next up previous
Next: Reliability and end-to-end flow Up: Protocol design Previous: Splitting messages into packets

Optimising throughput

The protocol supports the adaptive routing capabilities of the C104 switch to improve the network throughput performance[6]. However, adaptive routing can cause packets of the same message to arrive out-of-order at their destination.

The communication protocol of the T9000 transputer[18] uses a strict acknowledgement scheme to avoid out-of-order packet arrival. This scheme limits the maximum message throughput[7] due to the latency of the acknowledge. Measurements have shown that two T9000 transputers, interconnected via one C104, can achieve a throughput of 4.7 Mbytes/s over a single virtual link. Interconnecting them via five C104s, which increases the latency of the acknowledge by 4  s, reduces the maximum throughput to 2.3 Mbytes/s.

We use a sliding window protocol[19], together with an acknowledgement scheme without the limit in maximum throughput: each acknowledgement packet can acknowledge a sequence of packets. The protocol does not use piggybacking because we do not expect a performance gain from it, since DS link networks perform well for small packets, and the protocol requires only a few acknowledgements per message.



Marcel Boosten
Wed Mar 11 14:25:07 MET 1998