A general purpose HPC Template and Experiment management system
Hyper Performance Computation Clusters (HPCs) are designed around a timesharing principle and are powered by queue-based execution ecosystems such as SchedMD's SLURM and IBM's Platform Load Sharing Facility (LSF). While these ecosystems provide a great deal of control and extension for planning, scheduling, and batching jobs, they are limited in their ability to support parameterization in a scheduled task. While there are facilities in place to execute jobs on an Array, the ability to do permutation based experments are limited to what you integrate into your own batch script. Even then, parameterization of values are only made availabile as environment variables, which can be limited depending on your OS or selected programming language. In many cases limitations set by the deployment trhough the compute center also hinder optimal use while restrictions are placed on duration and number of parallel accessible resources. In some cases these restrictions are soo established that removing them is impractical and takes weks to implement on temporary basis.
Cloudmesh Experiment Executor (ee) is a framework that wraps the SLURM batch processor into a templated framework such that experiments can be generated based on configuration files focusing on the livecycle of generating many permutations of experiments with standard tooling, so that you can focus more on modeling your experiments than how to orchestrate them with tools. A number of batch scripts can be generated that than can be executed according to center policies.
When you install cloudmesh-ee, you will also be installing a
minimum baseline of the cms command (as part of the Cloudmesh
ecosystem). For more details on Cloudmesh, see its documentation on
read the docs. However
all instalation can be done thorugh pip. After instalation, you will
need to initialize cloudmesh with the command
$ cms helpWhile SLURM is not needed to run the cloudmesh ee command, the
generated output will not execute unless your system has slurm installed
and you are able to run jobs via the slurm sbatch command.
The cloudmesh ee command takes one of two forms of execution. It is started with
$ cms ee <command> <parameters>Where the command invokes a partiuclar action and parameters include a number of parameters for the command These commands allow you to inspect the generated output to confirm your parameterization functions as expected and as intended.
In general, configuration arguments that appear in multiple locations are prioritized in the following order (highest priority first)
- CLI Arguments with
cms ee - Configuration Files
- Preset values
The generate command is used to generate your experiments based upon either a passed
configuration file, or via CLI arguments. You can issue the command using
either of the below forms:
cms ee generate SOURCE --name=NAME [--verbose] [--mode=MODE] [--config=CONFIG] [--attributes=PARAMS] [--out=DESTINATION] [--dryrun] [--noos] [--nocm] [--dir=DIR] [--experiment=EXPERIMENT]
cms ee generate --setup=FILE [SOURCE] [--verbose] [--mode=MODE] [--config=CONFIG] [--attributes=PARAMS] [--out=DESTINATION] [--dryrun] [--noos] [--nocm] [--dir=DIR] [--experiment=EXPERIMENT] [--name=NAME]
If you have prepared a configuration file that conforms to the schema defined in Setup Config, then you can use the second form which overrides the default values.
-
--name=NAME- Supplies a name for this experiment. Note that the name must not match any existing files or directories where you are currently executing the command -
--verbose- Enables additional logging useful when troubleshooting the program. -
--mode=MODE- specifies how the output should be generated. One of: f,h,d.forflat- specifies a "flat" mode, where slurm scripts are generated in a flattened structure, all in one directory.horhierarchical- specifies a "hierarchical" mode, where experiments are nested into unique directories from each other.dordebug- instructs the command to not generate any output.
-
--config=CONFIG- specifies key-value pairs to be used across all files for substitution. This can be a python, yaml, or json file. -
--attributes=PARAMS- specifies key-value pairs that can be listed at the command line and used as substitution across all experiments. Note this command leverages cloudmesh's parameter expansion specification for different types of expansion rules. -
--out=DESTINATION- specifies the directory to write the generated scripts out to. -
--dryrun- Runs the command without performing any operations -
--noos- Prevents the interleaving of OS environemnt variables into the subsitution logic -
--dir=DIR- specifies the directory to write the generated scripts out to. -
--experiment=EXPERIMENT- specifies a listing of key-value parameters that establish a unique experiment for each combination of values (a cartisian product across all values for each key). -
--setup=FILE- provides all the above configuration options within a configuration file to simplify executions.
ee generate submit --name=NAME [--verbose]
This command uses the output of the generate command and generates a shell script that can be used to submit your previously generated outputs to SLURM as a sequence of sbatch commands.
--name=NAME- specifies the name used in the generate command. The generate command will inspect the<NAME>.jsonfile and build the necessary commands to run all permutations that the cloudmesh ee command generated.
Note that this command only generates the script, and you must run the outputted file in your shell for the commands to be issued to SLURM and run your jobs.
Sample YAML File
This command requires a YAML file which is configured for the host and gpu. The YAML file also points to the desired slurm template.
slurm_template: 'slurm_template.slurm'
ee_setup:
<hostname>-<gpu>:
- card_name: "a100"
- time: "05:00:00"
- num_cpus: 6
- num_gpus: 1
rivanna-v100:
- card_name: "v100"
- time: "06:00:00"
- num_cpus: 6
- num_gpus: 1
example:
cms ee slurm.in.sh --config=a.py,b.json,c.yaml --attributes=a=1,b=4 --noos --dir=example --experiment=\"epoch=[1-3] x=[1,4] y=[10,11]\"
ee slurm.in.sh --config=a.py,b.json,c.yaml --attributes=a=1,b=4 --noos --dir=example --experiment="epoch=[1-3] x=[1,4] y=[10,11]"
# ERROR: Importing python not yet implemented
epoch=1 x=1 y=10 sbatch example/slurm.sh
epoch=1 x=1 y=11 sbatch example/slurm.sh
epoch=1 x=4 y=10 sbatch example/slurm.sh
epoch=1 x=4 y=11 sbatch example/slurm.sh
epoch=2 x=1 y=10 sbatch example/slurm.sh
epoch=2 x=1 y=11 sbatch example/slurm.sh
epoch=2 x=4 y=10 sbatch example/slurm.sh
epoch=2 x=4 y=11 sbatch example/slurm.sh
epoch=3 x=1 y=10 sbatch example/slurm.sh
epoch=3 x=1 y=11 sbatch example/slurm.sh
epoch=3 x=4 y=10 sbatch example/slurm.sh
epoch=3 x=4 y=11 sbatch example/slurm.sh
Timer: 0.0022s Load: 0.0013s ee slurm.in.sh --config=a.py,b.json,c.yaml --attributes=a=1,b=4 --noos --dir=example --experiment="epoch=[1-3] x=[1,4] y=[10,11]"
32 Processors (threads)
sudo apt update -y
sudo apt install slurmd slurmctld -y
sudo chmod 777 /etc/slurm-llnl
# make sure to use the HOSTNAME
sudo cat << EOF > /etc/slurm-llnl/slurm.conf
# slurm.conf file generated by configurator.html.
# Put this file on all nodes of your cluster.
# See the slurm.conf man page for more information.
#
ClusterName=localcluster
SlurmctldHost=$HOSTNAME
MpiDefault=none
ProctrackType=proctrack/linuxproc
ReturnToService=2
SlurmctldPidFile=/var/run/slurmctld.pid
SlurmctldPort=6817
SlurmdPidFile=/var/run/slurmd.pid
SlurmdPort=6818
SlurmdSpoolDir=/var/lib/slurm-llnl/slurmd
SlurmUser=slurm
StateSaveLocation=/var/lib/slurm-llnl/slurmctld
SwitchType=switch/none
TaskPlugin=task/none
#
# TIMERS
InactiveLimit=0
KillWait=30
MinJobAge=300
SlurmctldTimeout=120
SlurmdTimeout=300
Waittime=0
# SCHEDULING
SchedulerType=sched/backfill
SelectType=select/cons_tres
SelectTypeParameters=CR_Core
#
#AccountingStoragePort=
AccountingStorageType=accounting_storage/none
JobCompType=jobcomp/none
JobAcctGatherFrequency=30
JobAcctGatherType=jobacct_gather/none
SlurmctldDebug=info
SlurmctldLogFile=/var/log/slurm-llnl/slurmctld.log
SlurmdDebug=info
SlurmdLogFile=/var/log/slurm-llnl/slurmd.log
#
# COMPUTE NODES # THis machine has 128GB main memory
NodeName=$HOSTNAME CPUs=32 RealMemory==128762 State=UNKNOWN
PartitionName=local Nodes=ALL Default=YES MaxTime=INFINITE State=UP
EOF
sudo chmod 755 /etc/slurm-llnl/sudo systemctl start slurmctld
sudo systemctl start slurmd
# sudo scontrol update nodename=$HOSTNAME state=idle
sudo scontrol update nodename=$HOSTNAME state=resume
sudo systemctl stop slurmd
sudo systemctl stop slurmctld
sinfo
sinfo -R
sinfo -a
save into gregor.slurm
#!/bin/bash
#SBATCH --job-name=gregors_test # Job name
#SBATCH --mail-type=END,FAIL # Mail events (NONE, BEGIN, END, FAIL, ALL)
#SBATCH [email protected] # Where to send mail
#SBATCH --ntasks=1 # Run on a single CPU
#### XBATCH --mem=1gb # Job memory request
#SBATCH --time=00:05:00 # Time limit hrs:min:sec
#SBATCH --output=sgregors_test_%j.log # Standard output and error log
pwd; hostname; date
echo "Gregors Test"
date
sleep(30)
date
Run with
sbatch gregor.slurm
watch -n 1 squeue
BUG
JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)
2 LocalQ gregors_ green PD 0:00 1 (Nodes required for job are DOWN, DRAINED or reserved for jobs in higher priority partitions)
start slurm daemons
cms ee slurm startstop surm deamons
cms ee slurm stopBUG:
srun gregor.slurm
srun: Required node not available (down, drained or reserved)
srun: job 7 queued and waiting for resourcessudo scontrol update nodename=localhost state=POWER_UP
Valid states are: NoResp DRAIN FAIL FUTURE RESUME POWER_DOWN POWER_UP UNDRAIN
Continued work was in part funded by the NSF CyberTraining: CIC: CyberTraining for Students and Technologies from Generation Z with the awadrd numbers 1829704 and 2200409.
Command ee
==========
::
Usage:
ee generate submit --name=NAME [--job_type=JOB_TYPE] [--verbose]
ee generate --source=SOURCE --name=NAME
[--out=OUT]
[--verbose]
[--mode=MODE]
[--config=CONFIG]
[--attributes=PARAMS]
[--output_dir=OUTPUT_DIR]
[--dryrun]
[--noos]
[--os=OS]
[--nocm]
[--source_dir=SOURCE_DIR]
[--experiment=EXPERIMENT]
[--flat]
[--copycode=CODE]
ee list [DIRECTORY]
ee slurm start
ee slurm stop
ee slurm info
ee seq --yaml=YAML|--json=JSON
Expermine Executor (ee) allows the creation of parameterized batch
scripts. The initial support includes slurm, but we intend
also to support LSF. Parameters can be specified on the
commandline or in configuration files. Configuration files
can be formulated as json,yaml, python, or jupyter
notebooks.
Parameters defined in this file are then used in the slurm
batch script and substituted with their values. A special
parameter called experiment defines a number of variables
that are permuted on when used allowing multiple batch
scripts to be defined easily to conduct parameter studies.
Please note that the setup flag is deprecated and is in
future versions fully covered while just using the config
file.
Arguments:
FILENAME name of a slurm script generated with ee
CONFIG_FILE yaml file with configuration
ACCOUNT account name for host system
SOURCE name for input script slurm.in.sh, lsf.in.sh,
script.in.sh or similar
PARAMS parameter lists for experimentation
GPU name of gpu
Options:
-h help
--copycode=CODE a list including files and directories to be copied into the destination dir
--config=CONFIG... a list of comma seperated configuration files in yaml or json format.
The endings must be .json or .yaml
--type=JOB_TYPE The method to generate submission scripts.
One of slurm, lsf. [default: slurm]
--attributes=PARAMS a list of coma separated attribute value pairs
to set parameters that are used. [default: None]
--output_dir=OUTPUT_DIR The directory where the result is written to
--source_dir=SOURCE_DIR location of the input directory [default: .]
--account=ACCOUNT TBD
--gpu=GPU The name of the GPU. Tyoically k80, v100, a100, rtx3090, rtx3080
--noos ignores environment variable substitution from the shell. This
can be helpfull when debugging as the list is quite lareg
--nocm cloudmesh as a variable dictionary build in. Any vaiable referred to
by cloudmesh. and its name is replaced from the
cloudmesh variables
--experiment=EXPERIMENT This specifies all parameters that are used to create
permutations of them.
They are comma separated key value pairs
--mode=MODE one of "debug", "hierachical". One can also just
use "d", "h" [default: h]
--name=NAME Name of the experiment configuration file
--os=OS Selected OS variables
--flat produce flatdict
--dryrun flag to do a dryrun and not create files and
directories [default: False]
--verbose Print more information when executing [default: False]
Description:
> Examples:
>
> cms ee generate slurm.in.sh --verbose \\
> --config=a.py,b.json,c.yaml \\
> --attributes=a=1,b=4 \\
> --dryrun --noos --input_dir=example \\
> --experiment=\"epoch=[1-3] x=[1,4] y=[10,11]\" \\
> --name=a --mode=h
>
> cms ee generate slurm.in.sh \\
> --config=a.py,b.json,c.yaml \\
> --attributes=a=1,b=4 \\
> --noos \\
> --input_dir=example \\
> --experiment=\"epoch=[1-3] x=[1,4] y=[10,11]\" \\
> --name=a \\
> --mode=h
> >
> cms ee generate slurm.in.sh --experiments-file=experiments.yaml --name=a
>
> cms ee generate submit --name=a