high performance computing on graphics processing units: hgpu.org

Traffic analysis is an essential part of capacity planning, quality of service assurance and reinforcement of security in current telecommunication networks. As the network speed increases so does the traffic volume and the analysis of large traffic traces is computationally intensive. This document presents a flow extraction software that allows obtaining TCP flow records at 4.4 Millions of packets per second in a single GPU directly from network. Such TCP flow records include number of retransmissions and duplicates per flow, which are very challenging to obtain at high-speed. Other records extracted are flag counters, zero window counters, duration and length of each TCP flow. The software is designed as a pipeline where the packet data is analyzed and compacted in each stage to result in a brief record for each TCP flow after the last stage. Some stages of the algorithm take place in the CPU using pthreads, but the lion’s share of the processing takes place in the GPU, freeing CPU cores for other uses. There is also a second program that helps with post-processing. The high performance of the software comes at the price of precision, which is lost due to the implementation of the algorithm and limitations of memory. Our tests show that the loss of precision is small, so this tradeoff makes sense. Our software significantly increases the processing power of the recently proposed high-speed sniffers based on commodity hardware and demonstrates the advantages of applying massively parallel processing devices for traffic analysis.