Speaker
Dr
Stefan Stancu
(University of California, Irvine)
Description
The ATLAS experiment will rely on Ethernet networks for several purposes. A control
network will provide infrastructure services and will also handle the traffic
associated with control and monitoring of trigger and data acquisition (TDAQ)
applications. Two independent data networks (dedicated TDAQ networks) will be used
exclusively for transferring the event data within the High Level Trigger and Data
Acquisition system, all the way from detector read-out to mass storage.
This article presents a networking architecture solution for the whole ATLAS TDAQ.
While the main requirements for the control network are connectivity and resiliency,
the data networks need to provide high-bandwidth, high-quality transfers with minimal
packet loss and latency.
As the networks size is large -- O(1000) end-nodes -- we propose to use a multilayer
topology, with an aggregation layer (typically at rack level) and a core layer. In
order to achieve high resiliency, we propose to distribute the core of each network
on multiple devices interconnected via high-speed links, and to deploy a protocol
that efficiently uses redundant traffic paths. In addition, geographical aspects
(e.g. distances requiring optical fibre instead of copper) are addressed. The
proposed network architecture will be mapped on typical commercial devices.
Sample performance evaluation results, meant to validate the data network from the
pre-series system (a 10% slice of the final TDAQ system). Traffic patterns similar to
the ones created by real applications have been used to determine the network
performance under TDAQ specific conditions.
Primary author
Dr
Stefan Stancu
(University of California, Irvine)
Co-authors
Mr
Brian Martin
(CERN)
Dr
Catalin Meirosu
(CERN and Politehnica University of Bucharest)
Mr
Lucian Leahu
(CERN and Politehnica University of Bucharest)
Mr
Matei Ciobotaru
(CERN and Politehnica University of Bucharest)
