Home // International Journal On Advances in Internet Technology, volume 11, numbers 3 and 4, 2018 // View article

DXNet: Scalable Messaging for Multi-Threaded Java-Applications Processing Big Data in Clouds

Authors:
Kevin Beineke
Stefan Nothaas
Michael Schoettner

Keywords: Message passing; Ethernet networks; InfiniBand; Java; Cloud computing.

Abstract:
Many big data and large-scale cloud applications are written in Java or are built using Java-based frameworks. Typically, application instances are running in a data center on many virtual machines which requires scalable and efficient network communication. In this paper, we present the practical experience of designing an extensible Java network subsystem, called DXNet, providing fast object de-/serialization, automatic connection management and zero-copy messaging. Additionally, we present a Java.nio-based transport for DXNet, called EthDXNet, to efficiently utilize Ethernet networks. The proposed design uses a zero-copy send and receive approach for asynchronous messages and requests/responses. DXNet is optimized for small messages (< 100 bytes) in order to support graph-based applications, but also works well with larger messages (e.g., 8 MB). DXNet is available on GitHub, and its modular design is open for different transport implementations currently supporting Ethernet and InfiniBand. EthDXNet is based on Java.nio socket channels complemented by application-level flow control to achieve low latency and high throughput for 10 GBit/s and faster Ethernet networks. Furthermore, a scalable automatic connection management and a low-overhead interest handling provide efficient network communication for dozens of servers, even for small messages (< 100 bytes) and an all-to-all communication pattern. The evaluation with micro-benchmarks and the Yahoo! Cloud Serving Benchmark (YCSB) shows the efficiency and scalability with up to 64 virtual machines in the Microsoft Azure cloud. Furthermore, DXNet achieves request-response latencies sub 10 µs (round trip) including object de-/serialization, as well as a maximum throughput of more than 9 GByte/s on a private cluster (using InfiniBand).

Pages: 117 to 135

Copyright: Copyright (c) to authors, 2018. Used with permission.

Publication date: December 30, 2018

Published in: journal

ISSN: 1942-2652