> While I agree with the LGTM to keep the test focused on testing just one
> thing, the fact that this issue occured in our QA means that we have no
> test enforcing a certain level of performance for forking off jobs.

To be fair, from looking at old logs, I would say that the forking plus 
handshake
was never significantly faster than 0.3 seconds, at least not when running
in debug mode (where we write 4kB of logs inbetween). It seems that
just our ssh connection submmitting jobs seemed to have been slightly slower
(about 5% or so) in the past.

> Should we introduce one?

We kind-of have one. The job-queue submission performance test submits
lots of short-lived (delay 0.1, I think) jobs. This measures an overall
performance of submission speed, forking throughput, and lock-handover
speed. And yes, it shows that in this combined discipline current Ganeti
is significantly slower than version 2.10; and we've always been slower.

So yes, I think it is worth working on startup speed, and if you have a
good idea for a performance test specifically targeted at forking
speed that would be very good.

Klaus

Reply via email to