Present: Giovanna, Roland, Enrico, Thijs, Milo, Eric, Phil, Frank
Re-ordering / compression (see Milo's slides):
- metadata indicating format: 8 bits available for bot re-ordering and compression, so can accommodate format "evolution"
- now also compressing header information, while still storing full frame header in case of any error
- QAT performance results obtained on np04-srv-017, still using PCIe card
- algorithm for on-board QAT is identical so expect same compression factor (but improved speed)
- factor >~ 3.5: this is assuming that header compression is not very effective (so somewhat pessimistic)
- suspect that still lower factor than for RCE is due to the use of a generic rather than a custom algorithm (in s/w can also achieve a factor between 5 and 7)
- root issue with results (compression factor down to ~ 2.5) observed last week: bug in QATzip resulting in use (always) of static rather than dynamic Huffman table
- bug has been fixed in updated QATzip
- Roland: can now run with 5 inputs at up to 70 Hz (limit seems compatible with 3 ms/event * 70 events / ( link * s) * 5 links = 1.05 ~ 1)
- Eric will read the note to see if there is any information that is still needed for uBooNE
Software & firmware (Enrico):
- s/w: Roland & Joern solved issue preventing two NIC interfaces from being usable by different CPUs (newer libfabric version)
- f/w:
- initialisation instabilities fixed (thanks to Andrea, Frans, Henk, Rene)
- presently attempting to see if making the "last" FIFO deeper helps mitigate hiccups due to PCIe unavailability
- new FELIX host will not arrive before mid October (lack of computer memory @vendor) (!)
- so may need to stay with two FELIX hosts for the time being (backpressure turns out to be related to the limited number of available cores)
- some improvement perhaps possible now that NIC issue is solved (idea: use 2 CPUs w/ 5 links / CPU)
Next meeting: tentatively on September 19
There are minutes attached to this event.
Show them.