-
Notifications
You must be signed in to change notification settings - Fork 47
Highstat Analysis
| hmc0 | All histories look correct. However, the MPI run shows a significantly higher variance than all other runs. |
| hmc1 | Again, all histories look compatible except for the new MPI run which shows a higher variance than all other runs. |
| hmc2 (incomplete) | All histories look compatible. This also applies to the rectangle history. |
| hmc3 (incomplete) | All histories look compatible but it appears that the new MPI run seems to have a slightly greater variance. |
| hmc_cloverdet | The MPI run shows the same problem as before. |
| hmc_tmcloverdet | As above. |
| hmc_tmcloverdetratio | As above. |
| hmc0 | All acceptance rates are 92.6 +- 0.1 except for the new MPI run which is 89.58 . |
| hmc1 | All acceptance rates are 89.5 +- 0.1 |
| hmc2 (incomplete) | no acceptance rate yet |
| hmc3 (incomplete) | no acceptance rate yet |
| hmc_cloverdet | All acceptance rates are 96.5 +- 0.1 |
| hmc_tmcloverdet | All acceptance rates are 96.4 +- 0.1 |
| hmc_tmcloverdetratio | All acceptance rates are 91.55 +- 0.05. The MPI run had an acceptance rate of 90.16... |
| hmc0 | All points except the new MPI run are compatible. It is many sigmas too high. The starting history for the MPI run is much noisier than the others. |
| hmc1 | All serial and openmp points are fully compatible. The 5.1.6_mpi and hybrid run are one sigma too high while the MPI run is many sigmas too high. |
| hmc2 (incomplete) | All points look compatible. |
| hmc3 (incomplete) | The MPI run expectation value is about 50 sigma from the reference value! |
| hmc_cloverdet | Given the coarseness of the reference result all runs are compatible but the hybrid and MPI run are slightly higher than the other two. |
| hmc_tmcloverdet | The MPI run is about 4 sigma above the reference result. The other runs are in principle compatible but the errors overlap only on the edge. |
| hmc_tmcloverdetratio | The MPI run is far too high again. The hybrid run is a little high. |
One feature shared by all suspicious points is that the median lies far above the mean. (several error bars worth)
With the few bugfixes since the last version the code now performs as expected. There is some variation between the different runs but in general it is within errors. For hmc3, the expectation values are above the reference result by one and a half sigma except for the new pure mpi run, which is two and a half sigma away! (perhaps a bug in the functions relating to rectangle gauge terms?)
The clover runs look good except tmcloverdetratio in which the mpi run is still a bit high.
The acceptance rates are now all compatible, as are the autocorrelation times. OpenMP is always a little shorter while MPI and hybrid are always a little longer.
Another run with a very different seed will clear this up.