Re: [slurm-users] srun and mpirun
On Saturday, 14 April 2018 1:33:13 AM AEST Mahmood Naderan wrote: > I tried with one of the NAS benchmarks (BT) with 121 threads since the > number of cores should be square. That's an IO benchmark, not going to help you for this. You need something that is compute bound & comms intensive to see differences. It also sounds like you've got a problem with either your Slurm job or Slurm itself from the error you posted. -- Chris Samuel : http://www.csamuel.org/ : Melbourne, VIC
Re: [slurm-users] srun and mpirun
The output is certainly not enough to judge, but my first guess would be that your MPI (what is it btw?) is not support PMI that is enabled in Slurm. Note also, that Slurm now supports 3 ways of doing PMI and from the info that you have provided it is not clear which one you are using. To judge with a reasonable level of confidence the following info is needed: * Which MPI implementation is used. * What version. * How it was configured. If you do not provide "--mpi" option to srun you are using PMI1 which is implemented in Slurm core library. There are other implementations available: * PMI2 plugin ("srun --mpi=pmi2") * PMIx plugin ("srun --mpi=pmix") Those are more performant options but need some additional work: for PMI2 plugin you need to install the library from contrib/pmi2, for PMIx you need to build slurm --with-pmix=. 2018-04-13 8:33 GMT-07:00 Mahmood Naderan : > I tried with one of the NAS benchmarks (BT) with 121 threads since the > number of cores should be square. With srun, I get > > WARNING: compiled for 121 processes > Number of active processes: 1 > >0 1 408 408 408 > Problem size too big for compiled array sizes > > > > > However, with mpirun, it seems to be fine > > Number of active processes: 121 > > Time step1 > Time step 20 > > > Regards, > Mahmood > > > > > On Fri, Apr 13, 2018 at 5:46 PM, Chris Samuel wrote: > > On 13/4/18 7:19 pm, Mahmood Naderan wrote: > > > >> I see some old posts on the web about performance comparison of srun vs. > >> mpirun. Is that still an issue? > > > > > > Just running an MPI hello world program is not going to test that. > > > > You need to run an actual application that is doing a lot of > > computation and communications instead. Something like NAMD > > or a synthetic benchmark like HPL. > > > > All the best, > > Chris > > -- > > Chris Samuel : http://www.csamuel.org/ : Melbourne, VIC > > > > -- С Уважением, Поляков Артем Юрьевич Best regards, Artem Y. Polyakov
Re: [slurm-users] srun and mpirun
I tried with one of the NAS benchmarks (BT) with 121 threads since the number of cores should be square. With srun, I get WARNING: compiled for 121 processes Number of active processes: 1 0 1 408 408 408 Problem size too big for compiled array sizes However, with mpirun, it seems to be fine Number of active processes: 121 Time step1 Time step 20 Regards, Mahmood On Fri, Apr 13, 2018 at 5:46 PM, Chris Samuel wrote: > On 13/4/18 7:19 pm, Mahmood Naderan wrote: > >> I see some old posts on the web about performance comparison of srun vs. >> mpirun. Is that still an issue? > > > Just running an MPI hello world program is not going to test that. > > You need to run an actual application that is doing a lot of > computation and communications instead. Something like NAMD > or a synthetic benchmark like HPL. > > All the best, > Chris > -- > Chris Samuel : http://www.csamuel.org/ : Melbourne, VIC >
Re: [slurm-users] srun and mpirun
On 13/4/18 7:19 pm, Mahmood Naderan wrote: I see some old posts on the web about performance comparison of srun vs. mpirun. Is that still an issue? Just running an MPI hello world program is not going to test that. You need to run an actual application that is doing a lot of computation and communications instead. Something like NAMD or a synthetic benchmark like HPL. All the best, Chris -- Chris Samuel : http://www.csamuel.org/ : Melbourne, VIC
Re: [slurm-users] srun and mpirun
On Fri, 13 Apr 2018 13:49:56 +0430 Mahmood Naderan wrote: > Hi, > I see some old posts on the web about performance comparison of srun > vs. mpirun. Is that still an issue? Both the following scripts works > for test programs and surely the performance concerns is not visible > here. ... > #SBATCH --time=10:00 > mpirun mpihello ... > #SBATCH --time=10:00 > srun mpihello The difference will depend on the MPI implementation. For example for some MPIs the mpirun will run ssh (which may end up placing processes really badly wrt job cgroups etc) while others may auto-detect slurm and use srun to spawn ranks or proxies. Then there's very different performance aspects that can be affected (or at least work differently). For example launch and setup is one thing but process pinning behavior is another... It's not a very simple question in the general case. /Peter K
[slurm-users] srun and mpirun
Hi, I see some old posts on the web about performance comparison of srun vs. mpirun. Is that still an issue? Both the following scripts works for test programs and surely the performance concerns is not visible here. #!/bin/bash #SBATCH --job-name=hello_mpi #SBATCH --output=hellompi.log #SBATCH --ntasks=120 #SBATCH --time=10:00 mpirun mpihello and #!/bin/bash #SBATCH --job-name=hello_mpi #SBATCH --output=hellompi.log #SBATCH --ntasks=120 #SBATCH --time=10:00 srun mpihello Regards, Mahmood