SoldierNET - Quality of Service (QoS) Management

SoldierNET installs a server on each site. One of the jobs the server does is to try and optimise the service that users get by balancing their different uses of the internet.

Outgoing priorities

The server is able to drastically improve the service that users get by prioritising outgoing traffic. When a large webpage is requested the website that is sending the page will break it up into many large packets. It will send a packet then wait for an acknowledgement from your PC/laptop before sending the next one. Any delays in sending these ACK packets have a drastic effect on the download speeds that the user will receive. The download can be very slow even though there is lots more download capacity available and unused.

Small Packets

The ACK packets are actually very small since there is hardly any payload in them. Just by putting these first in the outgoing queue can have a significant impact on the download speeds achieved.

Download Speed Monitoring

Every 5 minutes our servers download the home pages of 4 of the most popular websites. If all 4 sites show delays in download speeds then this suggests that our service is suffering. If just one site is showing a delay, then this is normally the website and not the local SoldierNET service that is being slow. If several SoldierNET sites are showing delays on the same website at the same time then it is easy to confirm that the delay is caused by that site. We regularly see the BBC site suffering delays at the same time on all of our networks which indicates that the www.bbc.co.uk site itself is slightly overloaded.

Bandwidth Monitoring

Our servers also monitor the outgoing and incoming bandwidth usage. It is possible to see when a site is using all available bandwidth or not.

Impact of QoS on Download Speeds

The following 2 graphs clearly show the impact that turning QoS management on has on download speeds. The first chart shows the bandwidth being used for downloads and uploads. The area above the horizontal axis is the download and below the axis the upload. This test site has a non-symetric service from the ISP i.e. the upload/outgoing limit is lower than the download limit. The maximum upload is about 130kbytes/sec. The second chart shows the download speeds of the 4 monitored sites.

At the start of the charts you can see a large download in progress. This runs from about Tue 1930hrs to 2300hrs. There is still some unused download capacity even during the large file download. The outgoing traffic is saturated and this is the bottleneck. During that time the dowload times for the 4 monitored web pages is very slow. The average is between 5 and 10 seconds and often peaks at much longer than that. Since all 4 pages are slow we can tell the problem is on our network not theirs. The maximum download times are pretty poor at 20-40 seconds and up to 490 secs for one BBC page.

At point 1 (Tue 2300 hrs) the download is finished. There is lots more unused download capacity, however the average speed of web page downloads is unimproved although the very large peaks have almost stopped. Performance of web page viewing is still really poor. Almost all of the outgoing bandwidth is file sharing (bittorrent) and this is preventing the small ACK packets from getting out quickly enough. We've left the tests running through the night, when no-one else is using the network and even then the download speeds are poor.

At point 2 (Wed 1000 hrs) we turn on the Quality of Service (QoS) management. You can see an immediate improvement in download speeds for web pages, with the slowest being 1.4 seconds. The outgoing bandwidth is reduced by at most 10% to ensure there is capacity for all the ACK packets and some others that significantly improve performance (such as DNS). There is still the occasional peak, especially on the BBC site. The peaks are now resticted to one site at a time which means that the problem is not on our network but theirs. We have found that the BBC site is so popular that it often suffers from such blips in performance. During some sporting events where coverage is available on iPlayer performance at the BBC suffers due to large numbers of users nationwide watching at the same time.

Multiple Protocols

The example above is based on bittorrent and simple http traffic (web pages). The QoS management facility configures and optimises about 12 different protocols and services.