MSFT boffins bust mobile data bottlenecks with iOS app

Microsoft Research boffins in Singapore and India have put up a novel idea: an app designed to identify where a smartphone's downloads are suffering and route around the problem without bothering the network operator.

The researchers reckon the mobile network has forgotten one of the prime directives of the Internet: to route around failure.

Moreover, the assumption that slow downloads are always caused by a congested base station isn't always true – and focussing on that link might mask problems, and possible fixes, upstream.

In a paper published here, researchers from the National University of Singapore (Nimantha Baranasuriya and Seth Gilbert) and MS Research India (Vishnu Navda and Venkata N. Padmanabhan) suggest that at a time when mobile networks deliver data they need to be more responsive to where bottlenecks are happening.

Testing a tool they call QProbe, the researchers find that the last hop is responsible for poor performance in nearly 70 per cent of 3G bottlenecks identified – but when you move to a 4G network, the last hop is only the bottleneck in 25.7 per cent of cases.

That offers a variety of strategies for addressing poor performance, particularly as the world moves to 4G, where the bottleneck might be in the WAN link rather than the last hop.

For example, they note, someone's downloads might be suffering because the content distribution network (CDN) server they're connected to is overloaded. In that case, switching to another server is a better solution than (for example) connecting to Wi-Fi instead of using the 4G link.

"On the other hand, if the cellular link is the bottleneck, the user may have to look for an alternate connection (e.g., WiFi) or have the application adapt (e.g., by downsizing media content)", the paper states.

While congestion identification is a well-understood problem on the Internet, things work differently on cellular networks, something the researchers had to take into account.

Cellular networks, they explain, use "proportional fair" scheduling in each base station, whereas the techniques used to identify bottlenecks in Internet routes assume FIFO queuing.

Instead, in QProbe the researchers send out probe packets from the user end – not too many, because mobile plans are capped – analysing things like timing and hop count to decide whether a bottleneck is caused by the last hop or the WAN link.

The probes produce a metric they call "stretch factor", which put simply describes how much a 3.5 kB packet sequence gets stretched between transmitter and receiver.

On a congested last hope, the researchers say, latency will be high but the stretch factor will be low. That's because the base station's queuing mechanism will save up all five QProbe packet trains and deliver them in a burst.

On the other hand, if the problem is further into the WAN, the packets are likely to get spread out, and a large stretch factor results.

QProbe was tested on applications deployed in Microsoft Azure servers in the USA, South America, Europe, Singapore, China, Japan and Australia, with QProbe servers on 51 PlanetLab servers.

The researchers' conclusion after a two-month test with more than 600 users in 33 countries (and on 51 carriers) is that the probe located bottlenecks in 3G and LTE networks with 85 per cent accuracy.

So far, the researchers have only published QProbe as an iOS application. ®