En studie av flytkontroll og bufferstørrelser i Ethernet som tett koblet nettverk

Abstract

In this thesis we study Ethernet as a technology for building interconnection networks. We start by changing the routing tables of traditional Ethernet to better utilize available resources in the interconnection network (in traditional Ethernet a link is not used if it would close a loop in the network topology). We then continue with a study of buffer usage and flow control in interconnection networks using input and output queued switches.

Through simulation we show how we can improve both throughput and latency in traditional Ethernet by taking control of the routing tables populating them using a version of up*/down*-routing. In doing this the throughput is as much as tripled while the mean latency is cut in half.

We continue with simulations testing throughput and latency in interconnection networks using input queued versus output queued switches (with flow control) concluding that the difference in performance is neglectable. While the problem of head of line blocking has been argued to lower performance of a single input queued switch (compared to an output queued switch) all our simulations show that this is not true for a network of switches. The performance is more or less the same for the two types of switches.

The last part of our studies focus on when to turn on and off pause based flow control. Our findings show that you can expect a small increase in throughout when keeping the fill ratio of the buffers high (and using big buffers). The penalty when it comes to mean latency however is huge when using big buffers with high fill ratio. The conclusion for interconnection networks is therefore to keep the buffers as small as possible still keeping a good flow of traffic. In our simulated networks this compares to using buffers in the size of two to four package sizes (using store and forward switches).