DAQ Column status
- Finished and presented results for the TDR.
- TTCrx fast enough on dual CPU computer. (Two polling loops.)
No data errors seen with the new card + firmware.
- Performance goal reached with new P4 machine.
- Performance problem when running with 32k blocks at certain event sizes.
- After various tests and a dozen new options the '32k performance problem' was traced to a
GM performance drop when using more than 64 MB memory.
- Try to run with two myrinet cards.
Still problems setting up fedkit in the new linux machines,
therefore only fake data yet (but previous measurements showed that we
were limited by the network, not the input).
The two cards work noticeably slower than separately, see plot.
The measurements were done with fake data and fake triggers on the
Dell and on the new Grand Champion (2 CPU) machines. The RU and the
EVM have two myrinet cards, while there are two BUs with one card
each, all connected to the same switch. For 'alias' (blue line) only
one card in the EVM was used.
- Working system cleaned up and checked in. Now we use the latest
fedkit and the latest XDAQ release.
- Measured dependence on number of resources. No significant
dependence between 4000 and 40 resources. This is expected since the
BU frees the resources inmediately after the event has been built.
The measurement was done with a bundling of 10 requests per message
both for the BU->EVM (available) and for the EVM->RU (send) messages,
16kB blocksize, random and fixed fragment size, with the Dell as the
RU. [ps] [pdf]
- Extended measurement to lower values, with the same
configuration, same bundling of 10 requests per message. The
degradation of the performance is observed starting at 20 resources
for small fragments, which is the double of the bundling
parameter. The system remains operational as long as the number of
resources is higher than the bundling parameter. The new plots:
- Comparison of first and second run with 40 resources for cross-check
- Fixed-size measurements
- Random-size measurements