I realize it varies massively (my current host affords me 1,000 GB a month and I'm lucky if I hit 10GB right now... but I also know people who push 3TB/month out of a single machine), but I was wondering about aggregate averages.

Has anyone ever found specs on what the actual average bandwidth per dedicated server is, across a data center? In other words, if I rented out to customers 100 dedicated servers with a 1,000 GB bandwidth quota per machine, what would my actual bandwidth usage be? Roughly 300 Mbps sustained if we assume flat traffic and that everyone uses exactly 1,000 GB, but I'm well-aware that both of these assumptions are dead wrong. And the fact that the two models go in opposite directions complicates guesses, too: since bandwidth increases as certain times, I might be looking at bills for more like 500 Mbps at 95th percentile. But at the same time, 1,000GB is more than most people use, so 100 Mbps might be overkill.

As with my question about finances, I'm not expecting anyone to reveal precise business benchmarks to me. I'm just wondering about a ballpark for actual usage data.

This is a very good question. We individually monitor dedicated and colo equipment through a tweaked version of Cacti. Offers fantastic reporting on 95th percentile bandwidth, CPU utilization, available RAM, etc.

You might think it would be enough to total the 95th reporting for each server, and then get a read for the total bandwidth flowing through a given subnet. But that is not the case.

You have to look at the total aggregate of all hardware on that subnet. As an example, three servers, each using 1Mbps 95th percentil bandwidth, does not mean it will total to 1Mbps + 1Mbps + 1Mbps = 3Mbps. More likely it could be as low as 1.3 or even 1.5Mbps. Why? Because the aggregate of all traffic is, itself run against the 95th percentile.

And that can be dramatically different than each individual machine. Further, we've found that for smaller customers, 95th percentile can be quite misleading. A few spikes here and there, and the reported bandwidth is much higher than actuals. On the flip side, much higher-traffic clients see a number of benefits from 95th percentile reporting, in that the overall higher level of traffic does help balance out traffic spikes (as there just seems to be fewer spikes and, as such, a lesser likelihood of pushing the 95th up very much at all).

So ultimately, to answer your question, the "actual bandwidth" or "average bandwidth" for a dedicated or colo customer largely depends on the usage profile (the usage habits) of your average customer:

1) do you typically host many, low-traffic, dedicated customers?

or

2) do you typically host higher-traffic customers?

Depending on the answer, you could find yourself with a lot of surplus bandwidth (due to 95th inaccuracy), OR you could find yourself actually having pay your fiber carrier 100% of the bandwidth your customers are churning through.

If that didn't make any sense, it's because it's been a very long day -- and so I offer any due apologies as necessary. :)

Thanks,

David