Conclusions

The objective was to provide Quality of service in a packet switched network. Two numbers characterize quality of service: Loss and Delay. There are a few factors that can influence their respective values:

The load, or the amount of work the server has to do, relative to the amount of time it has is the most important factor. It influences Loss and delay in the following way:

When the load is below 50%, loss is virtually zero and delay very close to the inevitable service delay.
Higher loads will always lead to higher losses and delays

The buffer size (the amount of work or packets a queue can store until the queue has time to service them) also influences loss and delay. However, it serves as an exchange point. When buffer space is added, loss is reduced, but delay will increase. The opposite is also possible, by reducing the buffer size the delay is also limited. Unfortunately, this leads to very high losses.

Other factors will also have influence. However, these two are very influential and exist in every queuing system.

The effect of priorities

The following conclusions can be drawn:

Priorities can improve the Quality of Service for a part of the traffic flow. Improving the Quality of Service one type of traffic receives will always decrease the quality for another type.
LDoLL Extended can provide a better performance for voice. It is possible to decrease the average delay. Also, a large part of the traffic will always have a delay lower than the average.

When bandwidth is reserved (like in MFQ), guarantees can also be given for well-behaved traffic (like constant bitrate). However, using time-slices can introduce strange effects in the final delay characteristics, so the size of time slices has to be chosen carefully.

A side effect of absolute priorities (as in VFQ) is that there is hardly any possibility to differentiate in service. One of the nice effects of LDoLL is that by adjusting the threshold setting, the delay and loss of the low loss traffic can more or less be fine-tuned. The resulting performance remains dependent of the load, but the relative performance of the traffic classes remains the same.

A major advantage of LDoLL Extended is the flexibility in assigning priorities. Although the link between LDoLL Extended's parameters (Threshold and time-slice settings) and service- and storage priorities is not completely one to one, loss and delay are adjustable over a wide range.

LDoLL Extended cannot guarantee the same performance as VFQ can for the traffic that receives absolute priority. However, with the proper settings, LDoLL Extended can approach the performance of VFQ very close, yielding delay on the order of the service time of the largest legal packet size.

The performance of LDoLL when using self-similar traffic

It is clear that the LDoLL principle is effective when using self-similar traffic. Although the effects of LDoLL are smaller when using Pareto distributed inter-arrival times (Compared with negative exponential inter-arrival times.), the effect still exists. And just like when using Poisson traffic, the effect depends on the level of the threshold.

Extending the LDoLL principle to more queues, as is done with LDoLL Extended is possible and more importantly, the principle remains useful. The only difference between normal LDoLL and LDoLL Extended is the level of the threshold.

Conclusions

The effect of priorities

The performance of LDoLL when using self-similar traffic

Quick Jumps