Quick User Guide
1. Hardware List
System Nodes Processor Cores per Threads Memory GPU GPU
Partition Socket /Core /Node Number Number
/Node
cpu6348 cpu6348n 2*Intel(R)Xeon( 28 - 512 - -
[1-2] R) Gold 6348
cpuhuge1t cpuhuge1t 4*Intel(R)Xeon( 24 - 1024 - -
R) Gold 6348H
gpu3090 gpu3090n 2*Intel(R)Xeon( 24 2 256 RTX3090 8
[1-6] R) Gold 6336Y
gpua800 gpua800n 2*Intel(R)Xeon( 24 2 512 A800 4
[1-2] R) Gold 6336Y
2. Connect and data transfer
SSH login:
ssh <username>@[Link]
Data Transfer Node:
ssh <username>@[Link]
Please note that the capacity limit of the /gpfs/home personal directory is 10GB. It is
recommended to use the /gpfs/work/personal directory for storage and save data in a timely
manner. For example: /gpfs/work/mits/sanzhang.
3. Software list
The cluster uses the environment module to load software (Lmod official document). Please
do not submit computing jobs at the login node. Computing jobs should be submitted to the
computing nodes).
Compilers, MPI libraries, math libraries, application tools
Software version
cudnn [Link]-11.8, [Link]-11.5, [Link]-11.6
gcc 8.5.0(default), 10.4.0, 11.4.0, 12.2.0, 13.1.0
g++ 8.5.0(default), 10.4.0, 11.4.0, 12.2.0, 13.1.0
gfortran 8.5.0(default), 10.4.0, 11.4.0, 12.2.0, 13.1.0
intel-mkl 2020.4.304
intel-oneapi-mkl 2021.4.0, 2022.2.1, 2023.1.0
intel-oneapi-compilers 2021.4.0, 2022.2.1, 2023.1.0
cmake 3.26.3
openmpi 4.1.4
CUDA 11.7(default), 11.8
python 3.7.15, 3.8.16, 3.9.16, 3.10.10
anaconda3 2022.10
Please view detailed installed software：
ml av
4. To submit job
System uses Slurm(Simple Linux Utility for Resource Management) for cluster/resource
management and job scheduling. Slurm is responsible for allocating resources to users,
providing a framework for starting, executing and monitoring work on allocated resources
and scheduling work for future execution.
Slurm official document
Slurm Function
sinfo Cluster status
squeue Queued job status
sbatch Submit job
scontrol Viewing and modifying job parameters
scancel Cancel job
GPU
Max
PartitionName QOS MaxTRESPU Max time (hrs)
nodes
gpu3090 gpu3090 6 4 168
gpua800 gpua800 2 2 168
CPU
Max
Max MaxTRES MaxJobs MaxSubmit
PartitionName QOS time
nodes PU PU PU
(hrs)
cpu6348
cpu6348 2 26 26 52 168
(default)
cpuhuge1t cpuhuge1t 1 46 46 92 168
Note:
MaxTRES
Maximum number of TRES each user is able to use.
PU
4.1 Submit job process:
1) View cluster resources - sinfo
2) Job script - vim [Link]
3) Submit job - sbatch [Link]
4) View job - squeue
4.2 Sample job script:
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --partition=cpu6348
#SBATCH --qos cpu6348
#SBATCH -n 8
#SBATCH --output=%[Link]
#SBATCH --error=%[Link]
Option Usage
Number of nodes -N or --nodes
Number of tasks -n or --ntasks
Number of processes (tasks) per node --ntasks-per-node
Total memory per node in megabytes --mem=
Memory per core in megabytes --mem-per-cpu
Wallclock time -t d-hh:mm:ss or --time=d-hh:mm:ss
Partition (queue) requested -p or --partition
QOS --qos
"--" No space between two minus signs.