bspprof - The Oxford BSP Toolset communication profiler for
bspprof [input-file] [output-file] [command-line-options]
The Oxford BSP toolset communication profiler analyses trace
files generated during the execution of a BSPlib program.
The profiler converts trace files into PostScript (level 2)
documents that are suitable for visualisation on a wide
variety of graphical devices. Trace files are generated by
adding the option -prof when compiling BSPlib programs with
The profiler produces three types of PostScript graphs as
1) Strata-graphs for either the number of distinct commun-
ications that occur over time, or the number of bytes
communicated. Two graphs are presented per profile:
a) the number of bytes (or communications) going in
to a processor.
b) the number of bytes (or communications) going out
from a processor.
2) Prediction graphs that are similar to 1) except that
the lower graph shows the predicted cost of the commun-
ication on the machine the profiler was run on. If the
- arch command is used to specify another architecture,
then that architectures BSP parameters will be used
instead. Alternatively, the BSP parameters of the
architecture which the prediction is for can be entered
using the -slg option.
3) A G graph, which is also based upon 1) except the bot-
tom graph shows the effect of g over time. For a true
BSP machine, g should be constant.
Options must be separate: `-dr' is quite different from `- d
- r '. Here is a summary of all the options, grouped by
type. Explanations can be found in the appropriate sections
-title string -zoom from,to -colo[u]r -letter -bytes -
packets -ignore -outline
Strata Graph Options
Prediction Graph Options
-predict -slg s,l,g,[nhalf] -arch string -nhalfmodel -
-ggraph -ceilingg integer -hrelgl
Use string as the title of the profile.
Zoom in on the section of the profile between times (in
seconds) from and to. If a superstep stradles either
from or to it is ignored.
Use coloured bars when producing the profile in prefer-
ence to the default grey-scale hashing.
Generate a postscript file that will print on American
letter format paper. The default print area is A4.
Displays the number of bytes transferred from each pro-
cess during each superstep.
Displays the number of distinct messages from each pro-
cess during each superstep.
Don't profile any supersteps from the file string.
Only show the accumulated message sizes for a super-
step. This option is useful to see the general shape of
the profile without generating large postscript files.
STRATA GRAPH OPTIONS
Produce a strata-graph showing two graphs with the
amounts of data entering and exiting each processor for
each superstep over time.
PREDICTION GRAPH OPTIONS
Produce a graph that contrasts the experimental results
with a prediction based upon the BSP cost model.
When generating the prediction profile, the options to
-slg are numbers which represent the BSP parameters:
s is the Megaflop/s rating for the machine.
l the minimum number of flops for a barrier syn-
g the asymptotic number of flops required for all
processors to communicate a single 32bit word to
the number of words required for the communication
performance to deteriorate to 2g.
See bspparam(1) for information concerning the BSP parameter
Specifies that the program was run on architecture
string; where string is one of the valid architecture
options that is returned from bsparch -help.
The formula h.g is used as the cost of communicating a
The formula h.g.(nhalf/h +1) is used as the cost of
communicating a h-relation. i.e., messages smaller than
nhalf are attributed a larger cost to take into effect
communication startup costs. This is the most accurate,
and therefore the default cost model used in the pro-
The formula h.g.nhalf is used as the cost of communi-
cating a h-relation. This will give an upper bound,
pessimistic cost of communication.
For each superstep, assume that l is fixed, and plot a
graph for the variance of g realised for the superstep.
Although the BSP model assumes a fixed g, this option
is useful when trying to visualise the effects of con-
tention in the implementation; or alternatively to
check if a machine is truly scalable!
Specify the maximum value of g to show in the lower G
graph. This option is useful when useful when trying to
scale multiple graphs to the same ranges.
In the ggraph, we assume that l is constant, and we
plot the variance of g. When the size of the h-relation
is much smaller than l/g, then a small variance at run-
time in l will produce a large variance in g. By de-
fault, when generating the profile for the G graph, h-
relations smaller than l/g are ignored. With this op-
tion, they are plotted as small white dots.
The file ROOTDIR/include/bsp_parameters.ascii contains the
BSP parameters for various platforms; where ROOTDIR is usu-
ally set in the top level Makefile of the BSPlib source.
bsplib(3), bspfront(1), bsparch(1)
``The theory, practice, and a tool Jonathan Hill, Paul
Crumpton, and David Burgess. In EuroPar'96. LNCS, Springer-
Verlag, August 1996. See:
The Oxford BSP toolset web pages can be found at:
Problems and bug reports should be mailed to bsplib-
Copyright 1995,1996,1997 University of Oxford
Permission to use, copy, modify, and distribute this
software, and to incorporate it, in whole or in part, into
other software, is hereby granted without fee, provided
i the above copyright notice and this permission notice
appear in all copies of the source code, and the above
copyright notice appear in clearly visible form on all
supporting documentation and distribution media;
ii modified versions of this software be accompanied by a
complete change history describing author, date, and
iii and any redistribution of the software, in original or
modified form, be without fee and subject to these same
Man(1) output converted with