Performance of the biosim program.

[Per Hammarlund]

We have been approached to give some details about the capabilities
and performance of the biosim program. (A program that enables
biologically realistic simulations of large neuronal networks on the
Connection Machine from TMC. The program has been developed by the
SANS group, Royal Institute of Technology; Technical Report
TRITA-NA-P9021.)

The biosim program is used, among other things, to simulate the
swimming rhythm generator of the lamprey. The present model includes
100 segments each consisting of 6 neurons who in turn are built up of
4 compartments. The segments have 12 internal synapses and are
interconnected by 4 synapses, giving a total of 2400 compartments and
1600 synapses.  The simulation, as presented in the TR, uses a time
step of 0.4 milliseconds. To simulate 2 seconds of real time with this
setup takes about 11 minutes or one time step in 0.135 seconds on our
8K machine. The system is thus running at a speed of about 0.3% of
real time.  The time step can often be doubled without significant
changes in simulation results. Obviously, changing the time step
directly effects the simulation time.

It should probably also be mentioned that the program does not put any
limitations on the number of compartments nor synapses per cell. These
numbers are set individually for every cell.

Due to the SIMD architecture of the Connection Machine, what is
simulated is actually at least 8192 compartments and equally many
synapses, since that is the number of processing elements in our
machine. It is therefore possible to simulate either a three times
larger model of the lamprey or three lampreys with the above
specifications simultaneously without ANY change in execution time. It
also turns out that up to 65536 compartments and synapses, i.e. 27
lampreys of the above type, can be simulated in roughly four times the
time. This is in some sense equivalent to simulating one lamprey at
about 2% of real time. For a full size Connection Machine with 64K
processing elements it is possible to simulate roughly eight times
larger models in the same time, i.e. using the same reasoning you get
16% of real time. All of the above shows that it is difficult to state
any single general nice and easy to grasp number regarding the
relation to real time but we hope that this has explained the issue.

Why would anyone want to simulate the same thing 27 or 216 times in
parallel?  An excellent use of such an ability is to try out different
settings of a number of parameters in one shot. A more obvious use of
the biosim program is to run a single large model. For instance the
number of cells in each segment of the lamprey model can be increased
giving even better correspondence with reality.


Bjorn Levin,        Per Hammarlund,     Anders Lansner
blevin at bion.kth.se, perham at nada.kth.se, ala at bion.kth.se


SANS-NADA
Royal Institute of Technology
S-100 44 Stockholm
Sweden