METAQ (pronounced "meta-queue") is a system for bundling tasks for supercomputing in an environment with a batch scheduler like SLURM or PBS.
In many cases, you might prefer to bundle many small tasks together to achieve better throughput or a charging discount. METAQ is designed to allow you to submit big bundled jobs easily, by creating a queue for your own work (tasks), that will run in a near-optimal way, wasting very little of your allocation.
METAQ doesn't care about what the individual tasks are. So, if you have multiple projects going at once, you can nevertheless use the same METAQ and the tasks will be queued and run side-by-side with no problem.
Essentially, this allows you to back-fill your tasks, and to submit jobs that build priority in the queue that you can add tasks to after submission. Different users can put tasks in the same METAQ and (as long as the file permissions are liberal enough) can share the load.
JOB SCRIPTS all do the same thing when launched by the BATCH SCHEDULER: look through your TASK SCRIPTS and start what's possible.
The BATCH SCHEDULER controls the computational resources. This is chosen by the administrator of your computer. Examples include SLURM and PBS.
JOB SCRIPTS are bash scripts that get submitted to the scheduler,
contain some amount of user-specified METAQ options,
convert BATCH SCHEDULER variables to METAQ variables, and
call METAQ/x/launch.sh
TASK SCRIPTS get put in the METAQ/todo and METAQ/priority folders,
describe their needed resources with #METAQ flags, and
contain the hard work you want to do.
- METAQ
- Quick Summary
- Table of Contents
- Basic Introduction
- Task Script Structure
- Folder Structure
- METAQ Flags
- Job Script Structure
- Miscellaneous
- METAQ Monitoring
- Reports
- Installation
- Known Bugs and Complaints
- License
Users create various task scripts to do some hard work (meaning supercomputing) as usual. These task scripts can be essentially anything, as long as they contain some METAQ markup. They can include setup steps and whatever else you want. It is these task scripts that contain the command that tells the scheduler to put work onto the compute nodes (for example, in a SLURM environment these task scripts contain srun commands).
These task scripts exist in the METAQ/todo folder, or they can be symbolically linked in. They will be moved to the METAQ/working folder and ultimately to the METAQ/finished folder upon completion.
Users also need to forumulate jobs script. These scripts are what actually gets submitted to the scheduler. So, it needs to have all the usual batch scheduler mark up. It needs to convert the batch scheduler variables to METAQ variables. Then, it executes METAQ/x/launch.sh, which is where all the magic happens.
METAQ/x/launch.sh essentially evaluates these steps:
loop over all possible remaining tasks until there are none:
Check if you currently have enough resources (nodes, GPUs, clock time, etc.) to perform the task
If so, move it to the working directory, deduct those resources from what's available, and launch it!
Else, skip it!
But, if it is impossible, don't count this job as "remaining".
Some examples of impossibility:
The task needs more nodes than are allocated to this job.
The remaining clock time isn't enough to complete the task.
METAQ/x/launch.sh logs the individual tasks separately in METAQ/jobs/${METAQ_JOB_ID}/log/${task}.log and tallies the available resources in METAQ/jobs/${METAQ_JOB_ID}/resources, where you can see which tasks got started when, and what resources they consumed. My personal preference is to direct the output log of METAQ/x/launch.sh itself to METAQ/jobs/${METAQ_JOB_ID}/log/ as well, but that is specified in the job script as an option to the batch scheduler.
Task scripts must be directly executable, so if you script in a language that isn't bash, you must include the correct shebang. It need not be bash. The structure of a task script is only very loosely constrained.
You tell METAQ about the task by incorporating flags that METAQ understands into the task script. A bare-bones script looks like this:
#!/bin/bash
#METAQ GPUS 2
#METAQ NODES 2
#METAQ MIN_WC_TIME 5:00
#METAQ PROJECT metaq.example.1
do setup
load modules
etc
echo "working hard"
aprun -n 2 [...]
echo "finished"However, I often make the task scripts themselves legal submittable scripts, from the point of view of the batch scheduler. This way, if something goes askew (for example, a job hits the wall clock time before a task is complete) you can manually clean up very easily (though METAQ provides some amount of automatic clean-up too).
So, a task script (in a PBS environment) might look like
#!/bin/bash
#PBS -A lgt100
#PBS -l nodes=2
#PBS -l walltime=5:00
#PBS -N NAME_OF_TASK
#METAQ GPUS 2
#METAQ NODES 2
#METAQ MIN_WC_TIME 5:00
#METAQ PROJECT metaq.example.1
do_setup
load_modules
etc
echo "working hard"
aprun -n 2 [...]
echo "finished"This script can be qsubbed directly to PBS, thanks to the #PBS flags, but can also be understood by METAQ as a METAQ task.
However, as shown in the first example, all of the #PBS lines are not strictly necessary. The essential features are that the script indicates to METAQ how many NODES and GPUS are required and the much time should remain on the job's wall clock before starting this task in MIN_WC_TIME.
Task scripts inherit the bash environment variables as the batch scheduler provides them, but do not have access to METAQ variables, and are not passed any parameters. They should be relatively self-contained.
Tasks should not be invisible files or reside in invisible folders. METAQ ignores any task file with '/.' in its absolute path.
By default METAQ looks in the ${METAQ}/priority and ${METAQ}/todo folders. However, by setting the optional METAQ_TASK_FOLDERS variable, the user may instead specify the folders that METAQ should look in for tasks.
Sometimes it makes sense to organize tasks by their computational requirements. For example it might make sense to put all the 32 NODE, 0-GPU tasks together, separate from the 8 NODE, 8 GPU tasks.
METAQ looks in each folder for a special file .metaq. It parses that file for #METAQ FLAGS (as it would a task script) to short-circuit the need to check every file in that folder. For example, if METAQ knows it only has 4 NODES available and it's currently looking through a folder whose .metaq file claims that the folder contains jobs that requires 8 NODES, it will skip to the next folder.
It's important to understand that METAQ doesn't enforce the consistency of the folder's .metaq file and the tasks that folder contains. So, METAQ might skip over a task that it could execute if the folder's .metaq claims the associated tasks are bigger than the particular task.
If the .metaq file isn't there, then METAQ will loop over every file in the folder no matter what.
Valid FLAGs for a folder are NODES, GPUS, MIN_WC_TIME.
METAQ flags are always passed each on their own line. They always have the structure
#METAQ FLAG VALUE
Whitespace is not important, but there is never whitespace in a FLAG or in a VALUE.
METAQ only reads the first instance of a flag. So if you put the same flag on two different lines, METAQ will pick the first one.
FLAGs are always CAPITALIZED.
So, in the above example, you see
#METAQ GPUS 2
#METAQ NODES 2
#METAQ MIN_WC_TIME 5:00
#METAQ PROJECT metaq.example.1What FLAGs are understood by METAQ and what do they mean?
This task requires N nodes.
This task requires G gpus.
This task requires the job to still have HH:MM:SS available before starting.
This helps avoid having work that fails due to interruption.
You may specify times in the format understood by METAQ/x/seconds, which converts the passed string into a number of seconds. So, you don't have to specify canonically-formatted times. For example, you can specify 90:00 or 5400 instead of 1:30:00.
Write the running log of this task to the specified path.
METAQ doesn't worry about the task's project. However, it does log projects to METAQ/jobs/${METAQ_JOB_ID}/resources. This is convenient if you have many comingled projects (or parts of projects) in the same METAQ.
It (probably) makes sense to have the string prefixed in order of generality (most specific detail last).
THIS CURRENTLY DOESN'T DO ANYTHING.
The intention is: If multiple machines can see the same METAQ but, for example, have different hardware (and thus incompatible binaries) you can nevertheless keep the METAQ/todo folder comingled.
Job scripts are bash scripts that get submitted to your batch scheduler (SLURM, PBS, etc.). The structure of a job script is as follows. Job scripts MUST be bash (sorry), because that is the language METAQ itself is written in.
The script below shows all the user-specifiable options. For the optional options, their default values are shown as well.
#!/bin/bash
#MSUB -A [your account]
#MSUB -l nodes=10
#MSUB -l walltime=15:00
#MSUB [whatever other options]
#MSUB -o /absolute/path/to/metaq/log/%J.log
#MSUB -N metaq_job
# REQUIRED USER-SPECIFIED OPTIONS
METAQ=/full/path/to/metaq # Specifies the full path to the metaq folder itself.
METAQ_JOB_ID=${SLURM_JOB_ID} # Any string. You don't have to use the batch scheduler ID,
# but you should ensure that the METAQ_JOB_ID is unique.
METAQ_NODES=${SLURM_NNODES} # Integer, which should be less than or equal to the nodes specified to the batch scheduler.
# But if it's less than, you're guaranteeing you're wasting resources.
METAQ_RUN_TIME=900 # Seconds, should match the above walltime=15:00.
# You may also specify times in the format for the #METAQ MIN_WC_TIME flag, [[HH:]MM:]SS.
METAQ_MACHINE=machine # Any string. Helps organize the working directory.
# Could in the future interact with METAQ MACHINE flag.
# OPTIONAL USER-SPECIFIED OPTIONS, with their defaults
METAQ_TASK_FOLDERS=( # A bash array of priority-ordered absolute paths in which to look for tasks.
$METAQ/priority # This allows a user to segregate tasks and order their importance based on any number of
$METAQ/todo # metrics. Our original use was to separate tasks by nodes, so that we could waste as little
) # time as possible looking for a "big" task.
METAQ_GPUS=0 # An integer describing how many GPUs are allocated to this job.
# How many GPUs to specify is a bit of a subtle business. See below for more discussion.
METAQ_MAX_LAUNCHES=1048576 # An integer that limits the number of tasks that can be successfully launched. Default is 2^20, essentially infinite.
METAQ_LOOP_FOREVER=false # Bash booleans {true,false}. Should you run out the wall clock?
# If METAQ_LOOP_FOREVER is true then METAQ will continue to look for remaining tasks,
# even if it finds none and it is not waiting for any tasks to finish.
METAQ_SLEEP_AFTER_LAUNCH=0 # seconds to sleep after a task is successfully launched.
# We suspected that slowing down the launching might solve some unusual behavior we
# experienced when multiple METAQ jobs ran simultaneously.
METAQ_SLEEPY_TIME=3 # Number of seconds to sleep before repeating the main task-attempting loop.
METAQ_VERBOSITY=2 # How much detail do you want to see?
# Levels of detail are offset by tabbing 4 spaces.
METAQ_SIMULTANEOUS_TASKS=1048576 # An integer that limits how many tasks can run concurrently.
# Some environments limit how many simultaneous tasks you can submit. For example,
# [on Titan, users are artificially limited to 100 simultaneous aprun processes](https://www.olcf.ornl.gov/kb_articles/using-the-aprun-command/).
METAQ_MIN_NODES=0 # Integers that puts a lower size limit on tasks.
METAQ_MIN_GPUS=0 # If the main loop decides that there were no possible tasks, it will halve these minimal
# values and loop again. It will only concede that there are truly no possible jobs when
# these minimal values are <= 1.
METAQ_MAX_NODES=${METAQ_NODES} # Integers that puts an upper size limit on jobs.
METAQ_MAX_GPUS=${METAQ_GPUS} # If the main loop decides that there were no possible jobs, it will double these maximal
# values and loop again. It will only concede that there are truly no possible jobs when
# these maximal values max out at METAQ_NODES and METAQ_GPUS respectively.
METAQ_SORT_TASKS=sort # How to sort the result of finding tasks.
# A good choice can be shuf, the command-line utility that shuffles input lines.
METAQ_SKIP_ON_STOLEN=false # {false,directory,reset}, what to do if a task is stolen.
# On directory, break out of this directory and proceed to the next one.
# On reset, begin the loop over METAQ_TASK_FOLDERS again.
# On false, or anything else, just look at the next task.
# ANYTHING ELSE YOU WANT TO DO BEFORE LAUNCHING.
# For example, you can have this script resubmit itself.
# Or, if you have a script for populating the METAQ/{todo,priority} folders, you can run it now.
# ATTACK THE QUEUE
source ${METAQ}/x/launch.sh # ANYTHING BELOW HERE IS NOT GUARANTEED TO RUN!
# For example, if the tasks run out the wall clock time, METAQ/x/launch.sh never finishes.
# AND THAT IS ALL!If you name your job scripts x/q_* then git will ignore your job script. This will protect them as you need to pull new versions of METAQ.
THIS IS A SUBTLE BUSINESS.
On many machines there are only a few GPUs and many CPUs on a single physical node.
It may be possible to overcommit, meaning that you can run one binary on a few CPUs and the GPUs and another binary on the remaining CPUs.
If you CANNOT overcommit (eg. on Titan), then you can forget about GPUs altogether and just worry about NODEs, and think of them as the physical nodes on the machine.
If you CAN overcommit (eg. on Surface), then potentially what you should mean by GPU is (one GPU and one controlling CPU), while a NODE is (#CPUS per node - #GPUs per node). For example, on LLNL's GPU machine Surface each node has 2 GPUs and 16 CPUs. So, from the point of view of METAQ there are 2 GPUs and 1 14-CPU node.
However, it may be that your binary is smart and knows how make use of all the resources on the whole physical node at once. In that case you can again forget about GPUs and just worry about NODEs.
If some of your tasks are smart and can use the whole node and some cannot, you need to be careful. Because METAQ makes no promises about where the available NODEs and GPUs are. It could be, for example, that there is 1 GPU free over HERE and a bunch of CPUs free over THERE and then it may not make sense to run a 1 NODE 1 GPU job. In this case you should again think of NODES as representing full physical nodes (CPUs and GPUs) and simply be OK with wasting some of the resources some of the time.
So, to summarize: GPU is really a stand-in for a way to partition the physical node. You should make sure your partition makes sense and is compatible between all your tasks.
Finally, you can set up a task that could run on different machines that seem the same METAQ. However, currently METAQ doesn't know how to read machine-dependent settings. In the example where we first understood this meature of METAQ, one machine had nodes with huge memory while the other didn't, meaning that on the first we could run on one physical node and on the other we needed four physical nodes. By writing a machine-aware task script, setting the #METAQ NODE flag to 1, and setting METAQ_NODES to (physical nodes / 4) on the smaller machine, we could circumvent this hardware requirement mismatch.
METAQ provides a number of small accessories to see what's going on.
Reports which jobs are in the working directory, machine-by-machine unless a machine is provided. The first column is the machine/job, the second is how many are currently working, the last is how many log files in total there are.
Reports based on tasks' PROJECT flag what's in the METAQ/{priority,todo,hold} folders.
Reports the current status of jobs in the working folder. This only works if you have set up
WARNING WARNING WARNING
SERIOUSLY
HANDLE WITH CARE!
If your queue contains only work for one machine, this is perfectly fine.
However, if METAQ_JOB_RUNNING (see below on interacting with the batch scheduler) could potentially miss the existence of a job then these commands can wreck havoc.
These scripts look in the working/${machineArgument} directory for jobs that are no longer running.
If it finds work for a METAQ_JOB_ID that is no longer running, x/abandon deletes the METAQ/working/${machineArgument}/METAQ_JOB_ID folder, x/reset moves the task scripts into the METAQ/priority folder first.
x/{abandon,reset} will compare $HOSTNAME to ^${machineArgument}.*$. If there's a match, the assumption is that running x/reset is probably OK. If there's a mismatch, the user will get a warning and prompted for confirmation.
For some accessories, METAQ needs to know about your batch scheduler.
It always looks in the shell file (or symbolic link) METAQ/x/batch.sh for information about your batch scheduler. Files for SLURM and PBS are provided.
If you use a different batch scheduler, then you can create your own file. Here are the specifications:
#!/bin/bash
METAQ_BATCH_SCHEDULER="SCHEDULER_NAME"
# A variable which indicates what batch scheduler you're using. Currently unused, but could be used in the future. Just set it for documentation's sake.
# I imagine some day this information can be used to create job script files.
function METAQ_JOB_RUNNING {
# JOB_ID --> STRING
# A function that returns the empty string if the job does not exist.
}
Because it tabulates consumed resources and keeps track of the associated task files, I hope to be able to generate some smart reports. As it stands, this is totally unimplemented.
Simply git clone into any directory that is readable and writable from the supercomputing nodes.
cd into that directory and run x/install. It will ask you what batch scheduler you are using. It will provide (hopefully) intelligent defaults.
You should rerun x/install after ever git pull.
You can perform a bare-bones test from the METAQ directory by running x/demo.sh.
Some of the accessory scripts detect their own location on disk and infer what METAQ they belong to. These include:
- x/current
- x/report
- x/reset
- x/running
- x/status
-
The business about NODEs and GPUs is more subtle than is ideal. But without making METAQ substantially more complicated I don't know how to solve the issue.
-
Tasks in the
METAQ/priorityfolder are only preferred at the beginning of a job.
METAQ grew out of finding collaborative HPC projects increasingly difficult to coordinate, and is a ground-up rewrite of scripts I originally crafted with Thorsten Kurth. The separation of job submission from
task description was emphasized as a useful trick by Chris Schroeder when I began working on lattice QCD at LLNL.
METAQ inspired my collaborators and I to work on mpi_jm, a C version that lives at the MPI level so that it stresses the head nodes of machines much more lightly (we crashed Titan with METAQ, although we were within the bounds of its user guide) and only relies on a single aprun (or srun, or equivalent). mpi_jm should be finished sooner rather than later, and I will link it here when it is in a presentable state, at which point development of METAQ will likely cease entirely, deprecated in favor of mpi_jm, though I am happy to accept pull requests. There will still be some situations where METAQ may be preferable---since mpi_jm will require recompiling against it as a library (and adding about 5 lines of code to your executable).
I posted a short document to the arXiv, 1702.06122. Here's a bibtex entry that I modified from inSPIRE I'd appreciate if you cited if you find METAQ useful
@article{Berkowitz:2017vcp,
author = "Berkowitz, Evan",
title = "{\texttt{METAQ}: Bundle Supercomputing Tasks}",
year = "2017",
eprint = "1702.06122",
archivePrefix = "arXiv",
primaryClass = "physics.comp-ph",
SLACcitation = "%%CITATION = ARXIV:1702.06122;%%",
url = {https://github.com/evanberkowitz/metaq},
howpublished = {\url{https://github.com/evanberkowitz/metaq}}
}
or drop me an email / buy me a beer at a conference if METAQ helped you push your project through the supercomputers efficiently. But, the license is GPLv3.0. I would love there to be a free software license that was as liberal as the GPL but had a binding citation requirement upon use (rather than upon redistribution).
METAQ
Bundle supercomputing tasks.
Copyright (C) 2016 Evan Berkowitz
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
I also published a proceedings for LATTICE 2017 the discussing METAQ and bundling of tasks more generally. The inSPIRE entry is
@article{Berkowitz:2017xna,
author = "Berkowitz, Evan and Jansen, Gustav R. and McElvain,
Kenneth and Walker-Loud, André",
title = "{Job Management and Task Bundling}",
booktitle = "{Proceedings, 35th International Symposium on Lattice
Field Theory (Lattice 2017): Granada, Spain, June 18-24,
2017}",
journal = "EPJ Web Conf.",
volume = "175",
year = "2018",
pages = "09007",
doi = "10.1051/epjconf/201817509007",
eprint = "1710.01986",
archivePrefix = "arXiv",
primaryClass = "hep-lat",
SLACcitation = "%%CITATION = ARXIV:1710.01986;%%"
}