[subexp-daq] dropping of data to analysis that cannot keep up

[Håkan T Johansson]

Hi,

when the point of DAQ being faster than online is reached, we face the 
issue of somehow dropping old data.

Left with default settings, ucesb is not really helpful.  Old data will 
pile up in the '--server' until that is full, and only then dropped.  If 
the server is configured with a large amount of memory for buffers 
('--server' option 'size='), this can be significant.  This leads to 
online analysis lagging reality, loosing the 'online' feeling and useful 
rapid feedback of changing conditions...

There is a 'dropold=N' (seconds) option which will not serve buffers to 
the client older than the specified time.  It may be reasonable to set 
this to at most two times the spill length, if the beam has a pulse 
structure.  Otherwise a few seconds ought to be enough.


A dry-run demo (3 instances), using the 'mbuffer' program to limit rates:

---

# Generate data (at 10 MB/s):

file_input/empty_file --lmd \
   | mbuffer -r 10M -q \
   | empty/empty --file=- --server=trans,size=1G

# Serve the data, with dropold (can be removed):

empty/empty --trans=localhost --server=trans:8000,size=1G,dropold=10s

# 'Analysis' (at 1 MB/s).
# Look at the data (buffer times)
# Note: dd is used here to strip the 16-byte transport protocol startup 
# message.

nc localhost 8000 \
   | mbuffer -r 1M -q \
   | dd bs=16 skip=1 \
   | empty/empty --file=- --print-buffer

---

Note: the buffer times only relates to the middle server, not the original 
data generation.  But in this case is good enough to show the difference.
If 'dropold' is not used in the middle instance, then the last instance 
(doing --print-buffer) will be seen to lag further and further.


Cheers,
Håkan