People present at meeting:
Hal Evans, Eric Hazen, Ulrich Heintz, John Hobbs, Bill Lee, Jim Linnemann
(via video-link), Meenakshi Narain,
Chuck Pancake, Wendy Taylor, Shouxiang Wu
Minutes:
(Notes taken by Eric Hazen)
Endian-ness discussion: Alphas are little-endian.
Final setting will probably need to be determined empirically!
(all outside-world interfaces should be switch-able)
L1CTT data padded to multiples of 4 words.
Min T/R size: 10 words Max T/R size: 54 words
First T/R word arrives at STC/TFC logic after about 3.5us
Errors: missing BoE: (two EoE in a row)
FRC constructs header with SCL BX number and
trailer with error bit set
missing EoE: (two BoE in a row)
FRC constructs trailer with error bit set
(timing of these is not yet well defined)
[ Jim Linneman again suggested that LVDS links should be protected scrupulously against missing header/trailer words ]
Word count is at end. All other details subject so possible change.
** sure would be nice to identify event boundaries uniquely **
** can we agree across the 3 boards on this ?? **
DB responsibility is to raise L3_READY, make data available for transfer, and to disconnect after transferring trailer.
Logic boards must provide sufficient FIFO space for one unbiased event plus a (few) normal events.
If this buffer gets full, need to (a) raise L1_BUSY and (b) dump data. SCL_INIT is to be avoided.
Logic boards signal to FRC that monitoring data is ready implicitly by processing and transferring L3 data as usual, then set mon_done[i] register for CPU to notice.
FRC interrupts CPU, which checks mon_done[], reads data, etc.
Jim L points out that the read/clear model breaks if a monitor readout cycle is missed. He prefers not clearing anything and letting the CPU handle rollovers. It seems to me that the February review told us to use the read/clear model, but opinions differ...
Hal's flowchart assumes that monitoring is done is one atomic set of accesses to the VME bus. It was suggested that it be possible to break the accesses up and permit a VBD transfer in the middle. Hal agrees to look at this.
"End-of-run" monitor request comes only when the system is completely quiet according to Jim L, but he will look into this.
3 types of monitoring:
** "live" plus buffered monitoring registers should both be
PCI-accessible for diagnostic purposes **
** LRB needs copy of COLLECT_STATUS in the trailer (or header) to
tell it to copy its monitoring registers.