Server Fault is a question and answer site for system and network administrators. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

First of all, apologies if this is the wrong StackExchange site for this question - it's about Ethernet network capacities, but not specifically about servers etc.

I am designing a network for a data acquisition system that will be outputting data via TCP at just over 30 Mbps. (To ward off the obvious first comment, this is definitely mega*bits* per second.)

I recall hearing somewhere that one should aim to keep a network's normal utilisation under 10% of its capacity, but I can't find any proper research to that effect. Is the 10% figure reasonable, and if so, is it appropriate for my data acquisition system or is it a figure that is intended for e.g. corporate networks that will be more 'bursty' than my constant 30 Mbps?

Would it be better to use Gigabit Ethernet, which would be running at about 3% capacity, compared to 100baseT devices that would be running at 30% capacity?

share|improve this question
up vote 2 down vote accepted

In my experience, you should try to keep your most-utilised link at "less than 90%" at peak. As to what your expected differences between "typical" and "peak" are, you are better placed than I am to answer that.

I guess "normal utilisation under 10%" is for networks taht experience a drastic difference between "typical" and "peak" or is based on large collision domains (coax-based and wifi, typically), where you can expect a "perfect" utilisation of roughly 80% (after that, retransmissions are taking up a larger and larger proportion of the available bandwidth, until all you have is essentially only retransmissions).

If it is a pretty constant 30 Mbps and not sharing infrastructure with a "bursty" network, I suspect 100-BaseT should be fine, but I'd also make sure to have monitoring set up (this will eat a little bandwidth and CPU on the network elements, but it's well worth it), that way you should be able to plan an upgrade to faster network links well in advance of your needs growing to the full capacity of the network.

share|improve this answer
+1: But 'peak' is less, sometimes a lot less than the theoretical max throughput. – symcbean Feb 23 '11 at 11:21

Statically, stay away from sizing to fail (;-)) As with everything, if you use more that 80% of your theoretical capacity, probability will make too many packets arrive at the same time and you'll zoom well past 100% demand on an instantaneous basis. [Really old "pup" Ethernet would misbehave at 30% utilization, causing an urban legend that that ethernets had to be over-provisioned massively]

Technically, you need to eliminate bufferbloat (see anything by Dave Taht), before doing anything dynamic.

Dynamically, after you have removed the hidden bottleneck of bufferbloat, plot packet drops. IP will drop packets to keep the actual maximum as close below the theoretical maximum as possible, and drops are how it does it. Every drop is an instantaneous overload being fixed. [This makes queue models tricky, by the way]

share|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.