In this experiment, we test a scenario where network bandwidth is higher than what end-systems can consume, and test the forwarding processing delay caused by ALMI processing. We used a Sun Ultra-1 attached to a 10Mb/s Ethernet network as a source sending data to several Pentium III - class PCs connected over a 100 Mb/s LAN. We vary the number of intermediate data relaying hops and measure the throughput at the last hop. In this experiment, we use TCP connection between nodes and confine the controller to connect members as a chain in order to capture the effect of ALMI member node forwarding.
From Table 1, we observe that the throughput achieved in all cases remains stable regardless of the number of intermediate hops. This shows that ALMI processing delay does not increase with the higher number of data relaying hops. From a scalability point of view, this means that the overall TCP throughput achieved in a session is decided by the slowest network path or intermediate hop, but is not affected by the aggregation of bottlenecks if there are multiple. On the other hand, if we look at Table 1 vertically, we see that the processing delay associated with each packet is relatively high, especially for small size packets. This is due to the fact that the Java virtual machine is still comparably slow even in the presence of JIT. However, we believe this gap will be reduced in the near future with the advances of better compilers and faster CPUs.