Cram

Cram runs many small MPI jobs inside one large MPI job.

Released under the GNU LGPL, LLNL-CODE-661100. See the LICENSE file for details.

Overview

Suppose you have a supercomputer with 1 million cores, and you want to run 1 million 1-process jobs. If your resource manager is scalable, that is most likely an easy thing to do. You might just submit 1 million jobs:

srun -n 1 my_mpi_application input.1.txt
srun -n 1 my_mpi_application input.2.txt
...
srun -n 1 my_mpi_application input.1048576.txt

Unfortunately, for some resource managers, that requires you to have 1 million srun processes running on the front-end node of your fancy supercomputer, which quickly blows out either memory or the front-end OS's process limit.

Cram fixes this problem by allowing you to lump all of those jobs into one large submission. You use Cram like this:

cram pack -f cram.job -n 1 -e my_mpi_application input.1.txt
cram pack -f cram.job -n 1 -e my_mpi_application input.2.txt
...
cram pack -f cram.job -n 1 -e my_mpi_application input.1048576.txt

This packs all those job invocations into a file called cram.job, and saves them for one big submission.

To use the file, either:

• Link your application with libcram.a if you are using C/C++.
• Link your application with libfcram.a if you are using Fortran.

Then, run your job like this:

env CRAM_FILE=/path/to/cram.job srun -n 1048576 my_mpi_application

That's it. The job launches, it splits itself into a million pieces. You'll see some output like this from your srun invocation if Cram is running properly:

===========================================================
 This job is running with Cram.

 Splitting this MPI job into 1048576 jobs.
 This will use  total processes.

 Successfully set up job:
   Job broadcast:   0.005161 sec
   MPI_Comm_split:  0.001973 sec
   Job broadcast:   0.078768 sec
   File open:       0.700172 sec
  --------------------------------------
   Total:           0.786074 sec

===========================================================

Each runs independently of the others, generates its own output and error files, then terminates as it normally would. All you need to make sure of is that the final, large submitted job has at least as many processes as all the small jobs combined.

When the job completes, you should see output files like this:

cram.0.out   cram.1.err    cram.3.out   cram.4.err
cram.0.err   cram.2.out    cram.3.err   cram.5.out   ... etc ...
cram.1.out   cram.2.err    cram.4.out   cram.5.err

If the jobs in your cram file had different working directories, then these files will appear in the working directories.

NOTE: By default, only rank 0 in each job in your cram file will write to the output and error files. If you don't like this, Cram has other options for output handling. See Output Options below for more on this.

Setup

This section covers how to use Cram at LLNL. If you're at another site where Cram isn't installed yet, see Build & Install below to build your own.

To run Cram, you need Python 2.7. Python and cram are available at LLNL through dotkits:

use python-2.7.3
use cram

Once you use cram, the CRAM_HOME environment variable is set for you. You can use this to locate libcram.a:

$ ls $CRAM_HOME/lib
libcram.a  python2.7/

That's the library you want to link your application with before you submit your large job.

Command Reference

Nearly everything in Cram is done through the cram command. Like svn, git, and various other UNIX commands, cram is made up of many subcommands. They're documented here.

cram pack

The most basic cram command -- this packs command line invocations into a file for batch submission.

usage: cram pack [-h] -n NPROCS -f FILE [-e EXE] ..

• -n NPROCS Number of processes this job should run with.

• -f FILE File to store command invocation in. Default is 'cram.job'

• -e EXE

  Optionally specify the executable name to use. By default, when you launch a cram job, The name of the launched binary is used (argv[0]). If you need to override this, you can supply -e with an argument.

• ... Command line arguments of the job to run, not including the executable.

cram info

This command is useful if you are trying to debug your run and you need to see what you actually submitted.

Usage: cram info [-h] [-a] [-j JOB] [-n NUM_LINES] cramfile

There are three modes:

1. Summary mode

cram info <cramfile> will show you a summary of the cram file, e.g.:

$ cram info test-cram.job
Name:            test-cram.job
Number of Jobs:              3
Total Procs:                82
Cram version:                1

Job command lines:
    0     35 procs    my_app foo bar 2 2 4
    1     35 procs    my_app foo bar 2 4 2
    2     12 procs    my_app foo bar 4 2 2

2. Job detail

cram info -j 5 <cramfile> will show you information about the job with index 5 inside the cram file. That includes:

1. Size (number of processes) of the job
2. Its working directory
3. Its command line
4. All of its environment variables.

Example:

cram info -j 2 test-cram.job
Job 2:
  Num procs: 12
  Working dir: /g/g21/gamblin2/test
  Arguments:
      my_app foo bar 4 2 2
  Environment:
      'LOG_DIRECTORY' : '/p/lscratcha/my_app/output'
      ... etc ...

3. All data

cram info -a <cramfile> will print out all information for all jobs in the file. This can be very verbose, so use it carefully.

cram test

Usage: cram test [-h] [-l] [-v] [names [names ...]]

Running cram test will run Cram's Python test suite. If you think something is wrong wtih Cram, use this command to see if it catches anything.

Packing lots of jobs

If you want to pack millions of jobs, you most likely do NOT want to call cram millions of times. Python starts up very slowly, and you don't want to pay that cost for each pack invocation. You can get around this by writing a simple Python script. Alongside the cram script, there is a cram-python script. You can use this to make an executable script like this one:

#!/usr/bin/env cram-python

import os
from cram import *

cf = CramFile('cram.job', 'w')
env = os.environ

# Pack cram invocations.
# Usage:
#   cf.pack(<num procs>, <working dir>, <command-line arguments>, <environment>)
env["MY_SPECIAL_VAR"] = "my_value"
cf.pack(1, '/home/youruser/ensemble/run-0', ['input.1.txt'], env)
cf.pack(1, '/home/youruser/ensemble/run-1', ['input.2.txt'], env)
# ...
env["MY_SPECIAL_VAR"] = "another_value"
cf.pack(1, '/home/youruser/ensemble/run-1048576', ['input.1048576.txt'], env)

cf.close()

This script will create a cram file, just like all those invocations of cram pack above, but it will run much faster because it runs in a single python session. Note that CramFile.pack() takes similar arguments as the cram pack command.

Here's a more realistic one, for creating a million jobs with different user-specific scratch directories:

#!/usr/bin/env cram-python

import os
import getpass
from cram import *

env  = os.environ
user = getpass.getuser()

cf = CramFile('cram.job', 'w')
for i in xrange(1048576):
    env["SCRATCH_DIR"] = "/p/lscratcha/%s/scratch-%08d" % (user, i)
    args = ["input.%08d" % i]
    cf.pack(1, '/home/%s/ensemble/run-%08d' % (user, i), args, env)
cf.close()

Output Options

By default, Cram will redirect the stdout and stderr streams to cram.<jobid>.out and cram.<jobid>.err ONLY for rank 0 of each cram job. Other processes have their output and error redirected to /dev/null. You can control this behavior using the CRAM_OUTPUT environment variable. The options are:

• NONE: No output. All processes redirecto to /dev/null.
• RANK0: Default behavior. Rank 0 writes to cram.<jobid>.out and cram.<jobid>.err. All other processes write to /dev/null.
• ALL: All processes write to unique files. e.g., rank 4 in job 1 writes to cram.1.4.out and cram.4.1.err.
• SYSTEM: Ouptut is not redirected and system defaults are used.

To set a particular output mode, set the CRAM_OUTPUT environment variable to one of the above values when you run. For example, to make every process write an output and error file, you would run like this:

env CRAM_OUTPUT=ALL CRAM_FILE=/path/to/cram.job srun -n 1048576 my_mpi_application

Error reporting

You may notice that in the NONE and RANK0 modes, some processes don't have a place to report errors. Cram tries its best to report errors regardless of the output mode the user has chosen.

In particular, it tries to report non-zero exit() calls and segmentation faults (SIGSEGV) even if not every process has an output file.

To do this, Cram will write error messages to the original error stream in all modes except ALL. So, if a process dies or exits, you'll see a message like this print out to the console:

Rank 1 on cram job 4 died with signal 11.

If you are running in ALL mode, Cram will print error messages like this out to the per-process cram.<job>.<rank>.err file.

Build & Install

To build your own version of Cram, you need CMake. On LLNL machines, you can get cmake by typing use cmake.

Basic build

To build, run this in the Cram root directory:

cmake -D CMAKE_INSTALL_PREFIX=/path/to/install .
make
make install

CMAKE_INSTALL_PREFIX is the location you want to install Cram, and . is the path to the top-level Cram source directory.

CMake supports out-of-source builds, so if you like to have separate build directories for different machines, you can run the above command in a subdirectory. Just remember to change the path to source to .. or similar when you do this, e.g.:

mkdir  $SYS_TYPE && cd $SYS_TYPE
cmake -DCMAKE_INSTALL_PREFIX=/path/to/install ..

Cross-compiling

On Blue Gene/Q, you also need to supply -DCMAKE_TOOLCHAIN_FILE to get cross compilation working:

cmake -DCMAKE_INSTALL_PREFIX=/path/to/install \
      -DCMAKE_TOOLCHAIN_FILE=cmake/Toolchain/BlueGeneQ-xl.cmake .
make
make install

If you want ot build with the GNU compilers, there is a BlueGeneQ-gnu.cmake toolchain file too. This package supports parallel make, so you are free to use make -j if you like.

Other notes

This tool would have been called "Clowncar", but it was decided that it should have a more serious name. Cram stands for "Clowncar Renamed to Appease Management".

Authors

Cram was written by Todd Gamblin ([email protected]).

Many contributions, bugfixes, and testing were provided by John Gyllenhaal ([email protected]).

References

[1] J. Gyllenhaal, T. Gamblin, A. Bertsch, and R. Musselman. Enabling High Job Throughput for Uncertainty Quantification on BG/Q. In IBM HPC Systems Scientific Computing User Group (ScicomP’14), Chicago, IL, 2014.

[2] J. Gyllenhaal, T. Gamblin, A. Bertsch, and R. Musselman. "Cramming" Sequoia Full of Jobs for Uncertainty Quantification. In ASC eNews Quarterly Newsletter. June 2014.