commit logs in tbsetup for more info.

error.

timestamp of the packet that made us realize we missed them.

This fixes the last error from my washington->oregon trace.

When we detect that we've lost packets in libpcap, compensate by
making some dummy packets to cover the sequence range we missed. Use
a constant to guess how much header we missed. Only flag missed
packets as errors if we miss a lot of them.

Allow for ACKs of partial packets - this is important with the above,
since the fake packets can be larger than the real packets that we
missed.

Mark SACKed packets so we can avoid double-counting them - future
SACKs or ACKs can cover those packets, but, in fact, are not really
ACKing them.

Most of the time spent working on this was dealing with STL semantics.
The more I use the STL, the less respect I have for it.

Added a limit of the maximum number of samples used by LeastSquaresThroughput for slope calculation. Throughput calculation still uses the previous metrics.

problem, which is that swapmod messes up the env that the template
code has carefully setup.

with the old monitor.  Backward compat be damned!

Added debugging information so that it is more clear what port the control server is using on the stub.

think the problem is that the archive is forked for a template
instance, but not committed before the first savepoint. This screws up
the archive cause the datastore directory has not been copied in yet,
and subversion pukes.

kludgy code in template_exprun that tries to ignore dead nodes during
start and stop run.

Rolled the AverageThroughputSensor into a special case of LeastSquaresThroughput. It still uses the same command line arguments/constants.

Bump plab rootball version.

Create event sequences to do the equiv. of start- and stop-experiment.
Rather than using those scripts, you now can use start-run and stop-run
which are just wrappers around tevc to trigger the sequences.

This checking is backward compatible in the sense that you can continue to
use start- and stop-experiment and should never notice a thing.

The new event model doesn't support a couple of things that the old script
based execution did, most notably, you cannot change command line arguments.
It does support real vs. fake runs.

script sticks around so the program agent has something to kill!

will not timeout in the dispatch routine.  It apparently will stay
in recvfrom() til it gets a packet.  This caused stop/snapshot events
(as well as "kill -TERM") to not do anything.

So we blast all the pthreads out of recvfrom with a signal after setting
the flag to force them to return from dispatch (pcap_breakloop).

PlanetLab node's load.

Added a new sensor which averages packets over a particular period of time rather than a particular number of acks. This provides a proof of concept of coarse-grained measurements reflecting changes in bandwidth. Increased the number of samples that LeastSquaresThroughput uses as a quick way to use coarse-grained measurements easily.

getting deleted.

* Copy template datastore to instance "datastore" directory, instead of
  "template_datastore" ... avoid mass confusion.

* Change xxx:// mappping from template datastore to the instance datastore.

has two problems:

 * not all LANs might be shaped so there may not be a delay-agent running
   to receive and acknowledge the RESET event, leading to a hang

 * we might wind up resetting a delay node more than once (once for each
   lan it is responsible for); probably not harmful, but might lead to
   again not sending a completion event and a hang.

As long as I can reset the one LAN I care about, or rather all its involved
delay nodes, I should be fine.