DiFX benchmarking

DiFX benchmarking can be done using vlbi_fake program in this directory plus
some wrapper scripts. Samples of these scripts are in the "sample" driectory.

A significant amount of manual tweaking is needed - probably the sample 
scripts should be highly modified for your local environment and actual tests
you want to run. The run.sh script included steps through running DiFX on 
1..n compute nodes, with 1..m threads per node. 

Files

vlbi_fake

A simple evlbi data generation program. For benchmarking only the TCP mode is
useful. This has only been tested assuming LVBDR data, mark5b may work as 
well (for benchmarking it should not matter what you use). The options you
are likely to use for benchmarking are:

 -help	  	       Summary of options
 -duration <SEC>       Amount of seconds to transfer (default 60)
 -host <HOSTNAME>      Hostname to connect to. Default localhost which is
       		       probably what should be used
 -port <PORT>	       TCP port to use. Needs to match DiFX
 -nchan <N>	       Number of (2 bit sampled) VLBI channels to generate
 -bandwidth <VAL>      Bandwidth to assume (default 16).
 -day <DAY>            Day of month of start time (default now)
 -month <MONTH>        Month of start time (default now)
 -dayno <DAYNO>        Day of year of start time (default now)
 -year <YEAR>          Year of start time (default now)
 -time <HH:MM:SS>      Year of start time (default now)
 -mark5b/mk5b          Send Mark5b format data. Untested


sample/vstart.sh

Simple script to tun vlbi_fake. This will need to be editted to set the 
proper path of vlbi_fake and the starttime/duration etc. Normally the easiest
thing is to copy it to $HOME on the i/o nodes.

sample/run.sh

Sample script to test the full maxtix of varying number of compute nodes and
threads per node. This launches vstart.sh on the i/o nodes, sets the thread
file appropriately and launche DiFX. The output from DiFX and the i/o nodes are all logged.

Setup

Create/find a DiFX input file with the mode you want to test (number 
channels, telescopes etc). If you want to test the number of threads create
a set of "thread" files (for the maximum number of compute nodes) for each
thread test. Create the host files with the maximum number of hosts you want 
to test. 

Analysis

The perl script medrianrate.pl computes the median transfer rate from the 
i/o log files.  This can me used as the "speed metric". Just running on the 
logs from a single i/o node should be enough.

Chris Phillips