Inconsistency in the slater_five_band example

[henhans]

Dear TRIQS team,

Description

I am testing the Slater five-band example on the website, https://triqs.github.io/cthyb/latest/guide/slater_five_band.html. On the website, it says:

"We have rebinned the data to 1000 points to reduce the noise. The calculation takes about an hour, and data was accumulated on 32 cores."

However, I tested the script using 32, 64, and 128 cores, but I got very noisy data. I have to use 256 cores to reproduce the figure shown in the example. I wonder if it is normal or if I have done something wrong. The calculation also shows the following warnings:

WARNING: Tail fit in extraction of delta(infty) has large error of: 4.04355e-05
WARNING: Direct Fourier cannot deduce the high-frequency moments of G(tau) due to noise or a coarse tau-grid.     Please specify the high-frequency moments for higher accuracy.
WARNING: Tau discontinuity of G_tau deviates appreciably from -1
     .... max_element |g(0) + g(beta) + 1| = 0.0122623

Please see the figures below for 32, 128, and 256 cores.

32 cores:

128 cores:

256 cores:

Steps to Reproduce

1. I used the same script, "slater_five_band.py," on the website and the up-to-date TRIQS and CTHYB on GitHub.
2. Tested the code with 32 - 256 cores.

Expected behavior:
Expected results with 32 cores:

Actual behavior:
32 cores:

128 cores:

256 cores:

Versions

You are using triqs_cthyb version 3.1.0
You are using triqs_cthyb git hash f86a78b based on triqs git hash 1da1cb4ed217dea552a9b4147e4ecdc67d0fd94c

[henhans]

Is it possible that the parameters p["n_warmup_cycles"] = 10000 and p["n_cycles"] = 100000 in the example script are too small to get good statistics using 32 cores? I hope that I didn't mess up something in my compilation, but the results do become better if I increase the number of cycles or the number of cores.

[Wentzell]

@henhans,

Thank you for pointing this out! We can confirm using the versions that you specify that the compute resources specified in the example do not suffice to generate the plot shown. We are looking into this and will get back to you.