This post is a summary of two papers co-authored by Amreesh Phokeer and Josiah Chavula of the Research and Innovation Department and were presented at:
- IEEE Africon 2017, Cape Town, South Africa - Won the Most Outstanding Paper Award
- IEEE Infocom 2018, Honolulu, Hawaii - Won the Best-in-Session Presentation Award
We often reduce Africa to a single country or economy but we tend to forget that it is made up of 54 different nations, more than in any other continent of the world. Africa is therefore a complex set of countries having different levels of connectivity, Internet penetration and infrastructure. The good news is that in the recent years, there have been massive investment in both undersea and terrestrial fibre optical cables, and as a result, Africa is one of fastest growing continent in terms of international internet bandwidth - growing at a compound annual rate of 44 percent from 2013 and 2017.
Internet penetration in the world
However, as revealed by this study, Africa remains significantly behind other regions in terms performance and quality of service. This obviously has many implications on the Quality of Service (QoS) of networks in general but also on the Quality of Experience (QoE) from the end-user’s perspective. While some parts of Africa are performing very well, we see considerably bad performance in different subregions and specific countries such as Angola and Ethiopia. To achieve our study we used SpeedChecker probes found in 52 different countries (~ 850 probes) as vantage points and OOKLA Speedtest servers as targets (~213 servers) in 42 different countries. In this study more than 300 ASes were probed which yielded 42K RTT samples and 31K traceroutes were captured over a period on 3 months.
What is the inter-country delay in Africa, and how is this impacted by topology and interconnection strategies?
Left: Location and number of SpeedChecker probes Right: Location of SpeedTest (OOKLA) servers
We executed the experiment 4 times a day to make sure we are gathering the data both in peak and non-peak hours (00.00, 06.00, 12.00, 18.00) to randomly selected African SpeedTest (Ookla) servers. We did the same for traceroute and analysed the different hops the packets were taking to reach the final destination. To determine which ASN any of the hop was, we used the RIPE RIS database and MaxMind GeoLite2-City service to geolocate the ASN. Geolocation helped us to understand the amount of tromboning happening in African networks.
Mean RTT between countries
With the RTT samples collected, we could calculate the mean in-country latency of all the networks probes in Africa which is around 78ms. As a matter of comparison, the mean in-country latency in Latin America is almost similar at 76ms. However in North America and Europe the average in-country latency is less than 45ms. This shows that there is still some work to be done to reach a low latency threshold on average. But this task is daunting! The main reason being we see lot of dependency on foreign upstream providers. By extracting the the first hop (different from the source hop) in a traceroute, we could infer who is the upstream provider. We found that more than 37.8% of all the traces captured, had their 1st hop outside of Africa.
Additionally, by analysing the traceroute samples collected, we found that almost 50% of the traces have their 1st hop outside of Africa. This shows the poor level of interconnectivity in between networks in Africa. Obviously, this result varies in between regions. For e.g. a big chunk of the traces shows that packet from the Northern region goes through Europe before going to other regions in Africa. This shows the clear preference of using overseas upstream providers for international connectivity. This phenomenon can also be explained by the prevalence of a very competitive market in Europe, the geographical proximity with EU countries (data centres) and the availability of undersea cables and landing ports in the Mediterranean. Finally, we have also seen that 14% of the traces captures have packets that go through more than four different hops before reaching their destination.
Usage of foreign upstream providers
Another notable finding of this study are the latency clusters. By running the latency matrix through a clustering algorithm, we obtain clusters of latencies. It basically means that countries in the same cluster share more or less the same latency between one another. We found four main clusters, together with some interesting corner cases. There are countries in East Africa clustered with Western African countries (for e.g. Somalia), which means that it is faster to send traffic from Somalia to Benin rather than to Kenya for example. Reasons can be multiple, but we suspect poor traffic engineering and lack of peering in between those countries. In that case, traffic will usually take international path, with added RTT overhead. Similarly, Madagascar is clustered with Northern Africa, while it is closer to South Africa.
This study shows that not all regions in Africa benefits from the same networking conditions as they prevail in the Northern, Southern and Eastern Africa. The availability of undersea cable capacity is definitely a game changer as we observed landlocked countries generally experiencing poorer latency level. But there are other drivers, such as a more open and competitive market and more peering in between African networks, as we see in the Southern and Eastern part of Africa. The issue however remains content as most African networks get their content from overseas (EU and US), which could explain the reason why there is little interconnectivity in between regions. There is no silver bullet for this situation, other than encouraging more deployment of local capacity such as terrestrial and undersea cables, data centres as well as reducing the price of access.
Four main clusters