On Mon, Dec 02, 2013 at 11:53:56AM +1300, Michael Hudson-Doyle wrote:
Michael Hudson-Doyle michael.hudson@linaro.org writes:
Antonio Terceiro antonio.terceiro@linaro.org writes:
For 2. I would use a specific device (such as a kvm) with a specific role of "data analysis node" and run my analysis code there. I can't see how LAVA could provide something useful for that (besides documenting this "Use a Separate Data Analysis Node" pattern).
Yeah, this had occurred to me and makes sense. Especially as an extrapolated version of my request might be to generate graphs out of the data, which would require installing packages such as matplotlib....
So I tried this and found something annoying: having this extra device around means that lava-sync is useless in the other bits of the test, because obviously this node that does not do anything until all the other nodes are done does not call lava-sync with the same arguments as the other nodes! I can work around this in a couple of ways for my use case, but it seems a bit of a blocker for a general "Use a Separate Data Analysis Node" pattern.
That's true! :-/
We probably want to use an existing node for data analysis, then, and we are back to your point of how to select a node to do that.
One guideline would be to use the one node that is different from the others in the test, e.g. in a "1 server to N clients" scenario, it would make sense to do the calculations on the server after the actual load has finished. In less trivial scenarios it is's not so obvious where to do the calculations, though.