Re: Mean vs Median
Paul Emmerich <[email protected]>
| Newsgroups | gmane.ietf.bmwg |
|---|---|
| Message-ID | <[email protected]> |
Hi, On 10.11.15 08:15, Marius Georgescu wrote: > For physical testing, the distribution tends to be normal (at least for my testing experiences). In this context the difference between mean and median would be negligible. > Since in the virtual world (the future) bimodal is the new normal (if i may say so :), median seems like a better choice. > I fully agree we should have an additional statistic, accounting for the variance somehow. I proposed the Margin of error. It can be the standard deviation, the standard error … normally distributed latencies are actually pretty rare when testing devices that do any part of the processing in software. It doesn't even have to be virtualized. The main reason for this is batch processing of packets which drastically improves the performance. Simple batching leads to a uniform distribution. However, there are different algorithms trying to optimize for different things in a typical software router which leads to really weird distributions. For example, Linux uses polling in the NAPI to reduce interrupts and increase the batch size, but many drivers like Intel's ixgbe also feature a second interrupt moderation algorithm that further reduces interrupts. Both systems interact and you get weird distributions. You can read more about interrupt handling for network processing in Linux in my paper about it at [1] if you are interested. (However, the paper does not yet look at distributions, it only states a few percentiles.) I've uploaded an excerpt from one of my papers here [2] which shows the latency of a Linux system running Open vSwitch without any virtualization. An excerpt from the same paper with virtualization is available at [3], this looks more like a normal distribution with a long tail. (The paper is currently under peer review for a journal and will hopefully be published soon, I can share a pre-print privately with anyone who is interested in the full paper.) Paul [1] http://www.net.in.tum.de/fileadmin/bibtex/publications/papers/SPECTS15NAPIoptimization.pdf [2] https://www.dropbox.com/s/dl9tmslubyd2j7m/Screenshot%202015-11-11%2010.17.05.png?dl=1 [3] https://www.dropbox.com/s/tia2lwhlt08hl8l/Screenshot%202015-11-11%2010.24.57.png?dl=1 _______________________________________________ bmwg mailing list [email protected] https://www.ietf.org/mailman/listinfo/bmwg