Speaker
R. Hughes-Jones
(THE UNIVERSITY OF MANCHESTER)
Description
How do we get High Throughput data transport to real users? The MB-NG project is a
major collaboration which brings together expertise from users, industry, equipment
providers and leading edge e-science application developers. Major successes in the
areas of Quality of Service (QoS) and managed bandwidth have provided a leading edge
U.K. Diffserv enabled network running at 2.5 Gbit/s. One of the central aims of MB-
NG is the investigation of high performance data transport mechanisms for Grid data
transfer across heterogeneous networks.
New transport stacks implement sender side modifications to the TCP algorithm which
enable increased bandwidth utilisation in long-delay high-bandwidth environments.
This allows a single stream of a modified TCP stack to transmit at rates that would
otherwise require multiple streams of standard RENO TCP. This paper reports on
investigations of the performance of these TCP stacks and their use with data
transfer applications such as GridFTP, BBFTP, BBCP and APACHE. End-host performance
behaviour was also examined in order to determine effects of the Network Interface,
PCI bus performance, and disk and RAID sub-systems.
In a Collaboration between the BaBar experiment and MB-NG we demonstrated high
performance data transport using these new TCP/IP transport protocol stacks and QoS
provisioning. We report on the benefits of this introduction of high speed networks
and advanced TCP stacks together with various levels of QoS to the BaBar computing
environment. The benefits achieved are contrasted with network behaviour and
application performance using today's "production" network.
Primary authors
Dr
N. Pezzi
(University College London)
R. Hughes-Jones
(THE UNIVERSITY OF MANCHESTER)
Dr
S. Dallison
(THE UNIVERSITY OF MANCHESTER)
Mr
Y-T. Li
(University College London)