The LHC beams collide 22 million times a second producing about 1 Terabyte of information in the LHCb detector. This volume is too great to be read out and written to disk and in any case, most of it is not terribly interesting. Consequently, a sophisticated
triggering system filters the data in real time reducing the output rate to disk to about 100K per second.
The is shown schematically on the right. Detector information is read out from the front-end electronics and switched to one node on a computer farm which runs algorithms on the data to decide whether to keep it or not.
Our task, on joining LHCb, was to test theoretically and practically that the switch could
pass and switch data at the required rate. LHCb's intention was to use standard network switches; however LHCb's data patterns are very different from regular network traffic being both sustained and coherent (more akin to a denial of service attack).
We implemented two simulations; one custom designed by Jean-Pierre Dufey; the other written by us using
Ptolemy. Details and results are presented in
Jonathan's talk
at CERN
and in
Aidan Smoker's MSc. thesis.
Hardware tests were performed using a small cluster at UCD and using the MAP supercomputer at the University of Liverpool. This collaboration was supported by a grant from the Royal Society.
At Liveprool, we were able to replicate the LHCb network configuration and intense data rates with some nodes of MAP acting as sources and some as data receivers joined together by a candidate switch, provided by
Force 10 Networks.
Soak testing the switch for up to an hour at LHCb style rates, we could experimentally conclude that this switch would be suitable for our data-taking.
You can view presentations made by
both
Jonathan
and
Karol
at CERN.
At the completion of the project
we published
an internal LHCb note
on our findings.
A
paper
was also presented at the
CHEP2004
conference held at Interlaken.
