Online Level-3: Performance Setup

Performance Test Setups

It is very difficult at present to simulate the proper load of the filter executable since we do not have real data nor do we have all the algorithms in place. The most realistic load can be obtained when using MonteCarlo events which contain the complete detector and thus allow all existing algorithms to run realistically.

Concering the control code the situation is somewhat better since most of the software is in its almost final version. The only difficulty is the proper reformatting since running from MonteCarlo files does not need any reformatting whatsoever. Therefore a fake reformatter option can be turned on which will do a reasonable amount of byte swaps and memory copies to simulate the real reformatting process.

The following setup is used to run the performance tests:

• All machines mentioned in the following are sitting on one 3Com fast Ethernet switch. Converter and Output Nodes are connected with two inputs to the switch. Processor nodes are connected with just one input.
• Events from a ttbar or bbbar MonteCarlo data sample are read from disk on nodes b0l3c07 or b0l3c08, respectively.
• The data are sent from the Converter Nodes to the Processor nodes where they are processed.
• After processing the data are sent to the Output Node where they are trashed.
• Processor nodes are:

   
   b0l3010 - 400MHz,  DPII, Mem: 128MB, Bus: 100MHz - l3 node standard
   b0l3011 - 400MHz,  DPII, Mem: 512MB, Bus: 100MHz - l3 node 512 MB 
   b0l3032 - 600MHz, DPIII, Mem: 512MB, Bus: 100MHz - Lancewood
   b0l3031 - 733MHz, DPIII, Mem: 512MB, Bus: 133MHz - Dell server