brainstorms (Re: dns based loadbalancing/failover)

Date: Sat, 6 Oct 2001 22:52:05 -0400
From: dsr@akamai.com

> If heavily enough distributed, congestion should be highly
> localized... if present at all. Let's assume that a "basic
> server" can service 10 Mbps of traffic. Install servers in
> various points, a la Akamai... any traffic sinks here ever manage
> to overload your Akamai boxen? If so, how often compared to a
> random traffic source.

Please don't start attacking your local Akamai box to prove
this, though. Brute-force is so inelegant.

!!!

My intended point was that Akamaized sites are less likely to be
overloaded than non-Akamaized sites. Traffic spikes from
incidents such as the release of the Starr report and the
bombings last month have demonstrated this.

Given enough hosting points, overloading becomes rare; I used
Akamai as a real-life example. Given that overloading is rare,
we can focus on failover, and worry less about load-balancing.
Given that we need only to focus on failover, the problem (and
hopefully the solution) becomes much simpler.

I'd hope that the average NANOG reader knows better than to try
any "let's see how much traffic before this breaks" stupidity...

Eddy

There are limitations. It moves the choke points around, in particular
the choke points between providers (upstream, downstream, and peers).
Overloading is a problem, but it is a distributed form of overloading
and harder for people outside to detect. Akamai gives excellent
protection from overloading as detected by services like Keynote.
To really measure how well it worked, you need a measurement
infrastructure as distributed as the delivery structure.

During the first couple of hours, the systems appeared to need a
lot of human intervention. Its unclear to me, as an outsider, what
happened during those first few hours. Was it a problem with people
not using the tools as well as they could, or if things needed to be
re-engineered on the fly.