GPU Accelerated Machine Learning
Author
Simonsen, Ulf Gaarde
Term
4. term
Education
Publication year
2017
Submitted on
2017-01-13
Pages
33
Abstract
This thesis investigates how GPUs can be used to scale and accelerate machine learning in big-data settings, with particular focus on integration with Apache Flink, which currently lacks GPU support. It reviews key background (MapReduce, Spark and Flink) and contrasts CPU and GPU architectures to identify the classes of algorithms suited to data-parallel execution. Related work, including HeteroSpark, is surveyed. On this basis, the thesis proposes an architecture for a GPU-enabled big-data framework, clarifying which components are provided by the framework versus the user, and introduces a model for automatic adaptation to heterogeneous cluster hardware. A general-purpose GPU machine learning library with a User API and backend is implemented. Local experiments show that GPUs can deliver speedups when datasets are sufficiently large, while also introducing overhead. Due to time constraints, integration and evaluation within a full big-data framework were not completed; instead, theoretical assessments of such tests are provided. The main contributions are (1) a conceptual model for GPU support in big-data frameworks and (2) a working GPU-based machine learning framework with API and backend, with directions for future work toward full Flink integration.
Dette speciale undersøger, hvordan GPU’er kan bruges til at skalere og accelerere maskinlæringsopgaver i en big data‑kontekst, med særligt fokus på integration med Apache Flink, som ikke har GPU-understøttelse. Arbejdet gennemgår centrale grundbegreber (MapReduce, Spark og Flink) og sammenligner CPU- og GPU-arkitektur for at identificere de algoritmeklasser, der egner sig til data-parallel udførelse. Relateret arbejde, bl.a. HeteroSpark, gennemgås. På den baggrund foreslås en arkitektur for et GPU-udvidet big data‑rammeværk, herunder afklaring af hvilke dele der håndteres af rammeværket versus brugeren, samt en model for automatisk tilpasning til heterogent cluster-hardware. Der implementeres et generelt GPU-baseret maskinlæringsbibliotek med et bruger-API og en backend. Lokale eksperimenter viser, at GPU’er kan give hastighedsgevinster, når datasættene er tilstrækkeligt store, men at der også introduceres overhead. På grund af tidsbegrænsninger er integration og test i et egentligt big data‑setup ikke gennemført; i stedet gives teoretiske vurderinger af forventet adfærd. Specialets hovedbidrag er (1) en konceptuel model for GPU-understøttelse i big data‑rammeværker og (2) en fungerende GPU-baseret maskinlæringsramme med API og backend, med forslag til fremtidigt arbejde mod fuld Flink-integration.
[This apstract has been generated with the help of AI directly from the project full text]