By Topic

TCP/IP Performance Near I/O Bus Bandwidth on Multi-Core Systems: 10-Gigabit Ethernet vs. Multi-Port Gigabit Ethernet

Sign In

Cookies must be enabled to login.After enabling cookies , please use refresh or reload or ctrl+f5 on the browser for the login options.

Formats Non-Member Member
$31 $13
Learn how you can qualify for the best price for this item!
Become an IEEE Member or Subscribe to
IEEE Xplore for exclusive pricing!
close button

puzzle piece

IEEE membership options for an individual and IEEE Xplore subscriptions for an organization offer the most affordable access to essential journal articles, conference papers, standards, eBooks, and eLearning courses.

Learn more about:

IEEE membership

IEEE Xplore subscriptions

3 Author(s)
Hyun-Wook Jin ; Dept. of Comput. Sci. & Eng., Konkuk Univ., Seoul ; Yeon-Ji Yun ; Hye-Churn Jang

With significant advances in network interfaces, I/O bus, and processor architecture of end node, innovative approaches are required to achieve high network bandwidth by fully utilizing available system resources. The issues related can be summarized into two: (i) Utilizing I/O bus bandwidth for high bandwidth network connection and (ii) Utilizing multiple cores for high packet processing throughput. In this paper, we conduct several experiments on a multi-core system with 10 GigE and multi-port 1 GigE network interfaces. We aim to show the impact of system configurations on the network performance and compare the performance of two different network interfaces. The experimental results show that, with the proper interrupt affinity configurations, the multi-port 1 GigE can achieve comparable bandwidth to 10 GigE. The peak bandwidth achieved by the multi-port 1 GigE is 6.7 Gbps, which is more than 80% of the theoretical maximum I/O bus bandwidth on the experimental system. We, however, also show that the multi-port 1 GigE can consume much more processor resource than 10 GigE. More importantly, we reveal that processing the packets on many cores can result in more resource consumption without much benefit. This can be because of locking overhead between softirqs running on different cores and lower cache efficiency. We show that the more tuning on the configuration cannot overcome this side effect.

Published in:

Parallel Processing - Workshops, 2008. ICPP-W '08. International Conference on

Date of Conference:

8-12 Sept. 2008