Dokument: High Performance Networking for Distributed Java Applications
| Titel: | High Performance Networking for Distributed Java Applications | |||||||
| URL für Lesezeichen: | https://docserv.uni-duesseldorf.de/servlets/DocumentServlet?id=65655 | |||||||
| URN (NBN): | urn:nbn:de:hbz:061-20240516-171545-1 | |||||||
| Kollektion: | Dissertationen | |||||||
| Sprache: | Englisch | |||||||
| Dokumententyp: | Wissenschaftliche Abschlussarbeiten » Dissertation | |||||||
| Medientyp: | Text | |||||||
| Autor: | Krakowski, Filip [Autor] | |||||||
| Dateien: |
| |||||||
| Beitragende: | Prof. Dr. Schöttner, Michael [Gutachter] Prof. Dr. Conrad, Stefan [Gutachter] | |||||||
| Stichwörter: | java,networking,distributed,high performance,panama,native,infiniband | |||||||
| Dokumententyp (erweitert): | Dissertation | |||||||
| Dewey Dezimal-Klassifikation: | 000 Informatik, Informationswissenschaft, allgemeine Werke » 004 Datenverarbeitung; Informatik | |||||||
| Beschreibung: | While data processing has become a core component of almost all systems in the digital world, some applications are designed for very large amounts of data or big data and must also be able to produce a result in a short time. Furthermore, big data applications have to be distributed, as a single computer is no longer sufficient due to its limited resources. Due to the ever-increasing complexity of data, they also need a suitable way to process it, which is offered in the form of various big data processing frameworks. Many of these big data frameworks have been developed within the Java ecosystem and the focus has therefore been placed on the Java Virtual Machine. This work pursues the goal of accelerating communication between distributed Java applications using modern technologies. Specifically, an integration of the InfiniBand technology developed by NVIDIA (formerly by Mellanox) within the Java ecosystem is being pursued. For this purpose, a connection to the Verbs API based on the Java Native Interface is first developed, which allows efficient access to InfiniBand hardware within Java applications. The developed solution shows very good results in benchmarks with throughputs of up to 14 million messages per second as well as reaching the maximum bandwidth achievable in practice of a 56 Gbit/s ConnectX-3 controller at 6 GB/s. Based on these findings and a cooperation with Oracle Labs, the focus is being shifted away from the Java Native Interface to the Foreign Function \& Memory API recently developed by Oracle. The resulting Infinileap project continues to pursue the goal of connecting InfiniBand hardware in Java applications, but relies on Oracle's Project Panama for native access and on the OpenUCX library, which provides an abstraction layer for high-performance networking hardware. It also aims at providing RDMA functions for Java applications. The benchmarks carried out in the context of the Infinileap project show that using Project Panama with OpenUCX is a good match. Using the Infinileap project and 100Gbit/s ConnectX-5 network controllers, the round-trip latency for send operations and RDMA read and write operations with small amounts of data of less than 256 bytes is around 2 microseconds. Similarly, atomic operations such as Compare \& Swap can be executed on remote memory in under 2 microseconds The round-trip latency for smaller messages of 8 bytes is even lower with 1.4 microseconds while the theoretically achievable minimum latency is 1.2 microseconds. The bandwidth of the network card is also reached at 100 Gbit/s when using larger messages starting at around 4 kilobytes. Finally, an integration into Java's NIO framework is developed, which allows existing network applications to use transparent InfiniBand controllers on the network level without major changes to the code. The benchmarks developed in this context show that the Infinileap project offers significant added value compared to other existing solutions such as JUCX and enables more stable operation overall. | |||||||
| Lizenz: |  Dieses Werk ist lizenziert unter einer Creative Commons Namensnennung 4.0 International Lizenz | |||||||
| Fachbereich / Einrichtung: | Mathematisch- Naturwissenschaftliche Fakultät » WE Informatik » Betriebssysteme | |||||||
| Dokument erstellt am: | 16.05.2024 | |||||||
| Dateien geändert am: | 16.05.2024 | |||||||
| Promotionsantrag am: | 11.01.2024 | |||||||
| Datum der Promotion: | 18.04.2024 |
