A 4x4 switch fabric with three FPC2s was used for the experiments. One FPC was simply an Ethernet gateway to allow booting of the other two, and played no part in the experiment. The other two were both connected to ORL YES V2 boards in DEC Station 5000/25s via approximately 50m of twin coax. Both DEC Stations were running Wanda, but with different user processes. One ran a traffic source process, the other a traffic sink.
All the FPCs were running kernels with gateways and RARP daemons spliced on. The host info was setup with MSNL routes for both fabric and transmission in the standard way. The two FPCs connected to DEC Stations were given gateway controls that allowed them to forward on behalf of their attached machines.
The DEC Station running the traffic source attempts to connect to a known port on the sink machine. The FPCs setup the connection, and give it a number of seconds to settle down (allowing print statements to occur etc). The traffic generator then produces a regular traffic pattern over the connection that speeds up slightly every 10 seconds.
The FPC2 connected to the traffic source must control the Xilinx to cause cells to be VCI re-mapped and then sent out on the fabric with the correct route byte. The FPC2 attached to the sink does no further work after the connection has been established - data arriving from the fabric passes through the passive side of the port controller out through the transmission interface to the traffic sink where it is dropped.
The traffic generator was capable of generating bursts of cells
every
s. These cell bursts were at the full transmission
line rate of 87.3 Mb/s. A particular setting of
and
was
repeated continuously for approximately ten seconds, during which time at least a
million cells are shipped. Should the FPC2 not be able to handle this
rate, its 2000 cell input buffer would begin to fill as cells arrive
at a faster rate than it can process. When the 2000 cell buffer
becomes full, this is detected by the FPC2 software which reports the
fact and halts. The previous data rate was the one that the FPC2 could
handle continuously. It is possible for this data rate to be an
over estimate - cells could be arriving slightly faster than leaving
and backing up slowly in the 2000 cell buffer. Given that we are
sending a million plus cells at each speed iteration, this over
estimate is bounded by 2000 over a million, 0.2%.
If the FPC2 can
successfully cope with the data rate, the parameter is
decremented, so that
cell bursts are sent every
s.
The reason for the `cells in burst' parameter was to see how the traffic profile affected the maximum sustainable throughput. The experiments were repeated with a range of burst sizes between 1 and 35 cells. Bursts longer than 35 cells were not easily generatable, and results show that the things had settled down by this size.