- Title
- GPU Accelerated protocol analysis for large and long-term traffic traces
- Creator
- Nottingham, Alastair Timothy
- ThesisAdvisor
- Irwin, Barry
- Date
- 2016
- Type
- Thesis
- Type
- Doctoral
- Type
- PhD
- Identifier
- http://hdl.handle.net/10962/910
- Identifier
- vital:20002
- Description
- This thesis describes the design and implementation of GPF+, a complete general packet classification system developed using Nvidia CUDA for Compute Capability 3.5+ GPUs. This system was developed with the aim of accelerating the analysis of arbitrary network protocols within network traffic traces using inexpensive, massively parallel commodity hardware. GPF+ and its supporting components are specifically intended to support the processing of large, long-term network packet traces such as those produced by network telescopes, which are currently difficult and time consuming to analyse. The GPF+ classifier is based on prior research in the field, which produced a prototype classifier called GPF, targeted at Compute Capability 1.3 GPUs. GPF+ greatly extends the GPF model, improving runtime flexibility and scalability, whilst maintaining high execution efficiency. GPF+ incorporates a compact, lightweight registerbased state machine that supports massively-parallel, multi-match filter predicate evaluation, as well as efficient arbitrary field extraction. GPF+ tracks packet composition during execution, and adjusts processing at runtime to avoid redundant memory transactions and unnecessary computation through warp-voting. GPF+ additionally incorporates a 128-bit in-thread cache, accelerated through register shuffling, to accelerate access to packet data in slow GPU global memory. GPF+ uses a high-level DSL to simplify protocol and filter creation, whilst better facilitating protocol reuse. The system is supported by a pipeline of multi-threaded high-performance host components, which communicate asynchronously through 0MQ messaging middleware to buffer, index, and dispatch packet data on the host system. The system was evaluated using high-end Kepler (Nvidia GTX Titan) and entry level Maxwell (Nvidia GTX 750) GPUs. The results of this evaluation showed high system performance, limited only by device side IO (600MBps) in all tests. GPF+ maintained high occupancy and device utilisation in all tests, without significant serialisation, and showed improved scaling to more complex filter sets. Results were used to visualise captures of up to 160 GB in seconds, and to extract and pre-filter captures small enough to be easily analysed in applications such as Wireshark.
- Format
- 339 leaves, pdf
- Publisher
- Rhodes University, Faculty of Science, Computer Science
- Language
- English
- Rights
- Nottingham, Alastair Timothy
- Hits: 933
- Visitors: 914
- Downloads: 116
Collections
RU Department of Computer Science
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | SOURCE1 | 6 MB | Adobe Acrobat PDF | View Details Download |