Tõhusad paralleel-algoritmid radarsatelliidipiltide töötluseks kasutades suuremahulisi hajusraamistikke
Abstract
Radarsatelliidipiltide töötlemine on märkimisväärse suurusega arvutusülesanne kuna piltide mõõtmed on äärmiselt suured. Hajusarvutust kasutatakse sageli et võimendada algoritme, mis jooksevad ühel arvutil liiga aeglaselt. Kuid on ebaselge, milliseid radaripiltide töötlusalgoritme on võimalik tõhusalt paralleelsetesse keskkondadesse ümber viia ning kuidas neid korrektselt implementeerida. Eelnevad tööd on keskendunud paralleelsele pilditöötlusele kui üldisele arvutusülesandele, kuid unikaalseid radarpiltide omadusi või uuemaid hajusarvutusraamistikke pole käsitletud või on käsitlus keskendunud mõnele üksikule algoritmile. Käesolev töö pakub välja potentsiaalselt paralleliseeritavate radaripiltide töötlusalgoritmide klassifikatsiooni. Iga algoritmide klassi uuritakse enimkasutatavate hajusraamistike ja -failisüsteemide omadustel. Kõige paremini mingeid klasse esindavad algoritmid implementeeritakse konkreetsetel tehnoloogiatel. Klassifikatsioon lihtsustab huvipakkuvate algoritmide võrdlust ja pakub üldisi implementatsioonisamme ning hõlbustab seeläbi hajusarvutuse rakendamist radarsatelliidipiltide töötlusel.
Processing radar satellite images is a considerable computing task due to large image sizes. Distributed computing can often be leveraged to speed up algorithms that are too time-consuming on a single machine. It is however unclear which radar image processing algorithms can be efficiently migrated to parallel environments and what is the proper way to implement them. Previous works have concentrated on parallel image processing as a general computing task but either the unique properties of radar images or newer distributed computing frameworks are not considered or only some specific algorithms have been examined. This thesis proposes a classification of radar image processing algorithms that can potentially be parallelized. Each class of algorithms is studied based on the properties of current popular distributed computing frameworks and file systems. Algorithms that best represent their respective classes are implemented using some concrete distributed computing framework. The classification simplifies the gauging of potential algorithms in terms of parallel speedup and provides general implementation steps, thus easing the task of leveraging distributed computing for radar image processing.
Processing radar satellite images is a considerable computing task due to large image sizes. Distributed computing can often be leveraged to speed up algorithms that are too time-consuming on a single machine. It is however unclear which radar image processing algorithms can be efficiently migrated to parallel environments and what is the proper way to implement them. Previous works have concentrated on parallel image processing as a general computing task but either the unique properties of radar images or newer distributed computing frameworks are not considered or only some specific algorithms have been examined. This thesis proposes a classification of radar image processing algorithms that can potentially be parallelized. Each class of algorithms is studied based on the properties of current popular distributed computing frameworks and file systems. Algorithms that best represent their respective classes are implemented using some concrete distributed computing framework. The classification simplifies the gauging of potential algorithms in terms of parallel speedup and provides general implementation steps, thus easing the task of leveraging distributed computing for radar image processing.