Distributed processing of a batch of tasks.

Overview

One common usage pattern for distributed computing involves processing a long list of commands (aka tasks):

cd /path/to/workdir ; myprog argsFor000 > task000.log 2>&1
cd /path/to/workdir ; myprog argsFor001 > task001.log 2>&1
... 
cd /path/to/workdir ; myprog argsFor998 > task998.log 2>&1
cd /path/to/workdir ; myprog argsFor999 > task999.log 2>&1

One could do this by submitting 1,000 separate jobs to a cluster computer, but that may present problems for the queuing system and can behave badly if the system is configured to handle jobs in a simple first come, first serve fashion.

Another alternative is to use a resource management system's native support for job arrays, but this often requires massaging the commands to reflect syntax specific to a particular system's implementation of job arrays.

And what if you don't have a cluster available, but do have a collection of networked computers?

In any event, when processing such a list of tasks, it is helpful to acquire metadata about the execution of each task: where it ran, how long it took, its exit return code, etc.

disBatch.py has been designed to support this usage in a simple and portable way, as well as to provide the sort of metadata that can be helpful for debugging and reissuing failed tasks.

It can take as input a file, each of whose lines is a task in the form of a command sequence (as in the above example). It launches the tasks one after the other until all specified execution resources are in use. Then as one executing task exits, the next task in the file is launched until all the lines in the file have been processed.

In the simplest case, working with a cluster managed by SLURM, all that needs to be done is to write the command sequences to a file and then submit a job like the following:

sbatch -n 20 --ntasks-per-node 5 --exclusive --wrap "disBatch.py TaskFileName"

This particular invocation will allocate sufficient resources to process 20 tasks at a time, with no more than five running concurrently on any given node. Every node allocated will be running only tasks associated with this submission. disBatch will use environment variables initialized by SLURM to determine the execution resources to use for the run.

Various log files will be created as the run unfolds:

• TaskFileName_134504_status.txt (assuming the SLURM Job ID is 134504): status of every task (details below)
• disBatch_134504_driver.txt (can be changed with -l), TaskFileName_134504_worker032_engine.txt: The disBatch log file contains details mostly of interest in case of a problem with disBatch itself. It can generally be ignored by end users (but keep it around in the event that something did go wrong---it will aid debugging). The ``*_engine.txt'' files contain similar information for each node acting as an execution resource.
• disBatch_134504_kvsinfo.txt: TCP address of invoked KVS server if any (for additional advanced status monitoring)

Installation

disBatch.py requires the kvsstcp package, which should be installed in python's path, or placed in this directory. You can simply clone this git repository with --recursive (or run git submodule update --init if you've already cloned it).

disBatch is designed to support a variety of execution environments, from a local collection of workstations to large clusters managed by job schedulers. It currently supports SLURM and can be executed from sbatch, but it is architected to make it simple to add support for other resource managers.

You can also run directly on one or more machines over ssh by setting an environment variable:

DISBATCH_SSH_NODELIST=host1:7,host2:3

This allows execution via ssh (or directly on localhost) without the need for a resource management system. In this example, disBatch is told it can use seven CPUs on host1 and three on host2. Assuming the default mapping of one task to one CPU applies in this example, seven tasks could be in progress at any given time on host1, and three on host2. Hosts used via ssh must be set up to allow ssh to work without a password.

Depending on your execution environment, the ability of disBatch to determine the location of itself and kvsstcp may be disrupted. If you get errors about not finding disBatch or kvsstcp, you may need to hard-code the paths for your setup into the script. You can do this automatically by running ./disBatch.py --fix-paths. This should only need to be done once.

Invocation

usage: disBatch.py [-h] [--fix-paths] [-p PATH] [-l FILE] [--mailFreq N]
                   [--mailTo ADDR] [-c N] [-t N] [-g] [-k COMMAND] [-K]
                   [-s HOST:COUNT] [-r STATUSFILE] [-R] [--force-resume] [-w]
                   [--kvsserver [HOST:PORT]] [--taskcommand COMMAND]
                   [--taskserver [HOST:PORT]]
                   [taskfile]

Use batch resources to process a file of tasks, one task per line.

positional arguments:
  taskfile              File with tasks, one task per line ("-" for stdin)

optional arguments:
  -h, --help            show this help message and exit
  --fix-paths           Configure fixed path to script and modules.
  -p PATH, --prefix PATH
                        Prefix path and name for log and status files
                        (default: ./TASKFILE_JOBID).
  -l FILE, --logfile FILE
                        Log file.
  --mailFreq N          Send email every N task completions (default: 1). "--
                        mailTo" must be given.
  --mailTo ADDR         Mail address for task completion notification(s).
  -c N, --cpusPerTask N
                        Number of cores used per task; may be fractional
                        (default: 1).
  -t N, --tasksPerNode N
                        Maximum concurrently executing tasks per node (up to
                        cores/cpusPerTask).
  -g, --gpu             Use assigned GPU resources
  -k COMMAND, --retire-cmd COMMAND
                        Shell command to run to retire a node (environment
                        includes $NODE being retired, remaining $ACTIVE node
                        list, $RETIRED node list; default based on batch
                        system).
  -K, --no-retire       Don't retire nodes from the batch system (e.g., if
                        running as part of a larger job); equivalent to -k ''.
  -s HOST:COUNT, --ssh-node HOST:COUNT
                        Run tasks over SSH on the given nodes (can be
                        specified multiple times for additional hosts;
                        equivalent to setting DISBATCH_SSH_NODELIST)
  -r STATUSFILE, --resume-from STATUSFILE
                        Read the status file from a previous run and skip any
                        completed tasks (may be specified multiple times).
  -R, --retry           With -r, also retry any tasks which failed in previous
                        runs (non-zero return).
  --force-resume        With -r, proceed even if task commands/lines are
                        different.
  -w, --web             Enable web interface.
  --kvsserver [HOST:PORT]
                        Use a running KVS server.
  --taskcommand COMMAND
                        Tasks will come from the command specified via the KVS
                        server (passed in the environment).
  --taskserver [HOST:PORT]
                        Tasks will come from the KVS server.

The options for mail will only work if your computing environment permits processes to access mail via SMTP.

A value for -c < 1 effectively allows you to run more tasks concurrently than CPUs specified for the run. This is somewhat unusual, and generally not recommended, but could be appropriate in some cases.

The -k and -K flags allow you to control what disBatch does when a node is no longer needed to run jobs. When running under slurm, disBatch will by default run the command:

scontrol update JobId="$SLURM_JOBID" NodeList="${DRIVER_NODE:+$DRIVER_NODE,}$ACTIVE"

which will tell slurm to release any nodes no longer being used. You can set this to run a different command, or nothing at all. While running this command, the follow environment variables will be set: NODE (the node that is no longer needed), ACTIVE (a comma-delimited list of nodes that are still active), RETIRED (a comma-delimited list of nodes that are no longer active, including $NODE), and possibly DRIVER_NODE (the node still running the main disBatch script, if it's not in ACTIVE).

The -g argument parses the CUDA environment varables (CUDA_VISIBLE_DEVICES, GPU_DEVICE_ORDINAL) provided on each node and divides the resources between the running tasks. For example, with slurm, if you want to run on n nodes, with t tasks per node, each using c CPUs and 1 GPU (that is, tc CPUs and t GPUs per node, or ntc CPUs and nt GPUs total), you can do:

sbatch -N$n -c$c --ntasks-per-node=$t --gres=gpu:$t -p gpu --wrap 'disBatch.py -g $taskfile'`

-r uses the status file of a previous run to determine what tasks to run during this disBatch invocation. Only those tasks that haven't yet run (or with -R, those that haven't run or did but returned a non-0 exit code) are run this time. By default, the numeric task identifier and the text of the command are used to determine if a current task is the same as one found in the status file. --force-resume restricts the comparison to just the numeric identifier.

--kvsserver, --taskcommand, and --taskserver implement advanced functionality (placing disBatch in an existing shared key store context and allowing for a programmatic rather than textual task interface). Contact the authors for more details.

Status file

The _status.txt file contains tab-delimited lines of the form:

314	315	-1	worker032	8016	0	10.0486528873	1458660919.78	1458660929.83	0	""	0	""	'cd /path/to/workdir ; myprog argsFor314 > task314.log 2>&1'

These fields are:

1. Flags: The first field, blank in this case, may contain E, O, R, B, or S flags. Each program/task should be invoked in such a way that standard error and standard output end up in appropriate files. If that's not the case E or O flags will be raised. R indicates that the task returned a non-zero exit code. B indicates a barrier (see below). S indicates the job was skipped (this may happen during "resume" runs).
2. Task ID: The 314 is the 0-based index of the task (starting from the beginning of the task file, incremented for each task, including repeats).
3. Line number: The 315 is the 1-based line from the task file. Blank lines, comments, directives and repeats may cause this to drift considerably from the value of Task ID.
4. Repeat index: The -1 is the repeat index (as in this example, -1 indicates this task was not part of a repeat directive).
5. Node: worker032 identifies the node on which the task ran.
6. PID: 8016 is the PID of the bash shell used to run the task.
7. Exit code: 0 is the exit code returned.
8. Elapsed time: 10.0486528873 (seconds),
9. Start time:1458660919.78 (epoch based),
10. Finish time: 1458660929.83 (epoch based).
11. Bytes of leaked output (not redirected to a file),
12. Output snippet (up to 80 bytes consisting of the prefix and suffix of the output),
13. Bytes of leaked error output,
14. Error snippet,
15. Command: cd ... is the text of the task (repeated from the task file, but see below).

Considerations for large runs

If you do submit jobs with order 10000 or more tasks, you should carefully consider how you want to organize the output (and error) files produced by each of the tasks. It is generally a bad idea to have more than a few thousand files in any one directory, so you will probably want to introduce at least one extra level of directory hierarchy so that the files can be divided into smaller groups. Intermediate directory 13, say, might hold all the files for tasks 13000 to 13999.

#DISBATCH directives

PREFIX and SUFFIX

In order to simplify task files, disBatch supports a couple of directives to specify common task prefix strings and suffix strings. It also provides environment variables to identify various aspects of the submission. Here's an example

# Note there is a space at the end of the next line.
#DISBATCH PREFIX cd /path/to/workdir ; 
#DISBATCH SUFFIX > ${DISBATCH_NAMETASKS}_${DISBATCH_JOBID}_${DISBATCH_TASKID}.log 2>&1

These are textually prepended and appended, respectively, to the text of each task line. If the suffix includes redirection and a task is a proper command sequence (a series of program invocations joined by ;), then the task should be wrapped in ( ... ) (which may also be added to suffix and prefix respectively) so that the standard error and standard output of the whole sequence will be redirected to the log file. If this is not done, only standard error and standard output for the last component of the command sequence will be captured. This is probably not what you want unless you have redirected these outputs for the previous individual parts of the command sequence.

Using these, the above commands could be replaced with:

myprog argsFor000
myprog argsFor001
 ... 
myprog argsFor998
myprog argsFor999

Note: the log files will have somewhat different, but still task specific, names.

Later occurrences of #DISBATCH PREFIX or #DISBATCH SUFFIX in a task file simply replace previous ones. When these are used, the tasks reported in the status file include the prefix and suffix in force at the time the task was launched.

BARRIER

If your tasks fall into groups where a later group should only begin after all tasks of the previous group have completely finished, you can use this directive:

#DISBATCH BARRIER

When disBatch encounters this directive, it will not launch another task until all tasks in progress have completed. Advanced feature: if BARRIER is followed by a string, that string is interpreted as a key name. When the barrier completes, the Task ID of the barrier will be put to that key.

REPEAT

For those problems that are easily handled via a job-array-like approach:

 #DISBATCH REPEAT 5 start 100 step 50 [command]

will expand into five tasks, each with the environment variable DISBATCH_REPEAT_INDEX set to one of 100, 150, 200, 250, or 300. The tasks will consist of the concatenation of the prefix, command (if provided), and the suffix currently in effect. start defaults to 0, step to 1. Note: the semantics here differ somewhat from many range constructs, the number immediately following REPEAT sets the number of tasks that will be executed; the next two numbers affect only the value that the repeat index will have in the environment for each of the repeat task instances. So, returning to our earlier example, the task file could be:

#DISBATCH PREFIX cd /path/to/workdir ; myprog argsFor$(printf "%03d" ${DISBATCH_REPEAT_INDEX}) > ${DISBATCH_NAMETASKS}_${DISBATCH_JOBID}_${DISBATCH_TASKID}.log 2>&1
#DISBATCH REPEAT 1000

License

Copyright 2017 Simons Foundation

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.