Internet Latency Part II : A casual look at Carriers and ISPs

The objective here is to? illustrate how performance can vary on the open Internet – showing that some smaller, regional vendors can deliver quality equal to some of the largest, reputable organizations in the market. We carefully reduced many thousands of Host names to a list of just over 60 organizations to which they belong. Averaging their performance over a seven day period seems to produce three different categories of performance. The first group averages low-latency (RTT), with few or no observed spikes. These organizations perform consistently well and seem to be both well-engineered and brilliantly led. The second group averages low-latency, but with occasional RTT spikes into questionable territory where a brief degradation might be experienced by Users. The third group is consistently horrible – with avg RTT above 400ms and spikes north of 2,000ms ! These organizations connect through a much higher average number of hops – going back to an issue we are trying to solve here. Recognition goes to the University of Toronto Computer Science Entrepreneur program, and the privilege of working with team leader Annanya Sharma in the design and promotion of this work.

We survey the open Internet - those hops across network backbones and the Core Layer that connect Users to their services. You may be tempted to think that it's a level playing field but some Internet Carriers appear to perform much better than others. We also discovered (without being able to explain yet), that connections under ~13 hops tend to have much lower latency. And connections going into the 14 - 20 hop range often encounter disproportionately higher latency (often 1,000 to 3,000 ms!). In other words, a shorter, simpler path from your Device to the Internet Service you connect with, seems to be the best guarantee of good, consistent performance.

We test a regular group of over 30,000 high-profile networks, and another group of millions of random IPs to build a profile of the open Internet. What's interesting is that when you break these connections down by hop, and identify the transit Device, less than 5,000 Device names are showing up. There are some routers out there through which dozens of ASNs connect to thousands of different destinations. Of the millions of tests we run every week, around 50% are routed through just 21 different Organizations. In other words, there could potentially be a lot of bottlenecking out there.

Here is a sampling of nineteen, low-latency, low-spike networks or service providers. Some are well-known while others are local/regional vendors with limited resources. This should help reinforce the idea that a smaller, well-managed network can perform significantly better than some of the bigger players. Before looking at the top performers, let’s have a quick look at a couple of the worst for comparison…

And a brief overview of selected top performers:

Over 4,000 tests to over 1,500 IPs on the Bell South network averaged 8ms while traversing ?up to three different ASNs to make the connection. No test exceeded 62ms.. very impressive.

Given the size of the Telus network both in subscribers and geography, it may not be fair to generalize their performance based on over 800 tests to 185 IPs. But 13ms average RTT with no real spikes does speak well of their network. We have watched them for many months now, and Telus rarely comes up with any kind of discernible issue.

We found T-Mobile to be performing around the same level as Telus – an excellent 13ms average from over 2,200 tests. Their average hop-count is only slightly higher.

Yyz.as54203 part of Strong Technology (ASN54203) – US based and with mixed reviews on their “Best streaming VPN” service. We tested 21 Ips in the Toronto region thousands of times at an unbelievable 1-ms avg latency. This may underscore the observation that the same organization performs differently in different markets.

Lucent may have merged with Alcatel over a decade ago, but their name still shows up in many host services. We tracked over 1,110 of their IPs connecting mainly through Level3 Devices at an impressive avg. 6ms. Nokia appears to be the current owner.

One hundred twenty IPs belonging to MyCingular tested at 9ms over the seven day period with a max observed spike to 88ms. Impressive.

Terrahost maintains Datacenters, Vitrual and Dedicated servers, cloud and unmanaged services. Over 6,000 tests to 27 IPs averaged10ms with just one observed spike t0 54ms. They tend to connect through Level3 Devices for clean, consistent performance.

We identified Asymptote ( AS53616 open peering) as a hostname for 36 IPs but exact ownership isn’t clear. Names like Velcrom (ONTAR-40), Ace Data Centers and others show up here. They tested consistently at around 11ms with one spike to 320ms (not bad).

Based out of Indiana, SurfAirWireless provides both fiber and wireless Internet with consistently great performance averaging 11ms over the entire week. They appear to have a lot of control over the Devices connecting Users to their network and a very low average number of hops to connect through.

Michigan-based Metro Wireless International? tested at an avg 17ms in over 2,000 tests against 27 of their IPs. They focus on Business Internet solutions, served by a dedicated high-energy team of entrepreneurs.

With locations in the US, Canada and Europe, Server Mania averaged a respectable 17ms in nearly 10,000 tests to over 2,300 IPs on their network. Connecting through Cogentco, NTT and their own backbone network,

Florida-based DCM-Cable averages good numbers without any concerning spikes over the week. But they may have changed ownership? Some of the 31 IPs with their name include references to Cogent, PSINet (serious?), and sometimes DCM-Cable. Connecting through Level3, NTT and twelve99 they average a higher number of hops per connection.

Centurylink Internet… over 2,000 tests to 468 IPs in their sizeable network at a respectable 18ms with no concerning spikes is above average. Connections (hops) are through reputable twelve99, Level3 and Comcast devices

CloudWebManage-Canada is a hosting provider with a safe, low-risk reputation, averaging 19ms on tests to 54 of their IPs.

Linodeuser (LINODE-US) is part of Akamai which is rated separately in this report. They give developers a place to build, run and secure their apps. Users may connect through Devices owned by Akamai and/or twelve99. The 4,500 tests we ran to over 400 of their IPs yielded a low 21ms average with some limited spikes above 200ms.

Rogers-Cable enjoys a clean, high level of performance that we have observed for many months, averaging 21ms. They rarely show up in any analytics related to network issues on this system.

Whitesky Communications offers Cable, Voice and Internet services for residence and businesses. They partner with residential developers (student housing, apartments, hotels etc..) and interconnect through a large number of other providers including Zayo, twelve99, ntt, Lightower, Comcast and others. They drive excellent numbers through this complexity. We ran nearly 16,000 tests to over 270 of their IPs averaging 22ms with only some limited spikes in the 200ms range.

Cox also tested well – over 4,100 times to 550 IPs averaging 24ms with little or no deviation.

Comcast users seem to connect directly to their massive network (no surprise) with an average 27ms RTT latency. The hop count in these cases is distinctly above the overall average in our work. Do the high hop-count and the integrity of the Comcast owned and managed backbone balance each other out to produce these good results?