GitLab

The values for these can be extracted from the environment so they can
be changed between Template runs.

Fixed a bug in how bufferFull was calculated (it was erroneously true most of the time). Changed LeastSquares to require both a buffer full and increasing delays to consider a measurement authoritative.

time a run starts. This info is stored in the DB forever, but a file
in the archive is a nice thing too.

apps/protocols that use X cpu on PlanetLab can also add network artifacts.

sfmiddleman sits in the middle as a compute-intensive proxy.  sfsender
sends the middleman n "messages" each composed of m "blocks", and can
pause for t1 and t2 times at microsecond granularity after each block send
and msg send.  sfmiddleman reads blocks and does operations on them
(default is decrypt/encrypt, can also do only encrypt), then forwards them
to all connected receivers.  You guessed it, sfreceiver connects to
sfmiddleman and reads blocks and prints timestamps.  For each op on a
block, sfmiddleman notes the number of microseconds taken by the crypto
ops, and the timestamp when it/they finished, and prints those.

There is one known bug.  I'll fix it in the morning.  If there are others,
I'll get them then too.

template world:
	set no_plab to 1
	set use_dbmonitor to 1
	put enough nodes in the plabnodes list
	use start-run and stop-run to do it

An unexpected exit code will cause the __all_programs HALT event to
fail and cause "Stop Run" to fail.

that pumps up bgmon.  To use:

	set USE_DBMONITOR non-zero
	modify DBMONITOR_INTERVAL from the default 10 seconds if desired
	make sure there are enough nodes listed in PLABNODES

These things can either be modified in the template, or in the instantiation
params.

Dan's message).

instance being replayed, not the template itself.

and issues tevc commands to delay nodes.  This version will look in the user
environment for plab nodes if they are not actually allocated to the
experiment.  This enables use of the no_plab option in the NS file so plab
nodes are not actually allocated to the experiment.

Can still get the per-node version with -r.

The centralized version uses bgmon (managerclient) to crank up the frequency
at which measurements are made.

Added idle/pcap loss avoidance to the system. Plus many miscellaneous debugging statements. Most notable is that there is not debugging output describing the throughput as seen at the offered load level (how much you can write to the socket buffer).

Changed slope calculation so that it is authoritative only if the slope is more than very slight. This is to remove the effects of rounding differences due to the banding at the 1ms level. This removes an artificial feedback effect. There are more potential feedback effects that I still need to investigate.

Fixed a bug causing underestimation of delay (in some cases a delay of 0). Now artificial packets are not counted when looking for the most recent packet.

stuff).

tblog code.

Run pings 4/second for 4 seconds, and measure loss at end. If loss is
100%, go onto highFreq state, and rest. If not 100%, go back to "normal"
state, but run latency tests at 10 sec periods for 2 minutes.

The result of the "lossCheck" is sent as a measurement result back to ops,
as either "lossy" or "down".

value of "down", insert a loss of "1" in the DB.

at this point.
- fixed -h option to correctly display the "help"

Added code to detect when a connection has closed but packets are still being captured for it and to not mark that as an error. Every time a new connection opens, a few dozen bytes are used up. But these bytes are not removed when the connection ends. So this is not yet safe for extremely long runs. TODO: Add a timer based removal.

 * all the user configurable variables become parameters (complete with
   descriptions--w00t!)

 * triggers a 'start' event sequence at time 0 which does what
   start-experiment does

In theory, there was going to be an auto-triggered 'stop' sequence as well,
but it turned out to be unneeded (the default stop-run actions are sufficient)