we have a machine with 2x64-core CPU, each core consists of 2 virtual cores, so in htop we see 256 distinct (virtual) CPUs. We configured Slurm quality of service to better manage CPU usage per user. I.e. we have defined a --qos=cpus50 which, as far as I understand it, gives me a budget of 50 virtual cores to compute my jobs. I created a test.sbatch script with an array of 100 jobs. Each job takes 10s to compute. So with the following config, I would hope that my jobs will be finished in 20s + some small overhead.
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --qos=cpus50
#SBATCH --array=1-100
#SBATCH --cpus-per-task=1
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --ntasks-per-core=1
#SBATCH --open-mode=append
#SBATCH --output=%x.out
python3 -c "import os; jobid = int(os.getenv('SLURM_ARRAY_TASK_ID')); start = f'JOBID:{jobid:04d} | Start {time.ctime()}'; time.sleep(10); print(f'{start} | End {time.ctime()} |')"
However,running the script above spawns only 25 jobs at once (according to squeue output) and finishes in 47seconds. (2x the desired duration). Running with --ntasks-per-core=2 results in the same behavior. Running with --ntasks=2 and --ntasks-per-core=2 results in the same behavior.
What am I doing wrong? I just want to run 50 jobs at once since I already have the virtual cores available. Thank you
Answering my own question. A member of our group found an answer here.
The problem was in Slurm configuration. In short, for our setup, we had to change the relevant part in slurm.conf from
SelectTypeParameters=CR_Core
NodeName=nodename CPUs=256 Sockets=2 CoresPerSocket=64 ThreadsPerCore=2
to
SelectTypeParameters=CR_CPU
NodeName=nodename CPUs=256
Now the sbatch script from the question spawns 50 jobs and finishes in a bit more than 20s as expected.
I am writing you after many attempts I have done on a CPU cluster so structured:
144 standard compute nodes
2× AMD EPYC 7742, 2× 64 cores, 2.25 GHz
256 (16× 16) GB DDR4, 3200 MHz
InfiniBand HDR100 (Connect-X6)
local disk for operating system (1× 240 GB SSD)
1 TB NVMe
Now, since my core-h are here limited, I want to maximize performance as much as I can.
I am doing some benchmarking with the following submission script:
#!/bin/bash -x
#SBATCH --account=XXXX
#SBATCH --ntasks=256
#SBATCH --output=mp-out.%j
#SBATCH --error=mpi-err.%j
#SBATCH --time=24:00:00
#SBATCH --partition=batch
srun ./myprogram
The program I am running si Gromacs2020 (MPI), a Software to perform Molecular Dynamic Simualtions.
In the machine manual I read about these keys:
--ntasks
--ntasks-per-node
--cpu-per-node
However, considering the very recently technology, I am getting mediocre performances. Indeed, in a 5-years older cluster, I get better performance with comparable resources.
So, do you envision a good combination of those keywords to maximize performance and avoid core-h wasting?
My system size is ~100K atoms (if it can help).
Any feedback would be very much appreciated,
Looking forward to hearing from your opinions.
Best Regards
VG
In your case, the 256 tasks have no constraints to run in the same rack, location or not. Slurm have no clues to schedule correctly the job on your cluster. It could be schedule 1 task on 256 different nodes, and that is not efficient at all.
To be sure that all is schedule correctly, perhaps you should force to locate the tasks on the node.
#!/bin/bash -x
#SBATCH --account=XXXX
#SBATCH --nodes=2
#SBATCH --ntasks=256
#SBATCH --ntasks-per-core=1
#SBATCH --tasks-per-node=128
#SBATCH --output=mp-out.%j
#SBATCH --error=mpi-err.%j
#SBATCH --time=24:00:00
#SBATCH --partition=batch
srun ./myprogram
And normally, each 256 tasks will be schedule on 1 core on each AMD socket. and located on 2 nodes. This will avoid oversubscribing and cpu cycles sharing which is inefficient.
To be sure and not be be disturb for benchmarking, ask --exclusive.
Hello to the community:
I recently started to use ORCA software for some quantum calculation but I have been having a lot of problems to lunch a parallel calculation in the cluster of my University.
To install Orca I used the static version:
orca_4_2_1_linux_x86-64_openmpi314.tar.xz.
In a shared direction of the cluster (/data/shared/opt/ORCA/).
And putted in my ~/.bash_profile:
export PATH="/data/shared/opt/ORCA/orca_4_2_1_linux_x86-64_openmpi314:$PATH"
export LD_LIBRARY_PATH="/data/shared/opt/ORCA/orca_4_2_1_linux_x86-64_openmpi314:$LD_LIBRARY_PATH"
For the installation of the corresponding OpenMPI version (3.1.4)
tar -xvf openmpi-3.1.4.tar.gz
cd openmpi-3.1.4
./configure --prefix="/data/shared/opt/ORCA/openmpi314/"
make -j 10
make install
When I use the frontend server all is wonderful:
With a .sh like this:
#! /bin/bash
export PATH="/data/shared/opt/ORCA/openmpi314/bin:$PATH"
export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/data/shared/opt/ORCA/openmpi314/lib"
$(which orca) test.inp > test.out
and an input like this:
# Computation of myjob at b3lyp/6-31+G(d,p)
%pal nprocs 10 end
%maxcore 8192
! RKS B3LYP 6-31+G(d,p)
! TightSCF Grid5 NoFinalGrid
! Opt
! Freq
%cpcm
smd true
SMDsolvent "water"
end
* xyz 0 1
C 0 0 0
O 0 0 1.5
*
The problem appears when I use the nodes:
.inp file:
#! Computation at RKS B3LYP/6-31+G(d,p) for cis1_bh267_m_Cell_152
%pal nprocs 12 end
%maxcore 8192
! RKS B3LYP 6-31+G(d,p)
! TightSCF Grid5 NoFinalGrid
! Opt
! Freq
%cpcm
smd true
SMDsolvent "water"
end
* xyz 0 1
C -4.38728130 0.21799058 0.17853303
C -3.02072869 0.82609890 -0.29733316
F -2.96869122 2.10937041 0.07179384
F -3.01136328 0.87651596 -1.63230798
C -1.82118365 0.05327804 0.23420220
O -2.26240947 -0.92805650 1.01540713
C -0.53557484 0.33394113 -0.05236121
C 0.54692198 -0.46942807 0.50027196
O 0.31128292 -1.43114232 1.22440290
C 1.93990391 -0.12927675 0.16510948
C 2.87355011 -1.15536140 -0.00858832
C 4.18738231 -0.82592189 -0.32880964
C 4.53045856 0.52514329 -0.45102225
N 3.63662927 1.52101319 -0.26705841
C 2.36381718 1.20228695 0.03146190
F -4.51788749 0.24084604 1.49796862
F -4.53935644 -1.04617745 -0.19111502
F -5.43718443 0.87033190 -0.30564680
H -1.46980819 -1.48461498 1.39034280
H -0.26291843 1.15748249 -0.71875720
H 2.57132559 -2.20300864 0.10283592
H 4.93858460 -1.60267627 -0.48060140
H 5.55483009 0.83859415 -0.70271364
H 1.67507560 2.05019549 0.17738396
*
.sh file (Slurm job):
#!/bin/bash
#SBATCH -p deflt #which partition I want
#SBATCH -o cis1_bh267_m_Cell_152_myjob.out #path for the slurm output
#SBATCH -e cis1_bh267_m_Cell_152_myjob.err #path for the slurm error output
#SBATCH -c 12 #number of cpu(logical cores)/task (task is normally an MPI process, default is one and the option to change it is -n)
#SBATCH -t 2-00:00 #how many time I want the resources (this impacts the job priority as well)
#SBATCH --job-name=cis1_bh267_m_Cell_152 #(to recognize your jobs when checking them with "squeue -u USERID")
#SBATCH -N 1 #number of node, usually 1 when no parallelization over nodes
#SBATCH --nice=0 #lowering your priority if >0
#SBATCH --gpus=0 #number of gpu you want
# This block is echoing some SLURM variables
echo "Jobid = $SLURM_JOBID"
echo "Host = $SLURM_JOB_NODELIST"
echo "Jobname = $SLURM_JOB_NAME"
echo "Subcwd = $SLURM_SUBMIT_DIR"
echo "SLURM_CPUS_PER_TASK = $SLURM_CPUS_PER_TASK"
# This block is for the execution of the program
export PATH="/data/shared/opt/ORCA/openmpi314/bin:$PATH"
export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/data/shared/opt/ORCA/openmpi314/lib"
$(which orca) ${SLURM_JOB_NAME}.inp > ${SLURM_JOB_NAME}.log --use-hwthread-cpus
I used the --use-hwthread-cpus flag as a recommendation but the same problem appears with and without this flag.
All the error is:
There are not enough slots available in the system to satisfy the 12 slots that were requested by the application: /data/shared/opt/ORCA/orca_4_2_1_linux_x86-64_openmpi314/orca_gtoint_mpi
Either request fewer slots for your application, or make more slots available for use. A "slot" is the Open MPI term for an allocatable unit where we can launch a process. The number of slots available are defined by the environment in which Open MPI processes are run:
1. Hostfile, via "slots=N" clauses (N defaults to number of processor cores if not provided)
2. The --host command line parameter, via a ":N" suffix on the hostname (N defaults to 1 if not provided)
3. Resource manager (e.g., SLURM, PBS/Torque, LSF, etc.)
4. If none of a hostfile, the --host command line parameter, or an RM is present, Open MPI defaults to the number of processor cores In all the above cases, if you want Open MPI to default to the number
of hardware threads instead of the number of processor cores, use the --use-hwthread-cpus option.
Alternatively, you can use the --oversubscribe option to ignore the number of available slots when deciding the number of processes to launch.
*[file orca_tools/qcmsg.cpp, line 458]:
.... aborting the run*
When I go to the output of the calculation, it looks like start to run but when launch the parallel jobs fail and give:
ORCA finished by error termination in GTOInt
Calling Command: mpirun -np 12 --use-hwthread-cpus /data/shared/opt/ORCA/orca_4_2_1_linux_x86-64_openmpi314/orca_gtoint_mpi cis1_bh267_m_Cell_448.int.tmp cis1_bh267_m_Cell_448
[file orca_tools/qcmsg.cpp, line 458]:
.... aborting the run
We have two kind of nodes on the cluster:
A punch of them are:
Xeon 6-core E-2136 # 3.30GHz (12 logical cores) and Nvidia GTX 1070Ti
And the other ones:
AMD Epyc 24-core (24 logical cores) and 4x Nvidia RTX 2080Ti
Using the command scontrol show node the details of one node of each group are:
First Group:
NodeName=fang1 Arch=x86_64 CoresPerSocket=6
CPUAlloc=12 CPUTot=12 CPULoad=12.00
AvailableFeatures=(null)
ActiveFeatures=(null)
Gres=gpu:gtx1070ti:1
NodeAddr=fang1 NodeHostName=fang1 Version=19.05.5
OS=Linux 5.7.12-arch1-1 #1 SMP PREEMPT Fri, 31 Jul 2020 17:38:22 +0000
RealMemory=15923 AllocMem=0 FreeMem=171 Sockets=1 Boards=1
State=ALLOCATED ThreadsPerCore=2 TmpDisk=7961 Weight=1 Owner=N/A MCS_label=N/A
Partitions=deflt,debug,long
BootTime=2020-10-27T09:56:18 SlurmdStartTime=2020-10-27T15:33:51
CfgTRES=cpu=12,mem=15923M,billing=12,gres/gpu=1,gres/gpu:gtx1070ti=1
AllocTRES=cpu=12,gres/gpu=1,gres/gpu:gtx1070ti=1
CapWatts=n/a
CurrentWatts=0 AveWatts=0
ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s
Second Group
NodeName=fang50 Arch=x86_64 CoresPerSocket=24
CPUAlloc=48 CPUTot=48 CPULoad=48.00
AvailableFeatures=(null)
ActiveFeatures=(null)
Gres=gpu:rtx2080ti:4
NodeAddr=fang50 NodeHostName=fang50 Version=19.05.5
OS=Linux 5.7.12-arch1-1 #1 SMP PREEMPT Fri, 31 Jul 2020 17:38:22 +0000
RealMemory=64245 AllocMem=0 FreeMem=807 Sockets=1 Boards=1
State=ALLOCATED ThreadsPerCore=2 TmpDisk=32122 Weight=1 Owner=N/A MCS_label=N/A
Partitions=deflt,long
BootTime=2020-12-15T10:09:43 SlurmdStartTime=2020-12-15T10:14:17
CfgTRES=cpu=48,mem=64245M,billing=48,gres/gpu=4,gres/gpu:rtx2080ti=4
AllocTRES=cpu=48,gres/gpu=4,gres/gpu:rtx2080ti=4
CapWatts=n/a
CurrentWatts=0 AveWatts=0
ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s
I use in the script of Slurm the flag -c, --cpus-per-task = integer; and in the input for Orca the command %pal nprocs integer end. I tested different combinations of this two parameters in order to see if I am using more CPU than the available:
-c, --cpus-per-task = integer
%pal nprocs integer end
None
6
None
3
None
2
1
2
1
12
2
6
3
4
12
12
With different amount of memories: 8000 MBi and 2000 MBi (my total memory is around 15 GBi). And in all the cases the same error appears. I am not an expert user neither in ORCA non in informatic (but maybe you guess this for the extension of the question), so maybe the solution is simple but I really don’t have it, Idon't know what's going on!
A lot of thanks in advance,
Alejandro.
Faced the same issue.
Explicit declaration --prefix ${OMPI_HOME} directly as ORCA parameter and using of static linked ORCA version helps me:
export RSH_COMMAND="/usr/bin/ssh"
export PARAMS="--mca routed direct --oversubscribe -machinefile ${HOSTS_FILE} --prefix ${OMPI_HOME}"
$ORCA_DIR/orca $WORKDIR/$JOBFILE.inp "$PARAMS" > $WORKDIR/$JOBFILE.out
Also, It's better to build OpenMPI 3.1.x with --disable-builtin-atomics flag.
Thank you #Alexey for your answer. And sorry for the wrong Tag, like I said, I am pretty rookie on this stuff.
The problem was not in the Orca or OpenMPI configuration but in the bash script used for scheduled the Slurm job.
I thought that the entire Orca job itself was what Slurm call a "task". For that reason I declared the flag --cpus-per-task equal to the number of parallel jobs that I want to do with Orca. But the problem is that each parallel Orca job (that is launch using OpenMPI) is a task for Slurm. Therefore with my Slurm script I was reserving a node with at least 12 CPU, but when Orca launch their parallel jobs, each one ask for 12 CPU, so: "There are not enough slots available ..." because I needed 144 CPU.
The rest of the cases in the table of my Question fails for another reason. I was launching at the same time 5 different Orca calculation. Now, because --cpus-per-task could be None, 1, 2 or 3; the five calculation might enter in the same node or in another node with this amount of free CPU, but when Orca ask for the parallel jobs, fail again because there are not this amount of CPU on the node.
The solution that I found is pretty simple. On the .sh script for Slurm I putted this:
#SBATCH --mincpus=n*m
#SBATCH --ntasks=n
#SBATCH --cpus-per-task m
Instead of only:
#SBATCH --cpus-per-task m
Where n will be equal to the number of parallel jobs specified on the Orca input (%pal nprocs n end) and m the number of CPU that you want to use for each parallel Orca job.
In my case I used n = 12, m = 1. With the flag --mincpus I ensured to take a node with at least 12 CPU and allocated them. With the --cpus-per-task is pretty evident what this flag do (even for me :-) ), which, by the way, has a default value of 1 and I don't know if more than 1 CPU for each OpenMPI Orca job improve the velocity of the calculation. And --ntasks gives the information to Slurm of how many task you will do.
Of course if you know the number of task and the CPU per task is easy to know how many CPU you need to reserve, but I don't know if this is easy to Slurm too :-). So, to be sure that I allocate the correct number of CPU i used --mincpus flag, but maybe is not needed. The thing is that it works now ^_^.
It is also important to take into account the amount of memory that you declare in the input of Orca in order of do not exceed the available memory. For example, if you have 12 task and a RAM of 15000 MBi, the right amount of memory to declared should be no more than 15000/12 = 1250 MBi
I had a similar problem with parallel jobs before. The slurm also output not enough slots error.
My solution is to change parallel threads into parallel processes. For my system is to change
#SBATCH -c 24
into
#SBATCH -n 24
and everything works just fine.
Hello everyone I'm actually using a soft called RepeatMasker, in this pipeline I can run parallelized job via slurm with the command -pa
here is a doc about this command :
RepeatMasker -h
-pa(rallel) [number]
The number of sequence batch jobs [50kb minimum] to run in parallel.
RepeatMasker will fork off this number of parallel jobs, each
running the search engine specified. For each search engine
invocation ( where applicable ) a fixed the number of cores/threads
is used:
RMBlast 4 cores
To estimate the number of cores a RepeatMasker run will use simply
multiply the -pa value by the number of cores the particular search
engine will use.
so in a slurm batch script I should add :
#SBATCH --cpus-per-task=8
RepeatMakser -pa 2, right?
since 8/4 =2
But I wondered if I should also add others #SBATCH parameters or if --cpus-per-task is sufficient ?
Thanks al ot
We have a 4 GPU nodes with 2 36-core CPUs and 200 GB of RAM available at our local cluster. When I'm trying to submit a job with the follwoing configuration:
#SBATCH --nodes=1
#SBATCH --ntasks=40
#SBATCH --cpus-per-task=1
#SBATCH --mem-per-cpu=1500MB
#SBATCH --gres=gpu:4
#SBATCH --time=0-10:00:00
I'm getting the following error:
sbatch: error: Batch job submission failed: Requested node configuration is not available
What might be the reason for this error? The nodes have exactly the kind of hardware that I need...
The CPUs are most likely 36-threads not 36-cores and Slurm is probably configured to allocate cores and not threads.
Check the output of scontrol show nodes to see what the nodes really offer.
You're requesting 40 tasks on nodes with 36 CPUs. The default SLURM configuration binds tasks to cores, so reducing the tasks to 36 or fewer may work. (Or increases nodes to 2, if your application can handle that)