Having some trouble running with the foss toolchain, else I'd include them instead.

Downloading and placing all required tarballs in the correct places
is likely to be tedious.

This is needed because OpenMPI on trusty fails with this error:

/usr/lib/openmpi/include/mpi_portable_platform.h:374:34: error: invalid suffix on literal; C++11 requires a space between literal and string macro [-Werror=literal-suffix]
              _STRINGIFY(__GNUC__)"."_STRINGIFY(__GNUC_MINOR__)"."_STRINGIFY(__GNUC_PATCHLEVEL__)

They're probably not fully complete

This has brought the speed of perftest2.sh from 2.3 kHz to 8.8 kHz, nearly
quadruple the speed!

There's some stuff still left to do in the master:

* queueX() should only need to be called once for each "stage"
  in each master. It accounts for 23% of cycles.
* initRefValues() accounts for 16% of cycles, perhaps there's still
  something that can be done to it.
* There's probably some overhead in message packing/unpacking to take care of.

Overall, the master spends about as much time inside itself as inside ZMQ.
Memory allocations seem to account for most of that not accounted for in
queueX() or initRefValues().

As for the server, there's not much else to do since 36% of time is spent in
FMI and 55% is spent inside ZMQ. The remaining 9% can be considered
acceptable overhead.

1% callgrind improvement

No change in walltime, but callgrind says 3% fewer cycles.

This reduces cachegrind cycles a bit, but no impact on walltime

31% Speedup in Release, woo!

4% Release speedup, and nicer code.

7% Release speedup.

12% Release speedup.
We're currently at 12 kHz on granular, compared to 5.5 kHz a day ago :]

This gives a 25% speedup in Release mode!

Speedup is 2% in Release, 5% in Debug

Sometimes it's good, sometimes not.
Since it modifies the master's behavior leaving it off seems best.

~1% speedup in Release, 3-5% in Debug

Speedup is a modest 3% in Release mode

It seem granular tops out around 68 kHz with the current architecture,
with fmigo-mpi being roughly an order of magnitude slower.

walltime down another 20%

This gives an MPI speedup of 27% in Debug mode with current perftest2.sh.

perftest2.sh results:

before:
real	0m55,910s
user	1m28,044s
sys	1m17,734s

after:
real	0m46,482s
user	1m2,104s
sys	1m12,223s

callgrind results were more impressive,
but it can't every fully represent reality.

Provide examples of TCP/IP command-line use while we're at it.
This fixes #2.

This fixes #1.