Last update: Apr 1, 2000
Speaker: Frans MeijersThe data acquisition system for the CMS experiment at the future LHC pp collider at CERN will require a large and high performance event building network. This switch fabric connects 500 readout units (RUs) to 500 builder units (BU). The RUs read out data from detector elements at a first level trigger rate of maximum 100 kHz. The expected average event size is 1 Mbyte, corresponding to event fragment sizes of 2 kbytes. With events of this size, the 500x500 event building network requires an effective aggregate bandwidth of 100 Gbyte/s. A software trigger running in Filter Units connected to the BU's reduces the rate to about 100 Hz of events to record on permanent storage. An event manager broadcasts the first level trigger information to all RUs and performs the destination assignment for each event.
Several switch technologies are currently being evaluated in order to compare different architectures for the event builder. One candidate is Myrinet. This paper describes the demonstrator which has been setup to study a small-scale (16x16) event builder based on PCs running Linux connected to Myrinet and Ethernet switches.
The software is designed in several layers. The application layer implements the high-level event building model for Readout Unit, Builder Unit and Event Manager. The presentation layer implements the protocol for remote message invocation. The network technology dependent part is confined to the transportation layer, which currently supports UDP, Myrinet and low level Ethernet. In order to minimise overheads, a zero-copy user-level message passing mechanism has been developed for Myrinet and Ethernet.
A detailed study of the Myrinet switch performance has been performed for various traffic conditions, including the behaviour of composite switches.
A number of event building architectures have been studied. In the push architecture, the event manager assigns a destination for each event and broadcasts the event identifier and destination to all sources. The sources then send their event fragments to the assigned destination. In the pull architecture, the destination processor initiates the data transfer by requesting event fragments from each of the sources. Measurements are presented on throughput, overhead and scaling. Traffic shaping techniques have also been studied.
Finally, results from the discrete-event simulation are presented and the extrapolation to a large scale system.
|| | Home | Bulletins | Committees | Scientific Program | Docs by topics | Social Event | Conference Location | Secretariat ||