Teaduslikus arvutusprotsessis riistvara virtualiseerimise hind
Date
2013
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Tartu Ülikool
Abstract
Selle töö eesmärk on uurida riistvara virtualiseerimise negatiivseid aspekte, kui paigaldatakse rakendusi pilve, mõõtes selle mõju täpselt teaduslike paralleelarvutus algoritmidega.
Virtualiseerimine annab pilveteenustele mitmeid eeliseid nagu seadistamise lihtsus, riistvara ja tarkvara lahtisidestus, väga kiire paigaldus ja konfiguratsiooni muutus ning elastsus. Kuid lisa virtualisatsioonikiht võib endaga kaasa tuua mitmeid puuduseid. Eriti ressursi nõudlikele teaduslikele algoritmidele, mis rakendavad paralleelarvutus tehnoloogiaid.
Selleks kasutame NASA välja töötatud spetsiaalset tarkvara paralleelsete süsteemide jõudlustestimiseks – „NAS Parallel Benchmarking“, mis töötab MPI tehnoloogial. Virtualiseerimiseks kasutada XEN ja KVM vabavaralist virtualiseerimistarkvara ning operatsioonisüsteemiks kasutada samuti vabavaralist „Ubuntu Linuxit“.
Laiame, et lihtsalt virtualisatsioonikihti lisades ei ole arvutusvõimsusele erilist mõju, küll aga olenevalt arvutite arvust, võib virtualiseerimine oluliselt mõjutada kõvaketta operatsioonide kiirust ja samuti on tuntav mõju võrgulatentsusele. Kokkuvõttes töötab KVM peaaegu kõikides jõudlustestides paremini kui Xen.
The goal of this thesis is to research what is the downside of using infrastructure virtualization for deploying applications on the cloud and to accurately measure its effect on parallel scientific computing algorithms. Virtualization provides numerous benefits for clouds such as ease of configuring, de-coupling machine and software stack, rapid deployment and configuration changes, elasticity etc. However additional virtualization layer introduces several disadvantages. Especially for resource demanding scientific algorithms that utilize parallel computing techniques. For this we deploy benchmarking algorithms designed to test distributed computing on different platforms such as NASA Parallel Benchmarking software. For virtualization we use XEN and KVM and for operating system we use Ubuntu. This thesis concludes that just by adding a virtualization layer, the computing power is not affected but depending on the number of machines the impact on the disk operations might be severe. Also networking capabilities are reduced for virtual machines. All in all KVM is better than Xen in almost all of the benchmarks.
The goal of this thesis is to research what is the downside of using infrastructure virtualization for deploying applications on the cloud and to accurately measure its effect on parallel scientific computing algorithms. Virtualization provides numerous benefits for clouds such as ease of configuring, de-coupling machine and software stack, rapid deployment and configuration changes, elasticity etc. However additional virtualization layer introduces several disadvantages. Especially for resource demanding scientific algorithms that utilize parallel computing techniques. For this we deploy benchmarking algorithms designed to test distributed computing on different platforms such as NASA Parallel Benchmarking software. For virtualization we use XEN and KVM and for operating system we use Ubuntu. This thesis concludes that just by adding a virtualization layer, the computing power is not affected but depending on the number of machines the impact on the disk operations might be severe. Also networking capabilities are reduced for virtual machines. All in all KVM is better than Xen in almost all of the benchmarks.