Developer's meeting

Wednesday 21 Feb 2018, 17:00 → 18:30 Europe/Zurich

Present: Roland, Milo, Giovanna, Eric, Frank

Discussion of integration issues observed recently

felixcore:

• is not publishing data from more than a single link
  • hopefully something that is relatively straightforward to fix
• of (much) more concern: necessary throughput is not reached (by a wide margin)
  • Roland carried out "bare" iperf tests to demonstrate that the issue is with the small packet size used for the NetIO transmission (no problem when using 128 kB size)
    • see https://docs.google.com/document/d/1fensIWDnNwnxmqMyEdeHUp3wCrcbNBTOxbiOtZ1aXn0/edit
    • this aspect was not tested by us before (previous throughput tests did not use felixcore) and we also did not have any indication that this would be an issue
  • FELIX developers say that packet coalescence should help
    • but also that in fact that should be used automatically by NetIO? At any rate it seems clear that we are not using it
• the near-term plan is to test more systematically (presumably mostly Roland):
  • throughput as function of packet size (TCP/IP)
  • throughput for the Inifiniband alternative
  • these tests could be done (to some extent -- at most 4 links) using the Full-mode emulator  

WIB data:

• RCE colleagues see COLDATA convert count issues (common at least to COLDATA chips connected to the same FEMB)
• we (presumably mostly Milo) should check whether we observe the same issues

 

timescale:

• it is not unlikely that the last APA that will allow for extensive cold-box tests (before insertion in the cryostat) is APA #4, which is to arrive in 2 weeks from now
• consequence: if we do not manage to sort out the above issues to (sufficient) satisfaction, we may be stuck for a longer period without opportunity to test
  • so it is important to attempt to fix issues (very) soon

17:00 → 17:20 EHN1 integration

Speakers: Enrico Gamberini (CERN), Roland Sipos (CERN)