I went to the Ottawa Linux Symposium and heard several good talks on testing and performance. One of the outstanding ones (to me) was the University of Waterloo and their testing apparatus and harness for doing testing for performance and the application of statistical analysis to the testing, as a one-time test may not show the true results.
They have a mechanism, and a lab, that allows people to build code for testing, run the tests with supplied data, and then statistically analyse the data to see.
It has a front end that allows you to supply the information necessary to build your project, then distributes it to a number of different machines to run for a selected number of times. Even though you may be running the same data through, they can show that various other factors will vary performance as much as 15% on a given run of the test data.
I asked them for a copy of their slides, which I have uploaded to a Google Drive and will make a link available to anyone who wishes to see them.
My thoughts on this are that Linaro might want to study what they have done and build this type of analysis into Lava, or perhaps the University of Waterloo would like to incorporate parts of Lava into their back end, or the two groups could work together. I have sent them pointers to Lava web pages from the Linaro site.