confidence in a sample size Firstly, I am not a math expert so I have limited knowledge and language here. bare with me :)
Currently, I am gathering metrics about servers that exist in a datacenter. the situation is that there maybe 100 servers of the same model that i need to sample. Yet I can only collect metrics on 50 of them for reasons that are not relevant. One metric happens to be power. So I collect power consumption on 50 servers and average the power data to come up with a number like 500 watts for example. So the average power consumption of a particular model of server is 500 watts. There are a few questions:


*

*Is there anyway to know that the sampling I have is significant enough to be considered? 

*How confident/probable am I that the next server I install of the same model type will be around 500 watts? 


I've been reviewing statistical significance, confidence interval, tolerance interval and the like but I want to be sure this is the right way to go. Maybe I'm completely off target on my research. So I am looking for some information to at least point me in the correct direction. 
 A: First off, all those confidence intervals usually assume that you take random samples from the population. It sounds to me like your are not selecting your servers randomly to collect measurements, but rather, "for reasons you think are not relevant (but maybe are)".
So before you compute any confidence intervals, you should either be certain that the servers you sample are no different (on average) than those you cannot sample---so look at those irrelevant reasons. Or, if you cannot, you need to control for that difference somehow.
In the first case, where you think sampled and nonsampled servers don't differ, you can indeed compute confidence intervals fairly easily. First, you compute the sample mean as usual, then you compute the standard error of the mean (using finite sample correction, since you want to sample about 50% of your population) and on this basis construct confidence intervals (CIs) using the t-distribution.
CIs can be used to answer your 2nd question. Suppose your standard error is 15W. Then you can construct a 95% confidence interval as
$$CI=\bar{x}\pm 1.96 SE=500W \pm 1.96*15W.$$
The 1.96 factor comes from the fact that, if the measured values are normally distributed (which they are for sufficient sample size), then 95% of the values lie within 1.96 of the standard error above and below the sample mean. As I mentioned, if the sample size is relatively small, you should instead use the t-distribution (then the factor will be $>1.96$, i.e., you estimate the values are more spread out). To sum up, after this calculation, you can choose confidence level $l$ yourself (in the example: $l=95$) to calculate the range in which $l$% of the server measurements will be. If you did everything right, not only do you know your confidence ($l$), but can also quantify "around" with the CIs. All of this, of course, relies on random sampling.
I don't understand your first question. Significant sampling? If you sample 50% of your servers, that would be a lot and you need not worry about small sample size. The more important thing about sampling, as stressed above, is that it is random, to rule out differences between the sampled and nonsampled units. We recently had a thread discussing what can happen with nonrandom sampling.
