Links to carriers

We have links to back-haul carriers BT and TT.

Our target is that we operate a uncongested links to the carriers. This means that if the general trend is that the link will be getting full we order more capacity from carriers. The capacity we order is expensive, takes time to change and has minimum terms. This means it can be difficult to perfectly manage the traffic. We allow some head room but not more than necessary. It is technically impossible to guarantee that the link is always uncongested (not without having a pipe the size of all internet links in the world added together) but by careful monitoring and allowing enough headroom we can aim for that target.

This table shows an analysis of how well we meet our target.

Unerrored seconds

The unerrored seconds report is the simplest statistic to understand. If we drop even a single packet in a second then this counts as an errored second. Given that we can be handling hundreds of thousands of packets every second, this can be a very sensitive measure of congestion. Bear in mind we will drop the larger packets first which will normally be TCP which re-sends the dropped plackets. This helps ensure VoIP and interactive uses of the internet are unaffected.

Daytime unerrored seconds are based only on 9am to 6pm Monday to Friday

Non-premium

We also consider 100 second samples. If we have reduced the throughput of non-premium lines at all (as a result of dropped packets for several seconds in a row) then the whole 100 seconds is considered to have affected non-premium customers. Because we are looking at 100 seconds at a time it is possible for very short busts of traffic to cause this to show a worse figure than the unerrored seconds. Even if traffic was reduced by 0.5% for part of the 100 second period the whole 100 seconds counts as errored.

Premium

Considering the 100 seconds samples, there is a limit to how much we will reduce the non-premium customers. If that limit is reached then we consider the whole 100 second sample is having an impact on premium lines. This is a sign of much more severe bursts of congestion for several minutes. Bear in mind, a single 100 second sample in a day is more than 0.1% of the day, and even if this does happen VoIP and interactive services should be unaffected.

Method of analysis

Every line is sent an LCP echo every second and loss and latency measured. This is summarised per 100 second sample and archived for every line for each whole day. The previous day is then analysed each morning and the above table is updated.

Each line is considered to pass, fail or be inconclusive. Inconclusive includes lines where samples have too much upload or download which could themselves cause higher loss or latency to be observed. Lines with less than 18 hours on-line are excluded totally.

Latency is considered against a reference base-line latency for the circuit, and this has to be present for several hours to be considered. As such, lines without clean latency monitoring (some makes of router) are excluded.

For an area to show in the above table there must be no lines that pass at all, and at least 50% of the lines must fail (the rest being inconclusive). Where we have very few lines we require 80% to fail (and none to pass) to show on this list. The area must also have enough lines, so some smaller exchanges will not be detected.

The size stated is what percentage of our total lines are in the affected area.

The block graph under each report for yesterday is based on an average of valid samples for each line for each hour, and then taking the lowest value of all lines in the affected area, so the loss and latency shown is something that is affecting every line in an area.