Ubuntu 16.04下mpirun无法调用全部4核CPU的技术求助

Hey there, let's work through this issue where your OpenMPI 3.0.0 setup isn't utilizing all 4 logical cores of your i3-6100U. You mentioned some commands fail while others work, and you've confirmed your system has 4 processors—let's get this sorted without oversubscribing.

1. First, Verify MPI Can See All 4 Cores #

Start by testing if MPI can actually spawn 4 processes on your machine. Run this simple command:

mpirun --np 4 --host localhost hostname

If this outputs your machine's name 4 times, MPI is capable of using all cores, and the issue is likely with how you're launching your program. If it errors out, we'll dig into configuration issues next.

2. Adjust Process Binding Settings #

OpenMPI's default process binding might be restricting core usage. Try either of these approaches to ensure processes spread across all cores:

• Disable binding entirely (quick test):

  mpirun --np 4 --bind-to none ./your_mpi_program
  

• Explicitly bind each process to a separate core (more reliable for performance):

  mpirun --np 4 --bind-to core ./your_mpi_program
  

3. Check Hostfile Configuration #

If OpenMPI is using a custom hostfile, it might be limiting the number of allowed slots (cores):

• Find where your default hostfile lives with:

  ompi_info | grep "hostfile"
  

• If you have a custom hostfile, make sure it lists localhost slots=4 (no extra limits). If you don't have one, you can explicitly specify the slot count when launching:

  mpirun --host localhost:4 ./your_mpi_program
  

4. Troubleshoot the Failing Command #

You noted some commands fail while others work—if you can share the exact error message, that helps narrow things down, but here are common fixes:

• If you're getting an oversubscription error, double-check you're not specifying more processes than cores (stick to -np 4).
• If the error relates to core binding, consider updating OpenMPI to 3.1.x (Ubuntu 16.04 allows source compilation or using a third-party PPA) since 3.0.0 had a few known binding bugs.

5. Confirm System & MPI Core Count Match #

Double-check that both your system and MPI agree on the number of available cores:

nproc  # This should return 4
ompi_info --parsable | grep mpi_built_with_thread_support:value  # Ensure this says "yes"

Quick Fix to Start Using All Cores #

If you want a straightforward solution right now, launch your program with these flags to explicitly use 4 cores and proper binding:

mpirun --np 4 --bind-to core ./your_mpi_program

内容的提问来源于stack exchange，提问作者James Smith