TB's mochitest is incredibly CPU-intensive and slow. TWICE AS SLOW AS |make mozmill| (!)

[ISHIKAWA,chiaki]

Hi,

Thunderbird |make mozmill| test suite is gone.

Instead mochitest is here to stay.

From a short experience for the last month or so, I have an issue,
though.

mochitest for TB is incredibly time consuming and SLOW under linux.

The whole run takes almost TWICE AS LONG as |make mozmill| on my local
PC.

This hurts developer's productivity.
It hurts LOCAL developer and presumably tryserver workload, too.

To wit, elapsed time in minutes.

For mochitest:
samples = 10
avg = 141.800000 +/- sigma = 24.330228

For mozmill:
samples = 52
avg = 69.673077 +/- sigma = 18.389714

(These numbers were obtained by rounding off the time reported as
"real" by time command for each whole test run.)

I am running locally created FULL DEBUG version of TB.

Now, I think each test is reasonably short.

*HOWEVER*, looking at the window where the mochitest draws TB GUI,
I learned that  mochitest harness spends a surprisingly long time each
time a test is finished.
Nothing happens for a quite a long time after the window goes blank.

Also, it spends long time after printing

 0:40.76 INFO checking window state

(This happens when the test seems to finish, but somehow the screen
was left as it was at the end of a test.)

I am afraid this slowdown is killing developer productivity and maybe
places undue workload on tryserver CPU farm.
(Well, at least, I can no longer reasonably expect how my local change
affects the outcome of whole test before going to bed. I could do it
when |make suite| finished in about an hour.)

I have filed a bugzilla:
Bug 1635009 TB's mochitest is incredibly slow
(There you can find the samples of real time usage for mochitest and
|make mozmill| from which the average times were derived.)

I would like to hear other people's experience.
Does anyone else notice this slowdown especially on local PC?

Any ideas to speed things up?
I am a bit puzzled why it uses so much CPU time, even.


TIA

Chiaki