123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384 |
- {% extends "base.html" %}
- {% block content %}
- <h2>Why look at latency?</h2>
- <h3>The problem of peering quality</h3>
- <p>Determining what is a "good" peering in dn42 is quite difficult: many
- criteria come into play, such as latency, jitter, capacity, packet loss,
- stability, or even the price your ISP will pay (peering vs. transit).</p>
- <p>On the other hand, a "bad" peering is easy to picture: if you are in
- Paris and peer with somebody in Australia, then you might end up doing
- Paris → Australia → Hamburg if you want to send packets to Germany. This
- does not feel very efficient. People usually solve this problem with
- policy routing (local preference and path-prepending). But it's still a
- good idea to build good links and avoid terrible links.</p>
- <p>Of course, you need to build long-distance links sometimes. Otherwise,
- dn42 would be made of small, independent islands. This tool can also help
- you to choose the best long-distance links.</p>
- <h3>Looking at latency</h3>
- <p>Latency is actually a good enough indicator of "distance". For
- instance, two machines located at the same ISP are expected to have low
- latency towards each other. On the other hand, a latency above 200 ms
- usually indicates that the two machines are quite far away geographically
- (but not always).</p>
- <p>Additionally, latency can vary widely for long-distance links,
- depending on the quality of transit and peering agreements between ISPs.
- For instance, to reach a specific destination in Singapore from France, we
- have the following latency as of September 2014:</p>
- <table>
- <tr>
- <th>ISP</th>
- <th>Latency</th>
- </tr>
- <tr>
- <td>Online</td>
- <td>177 ms</td>
- </tr>
- <tr>
- <td>tetaneutral.net</td>
- <td>264 ms</td>
- </tr>
- <tr>
- <td>SFR</td>
- <td>267 ms</td>
- </tr>
- <tr>
- <td>Free</td>
- <td>365 ms</td>
- </tr>
- <tr>
- <td>OVH</td>
- <td>402 ms</td>
- </tr>
- </table>
- <p>Of course, this is only a snapshot, and reflects the situation for a
- specific source and destination. Still, the latency more than doubles
- depending on the ISP, which in this case strongly favours a peering with
- somebody hosted by Online instead of OVH.</p>
- <h3>Situations this tool aims to solve</h3>
- <p>To sum up, this tool can help in several situations:
- <ol>
- <li>Detecting when somebody is in the same datacenter as you, so that
- it's mostly free to peer</li>
- <li>When you are far from everybody, find the peering with the lowest
- latency</li>
- </ol>
- </p>
- <p>By building low-latency links in dn42, it's actually possible to have
- lower latency in dn42 than over the Internet, for the same destination
- (it's called <em>detour routing</em>).</p>
- {% endblock %}
|