when advanced hardware is combined with a suitable software stack. In fact, the importance of software is rapidly growing. As a result, many international projects focus on the necessary software.

Abstrak :
This book presents the state-of-the-art in simulation on supercomputers. Leading researchers present results achieved on systems of the Gauss-Allianz, the association of High-Performance Computing centers in Germany. The reports cover all fields of computational science and engineering, ranging from CFD to Computational Physics and Biology to Computer Science, with a special emphasis on industrially relevant applications. Presenting results for large-scale parallel microprocessor-based systems and GPU and FPGA-supported systems, the book makes it possible to compare the performance levels and usability of various architectures. Its outstanding results in achieving the highest performance for production codes are of particular interest for both scientists and engineers. The book includes a wealth of color illustrations and tables.

Abstrak :
This proceedings volume contains a selection of papers presented at the Fourth International Conference on High performance scientific computing held at the Hanoi Institute of Mathematics, Vietnamese Academy of Science and Technology (VAST), March 2-6, 2009. The conference was organized by the Hanoi Institute of Mathematics, the Interdisciplinary Center for Scientific Computing (IWR), Heidelberg, and its Heidelberg Graduate School of Mathematical and Computational Methods for the Sciences, and Ho Chi Minh City University of Technology. The contributions cover the broad interdisciplinary spectrum of scientific computing and present recent advances in theory, development of methods, and applications in practice. Subjects covered are mathematical modelling, numerical simulation, methods for optimization and control, parallel computing, software development, applications of scientific computing in physics, mechanics, biology and medicine, engineering, hydrology problems, transport, communication networks, production scheduling, industrial and commercial problems.

Abstrak :
Contents : - Part I:Overview - High-Speed I/O Design and Test Review:From the Perspectives of Moore is Law and Multiple Gbps Data Rates - Part II:System Architecture and Performance - Transfer Functions for the Reference Clock Jitter in a Serial Link: Theory and Applications - Channel Compliance Testing Utilizing Novel Statistical Eye Methodology - Advances in High-Speed Design in Dispersively Attenuating Environments Such as Cables and Backplanes - Part III:Design Simulation and Modeling - Static Crosstalk Analysis - Modeling Loss and Jitter in High-Speed Serial Connects - Design and Modeling Methodology for High-Performance Power Distribution Systems - Source Synchronous Bus Design and Timing Analysis for High-Volume Manufacturable System Interconnects - Part IV:Design for High Performance - Eye Opening Techniques Enabled by Dispersion Compensation - Maximizing 10-Gbps Transmission Path Length in Copper Backplanes with and without Transceiver Technology - How to Make Optimal Use of Signal Conditioning in 40-Gbps Copper Interconnects - Design of a 6.25-Gbps Backplane SerDes with Top-Down Design Methodology - A Flexible Serial Link for 5-10 Gbps in Realistic Backplane Environments - Part V:Characterization and Test - Signal Integrity and Jitter:How to Measure Them Correctly - A Statistical and System Transfer Function Based Method for Jitter and Noise in Communication Design and Test - Total Jitter Measurement at Low Probability Levels,Using the Optimized BERT Scan Method - Comparison and Correlation of Signal-Integrity Measurement Techniques - Performance Evaluation of High-Speed Serial Links - Physical-Layer Design and Characterization of a 3.2 Gbps/Pair Memory Channel - Acronym Guide

Abstrak :
ABSTRAK
Teknologi di bidang perangkat lunak dan perangkat keras semakin berkembang cepat. Masalah keterbatasan kapasitas suatu komputer memicu berkembangnya sebuah inovasi yang disebut dengan High Performance Computing HPC . HPC merupakan sekumpulan komputer yang digabungkan dalam sebuah jaringan dan dikoordinasi oleh software khusus. Cloud Computing merupakan paradigma yang relatif baru dalam bidang komputasi. Pada penelitian ini dilakukan pengujian terhadap performansi High Performance Computing Cluster HPCC berbasis cloud menggunakan layanan OpenStack dalam menjalankan fungsi dasar Message Passing Interface. Pengujian dilakukan menggunakan program Mpptest dan SIMPLE-O. Penggunaan server yang tidak mendukung hypervisor KVM pada pengujian point-to-point communication dapat menurunkan performansi HPCC berbasis cloud sebesar 3,1 - 12,4 dibandingkan dengan HPCC berbasis non-cloud. Pada pengujian point-to-point communication dengan 2 server yang mendukung hypervisor KVM, HPCC berbasis cloud unggul dibandingkan HPCC berbasis non-cloud sebesar 1,6 ndash; 2,7 . Pada pengujian performansi HPCC dalam melakukan fungsi MPI collective communication tidak ditemukan perbedaan berarti antara kedua cluster dimana HPCC berbasis non-cloud mengungguli HPCC berbasis cloud sebesar 0 - 1,4 . Pada pengujian menggunakan program SIMPLE-O didapati performansi HPCC berbasis cloud dan non-cloud imbang jika semua instance dijalankan dengan server yang mendukung hypervisor KVM, apabila terdapat instance yang dijalankan server tanpa dukungan KVM maka HPCC berbasis non-cloud unggul 96,2 dibandingkan HPCC berbasis cloud. Ketersedian modul KVM pada server yang menjadi host suatu instance sangat berpengaruh terhadap performansi HPCC berbasis cloud.

---

ABSTRACT
Software and hardware technologies have been developing rapidly. Capacity limation problems found in computers triggered a development of a new innovation called High Performance Computing HPC . HPC is a cluster of computers in a network coordinated by a special software. Cloud Computing is a new paradigm in computation field. In this research, series of test are done to find out the performance of cloud and non cloud based High Performance Computing Cluster HPCC while running basic functions of Message Passing Interface. Tests are done using Mpptest and SIMPLE O program. By using a server that does not support KVM in point to point communication test could decrease the performance of cloud based HPCC by 3,1 to 12,4 compared to non cloud based HPCC. During the test of point to point communication using 2 servers that support KVM hypervisor, cloud based HPCC is ahead of non cloud based HPCC by 1,6 to 2,7 . During the test of collective communication, there are no significant differences between performances of the two cluster, with non cloud based HPCC is ahead by 0 to 1,4 compared to cloud based HPCC. During the test using SIMPLE O program, the two cluster is even in term of performance as long as every instance is run by servers that support KVM hypervisor, if there is an instance that is run by a server that does not support KVM hypervisor then the performance of non cloud based HPCC is still ahead by 96,2 compared to cloud based HPCC. During the performance testing of HPCC while running collective communication, noticable performance difference between cloud and non cloud based HPCC was not found. The availability of KVM module in a server that is used to host an instance is really essential to the cloud based HPCC performance.

Abstrak :
Proses training model membutuhkan sumber daya komputasi yang akan terus meningkat seiring dengan bertambahnya jumlah data dan jumlah iterasi yang telah dicapai. Hal ini dapat menimbulkan masalah ketika proses training model dilakukan pada lingkungan komputasi yang berbagi sumber daya seperti pada infrastruktur komputasi berbasis klaster. Masalah yang ditimbulkan terutama terkait dengan efisiensi, konkurensi, dan tingkat utilisasi sumber daya komputasi. Persoalan efisiensi muncul ketika sumber daya komputasi telah tersedia, tetapi belum mencukupi untuk kebutuhan job pada antrian ter- atas. Akibatnya sumber daya komputasi tersebut menganggur. Penggunaan sumber daya tersebut menjadi tidak efisien karena terdapat kemungkinan sumber daya tersebut cukup untuk mengeksekusi job lain pada antrian. Selain itu, pada cluster computing juga mem- butuhkan sistem monitoring untuk mengawasi dan menganalisis penggunaan sumber daya pada klaster. Penelitian ini bertujuan untuk menemukan resource manager yang sesuai untuk digunakan pada klaster komputasi yang memiliki GPU agar dapat meningkatkan efisiensi, implementasi sistem monitoring yang dapat membantu analisis penggunaan sumber daya sekaligus monitoring proses komputasi yang sedang dijalankan pada klaster, dan melayani inference untuk model machine learning. Penelitian dilakukan dengan cara menjalankan eksperimen penggunaan Slurm dan Kubernetes. Hasil yang diperoleh adalah Slurm dapat memenuhi kebutuhan untuk job scheduling dan mengatur penggunaan GPU dan resources lainnya pada klaster dapat digunakan oleh banyak pengguna sekaligus. Sedangkan untuk sistem monitoring, sistem yang dipilih adalah Prometheus, Grafana, dan Open OnDemand. Sementara itu, sistem yang digunakan untuk inference model adalah Flask dan Docker. ...... The amount of computational power needed for the model training process will keep rising along with the volume of data and the number of successful iterations. When the model training process is conducted in computing environments that share resources, such as on cluster-based computing infrastructure, this might lead to issues. Efficiency, competition, and the level of resource use are the three key issues discussed.Efficiency problems occur when there are already computing resources available, yet they are insufficient to meet the demands of high-level workloads. The power of the machine is subsequently wasted. The utilization of such resources becomes inefficient because it’s possible that they would be adequate to complete other tasks on the front lines. A monitoring system is also necessary for cluster computing in order to track and assess how resources are used on clusters. The project seeks to set up a monitoring system that can assist in analyzing the usage of resources while monitoring the com- puting processes running on the cluster and locate a suitable resource manager to be utilized on a computing cluster that has a GPU in order to increase efficiency, also serve inference model in production. Slurm and Kubernetes experiments were used to conduct the investigation. The findings show that Slurm can handle the demands of job scheduling, manage the utilization of GPUs, and allow for concurrent use of other cluster resources. Prometheus, Grafana, and Open OnDemand are the chosen moni- toring systems. Else, inference model is using Flask and Docker as its system constructor.

Abstrak :
Proses training model membutuhkan sumber daya komputasi yang akan terus meningkat seiring dengan bertambahnya jumlah data dan jumlah iterasi yang telah dicapai. Hal ini dapat menimbulkan masalah ketika proses training model dilakukan pada lingkungan komputasi yang berbagi sumber daya seperti pada infrastruktur komputasi berbasis klaster. Masalah yang ditimbulkan terutama terkait dengan efisiensi, konkurensi, dan tingkat utilisasi sumber daya komputasi. Persoalan efisiensi muncul ketika sumber daya komputasi telah tersedia, tetapi belum mencukupi untuk kebutuhan job pada antrian ter- atas. Akibatnya sumber daya komputasi tersebut menganggur. Penggunaan sumber daya tersebut menjadi tidak efisien karena terdapat kemungkinan sumber daya tersebut cukup untuk mengeksekusi job lain pada antrian. Selain itu, pada cluster computing juga mem- butuhkan sistem monitoring untuk mengawasi dan menganalisis penggunaan sumber daya pada klaster. Penelitian ini bertujuan untuk menemukan resource manager yang sesuai untuk digunakan pada klaster komputasi yang memiliki GPU agar dapat meningkatkan efisiensi, implementasi sistem monitoring yang dapat membantu analisis penggunaan sumber daya sekaligus monitoring proses komputasi yang sedang dijalankan pada klaster, dan melayani inference untuk model machine learning. Penelitian dilakukan dengan cara menjalankan eksperimen penggunaan Slurm dan Kubernetes. Hasil yang diperoleh adalah Slurm dapat memenuhi kebutuhan untuk job scheduling dan mengatur penggunaan GPU dan resources lainnya pada klaster dapat digunakan oleh banyak pengguna sekaligus. Sedangkan untuk sistem monitoring, sistem yang dipilih adalah Prometheus, Grafana, dan Open OnDemand. Sementara itu, sistem yang digunakan untuk inference model adalah Flask dan Docker. ...... The amount of computational power needed for the model training process will keep rising along with the volume of data and the number of successful iterations. When the model training process is conducted in computing environments that share resources, such as on cluster-based computing infrastructure, this might lead to issues. Efficiency, competition, and the level of resource use are the three key issues discussed.Efficiency problems occur when there are already computing resources available, yet they are insufficient to meet the demands of high-level workloads. The power of the machine is subsequently wasted. The utilization of such resources becomes inefficient because it’s possible that they would be adequate to complete other tasks on the front lines. A monitoring system is also necessary for cluster computing in order to track and assess how resources are used on clusters. The project seeks to set up a monitoring system that can assist in analyzing the usage of resources while monitoring the com- puting processes running on the cluster and locate a suitable resource manager to be utilized on a computing cluster that has a GPU in order to increase efficiency, also serve inference model in production. Slurm and Kubernetes experiments were used to conduct the investigation. The findings show that Slurm can handle the demands of job scheduling, manage the utilization of GPUs, and allow for concurrent use of other cluster resources. Prometheus, Grafana, and Open OnDemand are the chosen moni- toring systems. Else, inference model is using Flask and Docker as its system constructor.

Abstrak :
ABSTRACT In this paper after a short theoretical introduction about modern techniques used in parallel computing, we report a case study related to the design and development of the Caliban Linux High Performance Computing cluster, carried out by the author in the High Performance Computing Laboratory of the University of L Aquila. Finally we report some performance evaluation tests related to the Caliban cluster performed using HPL (High Performance Linpack) benchmarks.

Abstrak :
ABSTRAK
Pada skripsi ini telah dibangun infrastruktur cloud dengan menggunakan Openstack platform. Openstack menjanjikan infrastruktur yang scalable yang menjadikan platform cloud ini digemari banyak pengguna cloud. Tujuan dari skripsi ini adalah untuk mempelajari kinerja jaringan OpenStack berdasarkan implementasi Neutron dan memberikan rekomendasi rancangan jaringan optimal untuk integrasi high performance computing. Parameter kinerja jaringan seperti throughput, packet loss dan latency akan dievaluasi berdasarkan transmisi data TCP dan UDP dengan menggunakan alat benchmark IPerf. Hasil dari eksperimen menunjukkan kinerja Openstack hampir tidak memiliki hambatan dalam penggunaan bandwidth jaringan. Hasil penelitian juga menunjukkan bahwa lokasi virtual machine di instansiasi dan alamat jaringan akan mempengaruhi kinerja jaringan. Skripsi ini juga akan membahas alur yang digunakan untuk menganalisa perbedaan hasil kinerja jaringan virtual machine pada cloud serta menampilkan hasil pengolahan data kinerja jaringan yang kemudian akan memberikan rancangan cloud yang optimal untuk integrasi high performance computing.

---

ABSTRACT
In this thesis, a cloud infrastructure is built using Openstack platform. Openstack promises a scalable infrastructure that makes this platform as favourite for cloud users. The purpose of this thesis is to study the performance of Openstack network based on the implementation of Neutron to provide recommendations of the optimal network design for the integration of high performance computing. Network performance parameters such as throughput, packet loss and latency will be evaluated based on TCP and UDP transmission data using IPerf benchmarking tool. The result of the experiments show that Openstack performance have no network bandwidth bottleneck. The result also show that the location where virtual machine is instantiate and network address will affect network performance. This thesis will also discuss the flow used to analyze the differences in virtual machine network performance results in the cloud and display the results of virtual machine network performance which will then provide an optimal cloud design for the integration of high performance computing.

Abstrak :
Pada penelitian ini, dikembangkan HPC yang menerapkan multicore processing pada program Sistem Pendeteksi Plagiarisme dengan memanfaatkan infrastruktur komputasi awan berbasis OpenStack. Sistem Pendeteksi Plagiarisme merupakan program yang dikembangkan untuk mendeteksi tingkat plagiarisme dari suatu karya ilmiah. Algoritma program yang digunakan untuk penelitian kali ini adalah latent semantic analysis (LSA). Implementasi HPC dilakukan dengan bantuan library OpenMP yang didesain untuk bahasa pemrograman C. Diterapkan dua jenis paralelisme pada program, yaitu paralelisme fungsi dan paralelisme data. Setelah dilakukan pengujian, didapati hasil bahwa kedua metode paralelisme ini mempercepat eksekusi program. Paralelisme fungsi mempercepat waktu eksekusi hingga sebesar 1,03 kali waktu eksekusi serial dan paralelisme data mempercepat waktu eksekusi hingga 1,34 kali waktu eksekusi serial.

---

In this research, HPC with multicore processing is developed on Plagiarism Detection System using OpenStack based cloud computing infrastructure. Plagiarism Detection System is a software developed to detect plagiarism level of a scientific papers. The algorithm used in this program is latent semantic analysis (LSA). HPC implementation is done using OpenMP library which is designed to be used in C programming language. There are two types of paralelism in this program, which are function paralelism and data paralelism, both accelerate the execution time. Function paralelism accelerates program by up to 1,03 times of serial execution while data paralelism decreases the execution time by up to 1,34 times serial execution time.