Royal Holloway Centre for Particle Physics

Evaluation of network performance for triggering using a large switch Talk given at CHEP98 Chicago, Illinois
N Madsen, J A Strong et al
Abstract
Measurements have been made on a large switch with network traffic emulating that expected in the second level trigger of the ATLAS experiment at the LHC. The switch used in the measurements is part of a 1024-node DS-link based test-bed which can be configured into standard network topologies \026 grids, torii, Clos etc. The system is used to examine network performance as a function of data rates and distributions. In ATLAS, after an event is accepted by the first level trigger, data for the event are moved off the detector on parallel 1 Gbps links and stored as independent fragments in approximately 1600 data buffers. Proposals for the ATLAS second level trigger use networks to gather sets of data fragments into a processor from buffers connected to areas in the detector indicated by the first level trigger as being regions of interest (RoIs) for that event. A decision frequency of up to 100 kHz must be obtained and, with current estimates of event processing time in the second level trigger of between 1 and 10 ms, hundreds of processors will have to operate in parallel and many events will reside in the system concurrently. Results show that the required performance can be obtained in most cases but that due care and consideration must be given to the temporal and spatial distribution of the data presented to the network. Implications for commercial systems will be discussed.
Available format: M/S Powerpoint file
Also available as:
Contributed paper to CHEP98 Chicago, Illinois