Closed kostrzewa closed 11 years ago
By the way: why are the results for this "relative squared accuracy" so dependent on the seed? By changing the seed (for the hmc2 sample input file) I can make the accuracies vary from 10e5 to 10e-12 ... it seems a bit random and completely unexpected... I would think that the eigenvalue spectrum would be similar for two different "hot" configurations
can you start from a configuration from that ensemble? A Hot start is really the worst starting point for the polynomial: you'll likely have very small eigenvalues and they'll fluctuate a lot.
the input file looks fine, as far as I can see.
That's what I figured to be the problem, I just wasn't sure I had translated the timescales and integrator parameters correctly. I don't have grid access yet but I will ask someone who does for a configuration. I do have access to a few confs from the ensemble but they have the wrong hashes...
Might be one of the ensembles where all hashes are wrong due to a bug in the hmc version used for that run. You can still use it, if you check the plaquette value to be correct. And you need to se
DisableIOChecks = yes
in your input file...
that's solved by now, isn't it?
Yes, absolutely, let's close this.
Today I've begun looking at the HMC using the NDPOLY monomial. In particular, I've attempted to convert an input file for A40.24 to the new format.
The old input file: https://gist.github.com/3776270 The new input file: https://gist.github.com/3776251
During the computation of P and Ptilde the "relative squared accuracy in components" blows up both on Intel (pure MPI) and BGQ (hybrid run). As a result the heatbath for the polynomial generates NaNs.
Now, NDPOLY does seem to work fine because the hmc2 sample file runs well as long as the seed is picked "correctly", but maybe there's something subtle that broke in the last few months?
Could someone please:
Here's the output that I see from the initialisation: