Demonstrator for a High Performance Second Level Trigger

Paper: 440
Session: B (talk)
Speaker: Rensch, Bertram, Niels Bohr Institute, Copenhagen
Keywords: ?


Demonstrator for a High Performance Second Level Trigger
==========================================================

J.D. Hansen, J.R. Hansen, M. Dam and B. Rensch

NIELS BOHR INSTITUTE, COPENHAGEN


Abstract

We developed and evaluated a scenario for the implementation
of the Second Level Trigger(SLT) in the HERA-B experiment at DESY.
The architecture is charactrized by a global switching network
connecting the SHARC based Second Level Buffer(SLB) to a large
farm of commodity processors.
Data flow and resource allocation is controlled by a single
SuperVisor(SV) that also maintaines the globally managed buffer
in the SLB system. Events are processed in parallel leading
to an automatic load balancing at the processor level.

We have chosen the by now well proven DS-link technology for the
implementation of the global switching network.
A 512 terminal Clos (or indirect multistage) adaptive routing
network built from C104 asynchronous packet switches connects
the 134 SLBs to about 150 processors, leaving plenty of free
connections to add more processors or to multiply the bandwidth
to critical nodes.

To adapt the SLB system to the switch, we built a SHARC-to-DS
interface which handles the SHARC-link protocol by an FPGA that is
connected to a T9000 via two FIFOs, one for each direction,
and allows for a bidirectional message transfer rate of 50 kHz.

A commercially available DS-to-PCI interface has been chosen
to hook the processors to the switch. The T9000 on the interface
communicates with the SV via the switch, pulls the relevant
data from the SLB system and uses the attached processor as
a coprocessor to take a decision on the event.

The whole arrangement was emulated on MACRAME using the expected
HERA-B traffic pattern. Results show that the system can digest
an event rate in excess of the maximum possible of 100 kHz.

To demonstrate the validity of our approach we have built a
full vertical slice of the system including the SV to optimize
the data and control flow.