[LA-SiGMA-gpu] Cheng Cui - GPU Talk, 307 Frey - June 27

[sjlee at cct.lsu.edu]

Hello Everyone,

Cheng Cui, a candidate for replacing Zhifeng Yun, GPU Team Lead is scheduled
to give a talk Thursday, June 27th at 1:30pm in 307 Frey. The title of  
his talk is "The Performance of Transmission Control Protocol (TCP)  
over 10Gbps High Speed Networks with Tiny Buffers."

Abstract:
Recent years, the growth of physical bandwidth in high speed computer  
networks has exceeded the growth of bandwidth in computer internal  
I/O. Therefore, interconnecting with high speed networks becomes more  
important in distributed systems. Because of computing legacy and code  
compatibility, many applications still rely on networks over Ethernet.  
However, as a best effort delivery media, Ethernet depends on the  
TCP/IP protocol for reliability which is typically implemented in  
software and is sensitive to buffer size. For example, TCP requires a  
buffer size of bandwidth delay product in switches/routers to maintain  
nearly 100% link utilization. But such large buffer will challenge  
hardware design and power consumption, and will generate queuing delay  
and jitter which again cause problems. Therefore, for latency  
sensitive applications reduce network delay and improve data movement  
performance are top priorities.

This talk gives a procedure of studying TCP performance in 10Gbps high  
speed networks. First, a 10Gbps re-configurable optical networking  
platform is developed as a research environment. Second, a better  
traffic sniffing tool is developed for measuring and analyzing TCP  
performance for networks beyond 10Gbps. Based on observation, two  
basic reasons that cause performance problems are studied. Finally, a  
new TCP protocol called Multi-Channel TCP and its congestion control  
algorithm called Desynchronized Multi-Channel TCP (DMCTCP) are  
presented. The protocol and algorithm implementation take advantage of  
a potential parallelism from the Multi-Path TCP release in Linux. With  
a few dozen of packets as buffers, our experimental results confirm  
that bottleneck link utilization can be much better improved by DMCTCP  
than by many other TCP variants.

See attached flyer.

Regards,
Shelley Lee
Project Coordinator
578-0465
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Cheng Cui - Flyer.pdf
Type: application/pdf
Size: 76526 bytes
Desc: not available
Url : https://mail.loni.org/mailman/private/lasigma-gpu/attachments/20130625/bedc1dea/attachment-0001.pdf