I encountered an abnormal slowdown when restarting WRF from a wrfrst file.This is a single-domain, one-way nested WRF simulation using ndown, with 100 m resolution and adaptive time stepping.
In the continuous run, after the adaptive timestep reaches dt = 1.20 s, the model runs normally:
“Timing for main (dt= 1.20): time 2026-02-18_00:00:22 on domain 1: 1.29544 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-18_00:00:23 on domain 1: 1.26584 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-18_00:00:24 on domain 1: 1.25132 elapsed seconds”
The output interval is 30 minutes, and the wall-clock time is also about 30 minutes per 30-minute simulation output.
However, after stopping the run and restarting from:wrfrst_d01_2026-02-20_12:00:00,the model becomes much slower, although the timestep is still dt = 1.20 s:
"Timing for main (dt= 1.20): time 2026-02-20_12:00:01 on domain 1: 6.13815 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-20_12:00:02 on domain 1: 5.52845 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-20_12:00:03 on domain 1: 5.57714 elapsed seconds"
So for the same dt = 1.20 s, the continuous run takes about 1.2–1.3 s per step, while the restart run takes about 5.5–6.1 s per step. After restart, it takes nearly 2 hours of wall-clock time to produce 30 minutes of simulation output.
I have checked that:
1. The number of MPI processes is the same.
2. The namelist settings are the same.
3. The timestep is the same after adaptive time stepping reaches dt = 1.20 s.
4. The previous WRF job was stopped before restarting.
5. Similar behavior has also been observed in other simulations and on other servers.
The restart file appears to be read correctly:
LBC for restart: Found the correct bounding LBC time periods for restart time = 2026-02-20_12:00:00.
Any suggestions would be appreciated.
In the continuous run, after the adaptive timestep reaches dt = 1.20 s, the model runs normally:
“Timing for main (dt= 1.20): time 2026-02-18_00:00:22 on domain 1: 1.29544 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-18_00:00:23 on domain 1: 1.26584 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-18_00:00:24 on domain 1: 1.25132 elapsed seconds”
The output interval is 30 minutes, and the wall-clock time is also about 30 minutes per 30-minute simulation output.
However, after stopping the run and restarting from:wrfrst_d01_2026-02-20_12:00:00,the model becomes much slower, although the timestep is still dt = 1.20 s:
"Timing for main (dt= 1.20): time 2026-02-20_12:00:01 on domain 1: 6.13815 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-20_12:00:02 on domain 1: 5.52845 elapsed seconds
Timing for main (dt= 1.20): time 2026-02-20_12:00:03 on domain 1: 5.57714 elapsed seconds"
So for the same dt = 1.20 s, the continuous run takes about 1.2–1.3 s per step, while the restart run takes about 5.5–6.1 s per step. After restart, it takes nearly 2 hours of wall-clock time to produce 30 minutes of simulation output.
I have checked that:
1. The number of MPI processes is the same.
2. The namelist settings are the same.
3. The timestep is the same after adaptive time stepping reaches dt = 1.20 s.
4. The previous WRF job was stopped before restarting.
5. Similar behavior has also been observed in other simulations and on other servers.
The restart file appears to be read correctly:
LBC for restart: Found the correct bounding LBC time periods for restart time = 2026-02-20_12:00:00.
Any suggestions would be appreciated.