A25: Investigating Performance of Serialization Methods for Networked Data Transfer in HPC Applications
Student: Max Yang (Georgia Institute of Technology)
Supervisor: Thomas Stitt (Lawrence Livermore National Laboratory)
Abstract: Cluster-to-user data transfers present challenges with cross-platform endianness (byte-order) compatibility and handling a variety of numeric types, and may occur over suboptimal network links. Two serialization libraries, Protocol Buffers and Conduit, were selected for their ability to handle endianness and their cross-language support, and their performance in both size and speed was measured. It was found that the throughput of Protocol Buffers was significantly more than that of Conduit while exhibiting less protocol overhead. Adding a compression stage after serialization dramatically reduced the size of messages on certain types of data, but had some impact on throughput.
ACM-SRC Semi-Finalist: no
Poster: pdf
Two-page extended abstract: pdf
Poster Index