Can Raxml estimate how much time it needs to complete the task and inform users?

[lowie li]

Well.. I hate waiting in front of computer and wonder how many hours
left to obtain the results, and hope it's less than a century :p .
Could anyone tell me whether raxml provides the estimated time to
finish the job?
Thanks a lot !!

[JK]

Well, it is hard to tell for all the different kinds of analyses RAxML
can do, but from my experience, if you use rapid boostraping + ML
search (option '-f a') the bootstrapping phase (with 100 replicates)
takes something about 40-45% of the computation time with aminoacid
data, and 30-35% with nucleotide data (with StdDev of ~5%). Knowing
this, you might be able to make some rough estimates based off the
time that it takes to do a single bootstrap run.

Best,
-JK

[A-A_lowie li]

Thanks for your insight :)
Unfortunately, I only run a single run (without any bootstrap), and it has already taken 12 hours, and I have no idea about how much time it will take more :(
It would be nice that RaxML can make some very rough estimation of how much time it needs, and update them after a while.
Otherwise, I really don't know when could I have the results, or should I stop it after several days :)

Lowie

--
Without dream, life is incomplete!

[Jed]

Lowie-
RAxML does output a status report containing the time and log
likelihood values. You can guesstimate how long (order-of-magnitude)
convergence might take based on the sampling time and change in log
likelihood. For example, if you see this:
--snip--
602610.601201 -488387.444275
603517.544427 -488382.764550
604433.943700 -488380.917121
--snip--
you know you've still got a while to go (in this case another 13 hours
on an 180 hour run).

Jed

[Alexis]

Dear All,

In general, there is no way to predict the run-time of RAxML because
of the very nature of the search algorithm.
That is, a priori we do not know how many rounds of SPR moves will be
required for it to terminate. This seems
to be very much dataset dependent as well, i.e.. one dataset with 1000
taxa and 1000 sites may take much longer to converge
than another dataset with 1000 taxa and 1000 sites.

You can just try to derive this stuff empirically, i.e., as you add
more sites to your alignment the run time will increase linearly with
the number of sites.

Apart from that, the run time behavior is hard to predict as a
function of the number of taxa.

You'd probably need to just gather some benchmark data to maek
predictions.

Alexis