All the cell forwarding is carried out in the FIQ (Fast Interrupt Request) handler in the Wanda kernel. For more details look in the Xi3 bits document, (section 18).
The FIQ handler is entered on receipt of a cell. It attempts to forward the cell, placing it on the transmit queue, and then checks to see whether another cell has arrived. The code to do this is a small piece of hand crafted assembler written by Richard Black. Under heavy load, the FPC2 never leaves the FIQ routine as there is always another cell ready. When in this state, the port controller is unable to process any management cells that may have arrived. They just sit on a `cells for Wanda' queue in the 2000 cell buffer.
The time spent in the FIQ can be observed by attaching a probe to the FPC2's yellow LED. The FIQ illuminates the LED on entry and switches it off on exit. The original purpose of this LED was to detect crashes in the FIQ code. While the maximum throughput experiments were in progress, the data rate at which the FPC2 began spending all of its time in the FIQ was noted.
All experiments were performed with the same Xi3 Xilinx firmware bits, and the same Wanda kernels. Attempts were made later on in the series of experiments to achieve a 20 MHz switch fabric speed instead of the usual 18 MHz, but this proved unreliable and was abandoned.