Utilities for building CASPER and running experiments.

Clone this repository

This repository needs to be cloned carefully, because it makes use of:

• submodules to depend on repositories that need to be separate
• git LFS (Large File Storage) to store source tarballs

The cloning proceedure is:

1. clone the repository non-recursively from the "closest" clone (aka. mirror)
2. run the bootstrap script that initializes submodules and git lfs objects

On supported HPC clusters, this repository along with LFS storage is available on respective shared file systems, and it should be cloned from there, so that the tarballs are copied quickly rather than downloaded.

The clone commands below select a release rather than master tip. The workflow is: (1) build a release, (2) update to master tip (accoding to instructions in the Update section below). This is necessary because only releases have been tested to build without errors, while the master tip is very likely to be broken (because it moves quickly and is not tested on all platforms on every commit). Not all releases have been tested on all platforms, select the release tested on your platform from the ones available on releases page on GitHub.

In the following commands, replace RELEASE_ID with the tag name chosen above.

Take care to include the file:// URI prefix, since required by git-lfs.

• On OLCF Summit:

    $ git clone -b RELEASE_ID file:///gpfs/alpine/csc436/proj-shared/repos/casper-utils.git
    $ cd casper-utils
    $ ./bootstrap.sh
  

• On ANL Theta:

    $ git clone -b RELEASE_ID file:///lus/theta-fs0/projects/CASPER/repos/casper-utils.git
    $ cd casper-utils
    $ ./bootstrap.sh
  

• On ISI gpuk40:

    $ git clone -b RELEASE_ID file:///home/pub/casper/repos/casper-utils.git
    $ cd casper-utils
    $ ./bootstrap.sh
  

• On Generic Linux system: no public LFS server is available yet, and LFS cannot transfer objects over SSH (only over HTTP or over local file://), this leaves the following options:

  • clone the repo without LFS objects (won't be able to build the Prefix):

    $ GIT_LFS_SKIP_SMUDGE=1 git clone --recursive -b RELEASE_ID https://github.com/ISI-apex/casper-utils.git
    

  • if you have SSH access to one of the above systems, transfer the bare repo to your system and clone from it:

    $ rsync -aP USER@HOST:/PATH/TO/casper-utils.git .
    $ git clone file://$PWD/casper-utils.git casper-utils
    $ cd casper-utils
    $ ./bootstrap.sh
    

  • if you don't have access yourself, then you could ask someone who does to make you a tarball of the bare repo directory casper-utils.git/.

    $ tar xf casper-utils.git.tar
    $ git clone file://$PWD/casper-utils.git casper-utils
    $ cd casper-utils
    $ ./bootstrap.sh
    

  • if you have SSH access to one of the above systems, mount the remote file system using sshfs (sshfs need to be installed as root on your system, because fusermount needs to have setuid root permission):

    $ mkdir -p /mnt/host/repos
    $ sshfs USER@HOST:/PATH/TO/repos /mnt/host/repos
    $ git clone file:///mnt/host/repos/casper-utils.git casper-utils
    $ cd casper-utils
    $ ./bootstrap.sh
    $ fusermount -u /mnt/host/repos
    

Note: The bootstrapped git-lsf is inaccessible outside the bootstrap script deliberately. After you build the Prefix, do all git operations from within the Prefix to use the git and git-lfs that's installed in the Prefix. This has the added benefit of a recent git version that works faster with submodules than the ancient git versions installed on HPC clusters.

Note: The clones spread around various platforms are all subordinate to the master clone on GitHub. When pushing changes, you must push to GitHub clone first, and only after that you can push to the other clones. Having a designated master clone is critical, because a synchronization funnel is needed among all the committers, otherwise clones could diverge.

Note: We do not use GitHub's free-tier LFS storage, but there's no way to disable it for the repository on Github.com [1], no way to disable per remote on the client [2], and even no working way to override the LFS storage endpoint per remote on the client (git config remote.REMOTENAME.lfsurl is not honored). So, the workaround is: before pushing to GitHub make sure to disable LFS.

Assuming you cloned from a clone other than the GitHub clone (see above), add GitHub as a remote clone under the name of gh (only need to this once after you cloned above):

$ git remote add gh [email protected]:ISI-apex/casper-utils.git

Then, when you want to push to GitHub, temporarily disable LFS (note, this assumes git-lfs is not installed globally):

$ git lfs uninstall --local
$ git push gh master
$ git lfs install --local

Build Gentoo Prefix with CASPER and dependencies

A Gentoo Prefix is a way to manage dependencies installed from source with a full-featured package manager without root on a system, hence particularly useful on HPC clusters. Conceptually, in essence, it's a wrapper around ./configure && make && make install. Binaries built within the prefix may be called from the host as any other binary (although it's usually useful to work in a prefix shell, see below), and host binaries may be invoked from within the prefix seemlessly.

A Prefix allows everything to be independent of the host distribution (prefix bootstrap is very robust and should boostrap on any Linux system) and thus the versions of any of the software libraries, toolchains, and applications may be chosen at will. A Prefix is a lot more useful than Singularity or similar containers based on immutable images. The following sections describe how to build a prefix on USC HPCC cluster and on other hosts.

NOTE: Never run make install or similar in the prefix, because this will write and override files everywhere without any way to track then nor remove them, which will brake your prefix in the strangest ways imaginable, without any simple way to recover (you are not protected by root, since all paths in the prefix are user-writable, so be careful what commands you run). To install software into the prefix, it must be packaged in an .ebuild. It is usually very quick and easy to write a new .ebuild for some new software library or app, starting from an existing .ebuild as a template, and there are tens of thousands of packages already available (see eix searchstring).

Run build job

In all sections that follow:

• the first argument (PREFIX_PATH) is a relative or absolute folder where the prefix will be built (ensure several GB of space, and note that after prefix is built it cannot be relocated to another path)
• the second argument is a Gentoo profile name specific to the given platform

NOTE: before building, remember to checkout the latest release verified to build on the system that you're building on (see details in Step 0).

The build script is incremental, so in case of errors, troubleshoot them, and then simply restart the build job by the same method you used to start it.

USC HPCC or Discovery

Use this wrapper script to launch the build job on worker nodes:

$ casper-utils/jobs/usc/gpref.job.sh PREFIX_PATH casper-usc-CLUSTER-ARCH CLUSTER[:PARTITION] ARCH

• the second argument is a Gentoo profile name where CLUSTER identifies the HPC cluster (usc-hpcc or usc-discovery) and ARCH identifies the CPU family for which to optimize via the -march,-mtune compiler flags (for the generic unoptimized use amd64; for supported clusters and cpu families see ebuilds/profiles/casper-usc-*-*),
• the third argument is the cluster name: either discovery or hpcc, optionally followed by colon and a partition name, e.g. oneweek.
• the fourth argument ARCH again, but cannot be amd64; even if you want a generic (unoptimized) build, you still have to choose a CPU family for the build host (sandybridge is a reasonable choice for building a generic prefix, see notes below).

ANL Theta

On ANL Theta, the build can be done on a login machine:

$ casper-utils/jobs/gpref.sh PREFIX_PATH casper-anl-theta-knl

OLCF Summit

On OLCF Summit, choose PREFIX_PATH to be in the project work area backed by the Spectrum Scale parallel filesystem (aka. "scratch"), not project or user home file systems backed by NFS. This is because the work area has massive amounts of space, and because NFS (used for home) creates problems with stale .nfsX files and may not be the fastest when accessed from worker nodes.

The project work area has a 90 day file retention period, which means files are deleted after they haven't been accessed for this long. To refresh file access times and thus prevent the purge, run this every couple of months:

$ find PREFIX_PATH -execdir touch -a {} \;

The build can be done on a worker node (preferred), run this command on the login machine to submit the job:

$ casper-utils/jobs/olcf-summit/gpref.job.sh PREFIX_PATH casper-olcf-summit

Once the job starts (see bjobs), you can monitor the log with:

$ ls -t PREFIX_PATH/var/log/prefix/gpref-*.*.log | head -1 | xargs tail -f

Note that there are two parts to the job, each with its own log file, so after the first half completes, re-run the above command to start monitoring the second log.

The build job is submitted to Summit's killable queue, which is the only queue that can support single-node reservations of several hours. Jobs in this queue may be pre-empted by Summit's job scheduler (which terminates the job with a SIGTERM but automatically re-enqueues it). The build job is written such that is incremental, so it's ok if it is preempted and restarted.

Or, the build can also be done directly on the login macine:

    $ casper-utils/jobs/gpref.sh PREFIX_PATH casper-olcf-summit

Other Linux hosts

You might want to override the default for number of processors to use (the default is what nproc returns).

$ export NPROC=16

You may also want to override the default temporary build directory (the default is the first one that has enough space from the list: /tmp, /dev/shm):

$ export TMPDIR=/path/to/custom/temp/dir

Run the build script directly:

$ casper-utils/jobs/gpref.sh PREFIX_PATH casper-generic-ARCH

where the second argument is a Gentoo profile name where ARCH identifies the CPU family for which to optimize via the -march,-mtune compiler flags. For the generic unoptimized use amd64; for gpuk40 (Xeon E5-2670) use sandybridge; for supported cpu families see ebuilds/profiles/casper-generic-*).

Tips and notes

A useful oneliner for monitoring the log from the latest job invocation (especially useful when you have to re-invoke the job after fixing failures):

ls -t PREFIX_PATH/var/log/prefix/*.{out,err} | head -2 | xargs tail -f

The job logs are saved in PREFIX_PATH/var/log/prefix/, and logs from the bootstrap step are also in PREFIX_PATH/stage{1,2,3}.log.

Some notes:

• Prefix with all CASPER dependencies takes about 6 hours to build on average 16 core box with a normal disk file system. On USC HPCC the same exact job takes about 24-48 hours due to slow shared filesystem (despite the build working directories being in fast local /tmp directory).
• Due to imperfections in the build recipes for some libraries, the cpu family of the build host must be the same as that of the target host, i.e. if you want a prefix optimized for Sandy Bridge, you have to build it on a Sandy Bridge node. (This could be fixed by tracking down the offending packages and fixing each to not use target compiler flags for tools that need to be run on the build host.)
• Some packages optimize for the build host even if you did not request any optimization, so you can't build a generic unoptimized prefix on a new CPU family; use Sandy Bridge nodes to build generic prefixes that should mostly work on newer CPU families too.
• Builds of numerical libraries on AMD Opteron node appear to be broken when optimized for opteron CPU family (even with -march=native).
• By default the temporary work directory is removed automatically (including when job fails). To keep the dir, set KEEP_TMPDIR env var to a non-empty value. This is not useful when the build runs on a worker node (because not easy to get to the local file system of that node after job exists), as is the case on USC HPCC.

Enter the prefix

To build and run applications (or anything else) inside the prefix, it is in theory sufficient to invoke the application by its full path, however it is usually convenient to "enter" into the prefix, which adds the prefix binary PATHs to the PATH env var and does other setup.

Generic Linux host

To enter the prefix on a generic Linux host:

$ PREFIX_PATH/startprefix

USC HPCC

To enter the prefix on a USC HPCC host (login or worker with interactive shell):

$ PREFIX_PATH/ptools/pstart

To enqueue a job inside the prefix on a USC HPCC worker node:

$ PREFIX_PATH/ptools/psbatch CLUSTER[:PARTITION] ARCH[:GPU:GPU_COUNT] MAX_MEM_PER_TASK \
NUM_NODES NUM_TASKS_PER_NODE TIME_LIMIT command arg arg...

for example:

$ PREFIX_PATH/ptools/psbatch hpcc sandybridge:k20:1 all 1 1 00:10:00 python --version

The keyword 'all' for MAX_MEM_PER_TASK grants all memory on the node ("per task" does not apply anymore).

Test the prefix

Minimal smoke test

The minimal test is to enter the prefix shell (more details on entering the prefix are in a dedicated section below):

$ cd PREFIX_PATH
$ ./startprefix

A new shell should be opened, now try run a program within the prefix:

$  emerge --version
Portage 2.3.100 ...

If this smoke test was not successful, check the logs for any errors (see previous step for log file locations). Errors in later stages, may be fixed by entering the prefix shell and using portage (emerge) to diagnose. Then, the top-level job may be restarted (as described in previous step), and it should resume incrementally.

MPI hello-world test

First, build the mpitest application in the prefix:

$ cd casper-utils/exp/apps/mpitest
$ PREFIX_PATH/startprefix
$ make

Check that it was linked correctly against libraries strictly only within the prefix path:

$ ldd ./mpitest

Then, while your shell is still in the prefix, run the app with the following test scripts.

The test is scripted in test-mpi.sh, and instructions for running it are below. The script takes one optional argument that sets the number of nodes. By default only one node will be assumed, so the test will execute multiple ranks on a single node.

On a generic host

On a host (compatible with the ARCH for which the prefix was built):

$ cd casper-utils/exp/test-prefix
$ bash test-mpi.sh

On USC HPCC and Discovery

To launch a job on USC HPCC worker node, run this launcher script on the login node:

$ bash exp/test-prefix/usc/test-mpi.job.sh PREFIX_PATH CLUSTER ARCH

where PREFIX_PATH is the directory with the prefix, CLUSTER is the name of the compute cluster (e.g. discovery see Step 2 above), ARCH is the CPU Family (e.g. sandybridge see Step 2 above) with which the prefix is compatible and GPU is the GPU model (e.g. p100), run psfeat tool from casper-utils/bin/ to see the nodes and their GPU resources (do not include the gpu: prefix and do not include the :N count suffix).

These scripts will keep running (watching the log files) even after the job completes (they do not detect job completion), so you have to Ctrl-C when you see that the job has finished. To check success, check the end of the log for (but also check for expected output and lack of any errors):

Leaving Gentoo Prefix with exit status 0

Also, the script just watches the log file, after the job has launched, you can kill the script with Ctrl-C, and the job will keep running. You can then manually check the queue with squeue and monitor the job log files the paths to which where printed when the job was run (look for --output and --error arguments in the log).

On Argonne Theta

Enqueue a job for an in interactive mode:

qsub -A CASPER -n 2 -t 10 -q debug-cache-quad -I

In the job's shell on the "MOM" node, enter the prefix and run the script as you would on a generic host, see instructions above.

On Summit

Enqueue a job for an in interactive mode:

bsub -P csc436 -nnodes 2 -W 10 -q debug -Is /bin/bash

In the job's shell on the "batch" node, enter the prefix and run the script as you would on a generic host, see instructions above.

CUDA smoke test

First, check that the CUDA compiler is installed:

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver

Switch to a gcc version that's compatible with CUDA:

$ cuda-config -s
... 8.4
$ gcc-config -l
[1] ...-8.4
[2] ...-10.2
$ gcc-config 1
$ gcc --version
...8.4

There is a hello-world CUDA application in exp/apps/cuda: It has currently been tested only on a generic Linux host (with an Nvidia K40 GPU), not on any HPC clusters.

$ cd exp/apps/cuda
$ make
$ ./add_managed

Switch back to the latest gcc, or else other builds will fail:

$ gcc-config 2
$ gcc --version
...10.2

Known issues

• Only add_managed (unified memory) not add_copy works.

CFD application test: Cahn-Hilliard

The Cahn-Hilliard application in Firedrake can be invoked via a Makefile-based infrastructure, either directly in the shell or through launching a job.

First, enter the directory for the relevant CLUSTER:

$ cd exp/cahn-hilliard/CLUSTER

To invoke directly on the node, either a geneirc linux host, or, in interactive job (on Theta this would be the MOM node, and on Summit the batch node):

1. Launch the persistent PRRTE DVM on all the nodes in your job allocation:

    $ prte --daemonize
   

2. Invoke the application (you may change the parameters in the target name to execute for a different mesh size, ranks, etc.; application output and time info will be in dat/test-1/ directory):

    $ make IN=1 dat/test-1/ch_mesh-64_ranks-4_tpn-16_steps-5.csv
   

3. Terminate the DVM (you may invoke multiple jobs without restarting the DVM; the invocations may even be concurrent in theory, but this is broken in practice, so run only one job at a time):

    $ pterm
   

CFD application test: Lid-Driven cavity (old scripts)

There is a script that runs a CFD application in

• Firedrake FEM framework
• FEniCS (aka. DOLFIN) FEM framework [ currently disabled ]

The script cycles through several direct solvers to test them.

Instructions are similar to the MPI hello-world test:

$ cd casper-utils/exp/test-prefix
$ bash test-cfd.sh

The test-cfd.sh script takes one optional argument that enables the test on GPU (besides the non-GPU tests) when non-empty.

On USC HPCC and Discovery

For USC HPCC or Discovery clusters, there's a script that submits the test script as a job:

$ bash exp/test-prefix/usc/test-cfd.job.sh PREFIX_PATH CLUSTER ARCH:GPU

On Theta and Summit

You must set NOLOCAL environment variable so that the launch node (MOM on Theta, batch on Summit) is not used:

$ NOLOCAL=1 bash test-cfd.sh

CASPER compiler test

Enter the prefix, if you're not already in the prefix:

$ PREFIX_PATH/startprefix

Make sure you have the latest GCC version selected (otherwise you might get erros from the ld linker, which is used with Clang too):

$ gcc-config -l
[1] ...-8.4
[2] ...-10.2
$ gcc-config 2
$ gcc --version

Build the CASPER compiler:

$ cd casper-utils/compiler
$ mkdir build
$ cd build
$ CC=clang CXX=clang++ cmake ..
$ make -j6

SAR app

Build the SAR app:

$ cd casper-utils/exp/apps/casper/sar
$ mkdir build
$ cd build
$ CC=clang CXX=clang++ cmake ..

On Summit, a manual workaround is required to compensate for some unimplemented functionality:

$ cp ../tuned-params-summit.ini tuned-params.ini

Continue with the build:

$ make -j6

Run the SAR app (On Theta, must be on a worker node (not login, not MOM node)!):

$ make sarbp.run

Output image is in the current directory (build directory) output_image.png.

CFD app

On Theta, the following build (not run) steps must be done on a worker node, because CASPER compiler does not support cross-compiling for apps that use the Firedrake/UFL DSL yet. To get a shell on a worker node on Theta, first get an interactive session:

$ qsub -A CASPER -n 2 -t 30 -q debug-cache-quad -I --attrs enable_ssh=1

Then, from the MOM node, get the hostname of a worker node and login to it (if this hangs, you can also check qstat debug-cache-quad output for node names):

$ aprun -n 1 hostname
nidXXXXX
$ ssh nidXXXXX

Build the CFD app (On Theta, must do this on worker node! On Summit, must do tihs on the batch node!):

$ cd casper-utils/exp/apps/casper/cahnhilliard
$ mkdir build
$ cd build
$ CC=clang CXX=clang++ cmake ..
$ make -j6

Run the CFD app (On Theta, must do this on the MOM node! On Sumit, must do this on the batch node!):

$ RANKS=2 MAPBY=slot make ch.mpirun

Run without MPI (this might work on a generic Linux host, but not on HPC clusters, because MPI is imported within Python, and hence the app must be invoked via the mpi launcher):

$ make ch.run

You can add make VERBOSE=1 to the make command to see the invoked command line.

The output is in ch-sol.pvd file in the current directory (build directory), and can be viwed with ParaView GUI software.

Known Issues

The app is likely to fail on Theta with DIVERGED_LINEAR_SOLVE error. This is a problem related to floating point differences introduced by the Xeon Phi architecture, it is an open issue that needs resolving.

On Summit, runs with more than one rank hang in the mass task. This is a new open issue and needs to be debugged.

Running experiments

The above section on testing covers some experiments, but this section goes into more detail about the infrastructure available for invoking apps in order to collect measurements.

So far, not all experiments are using this infrastructure:

• The experiment that does use this infrastructure is the one for collecing scaling data for the CFD app implemented in Firedrake, it lives in exp/cahnilliard, and there is a subdirectory per platform, with a makefile that has useful functionality for submitting jobs described next.

• The experiments that do not, exist in exp/some-experiment subdiectories, and use ad-hoc scripts or are completely manual -- each such and has its own README with instructions.

This infrastructure allows to invoke apps (on various input sizes, with various rank counts) by specifying parameters on in a target name and invoking that target. The infrastructure also makes it easy to enqueue the target invocations into jobs on the supported HPC clusters.

The general idea is that you can time a particular run of an application with a command like:

$ cd casper-utils/exp/cahn-hilliard/CLUSTER
$ make dat/test-1/ch_mesh-64_ranks-2_tpn-16_trial-1.csv

This command would create this file, if it doesn't already exist. The underlying command for the app is specified in each Makefile and determines how this file is created and what it contains (e.g. timings), and what other files might be created along with it (e.g. .sol with solution data).

The path must start with two levels of subdirectories dat/test-1. The file name is a key-value map (e.g. {mesh=64, ranks=1, ....} above) and this map is used to construct the command by which the app is invoked. Any keys can be added, but some keys are required, depending on the Makefile for an individual experiment.

Some job parameters may be specified via make variables:

   $ make MAX_TIME=60 QUEUE=debug dat/test-1/ch_mesh-64_ranks-2_tpn-16.csv

There are two main usage modes:

A. wrapped into a job and enqueued to the job manager B. direct invocation in the current shell (enabled by IN=1 make var)

Mode A makes sense only on HPC clusters with a job manager. Mode B makes sense only on in an interactive job session on a "MOM" node on Theta cluster or a "batch" node on Summit, or on a generic host without a resource manager.

The infrastructure uses the persistent DVM (Distributed Virtual Machine) provided by the OpenMPI stack. This sets up the OpenMPI daemons on each node in the job allocation only onces, and allows jobs to be invoked without having to redo this setup for each job.

In Mode A, launching of the DVM is part of the job. To submit a job, you would just run:

$ make IN=1 dat/test-12/ch_mesh-64_ranks-2_tpn-16_trial-1.csv

But, in Mode B, you need to launch the DVM before invoking the target:

$ prte --daemonize
$ make IN=1 dat/test-12/ch_mesh-64_ranks-2_tpn-16_trial-1.csv
$ make IN=1 dat/test-12/ch_mesh-64_ranks-2_tpn-16_trial-2.csv
$ pterm

In Mode B, in theory, you can invoke multiple jobs in the same DVM in parallel (e.g. the two make commands above might be put into background with &), but this doesn't always work due to bugs in PRRTE/OpenMPI, so prefer to run one job at a time.

In Mode A, in theory, you can invoke multiple targets (one job per each) and let them run concurrently. However, due to bugs in PRRTE/OpenMPI, when a job in one DVM dies it takes down the other DVMs with it (observed on Summit). So, prefer to have only one job running at a time. This is very annoying, since HPC clusters do let users submit and run at least several jobs concurrently. But, the only way is to fix the bugs.

Some useful features (but beware of the bugs mentioned above!) (dry-mode make -n ... is useful to see what exactly each command woudl do):

• When multiple targets are specified in the make command, a separate job is enqued for each target (Mode A) or a separate prun is invoked for each target (Mode B), for example using shell globbing you can collect data for two rank counts:

    $ make dat/test-1/ch_mesh-64_ranks-{2,4}_tpn-16_trial-1.csv
  

• If you want to create one job that will invoke two prun instances, combine the targets with a + sign (can't use globbing anymore):

    $ make dat/test-1/ch_mesh-64_ranks-2_tpn-16_trial-1.csv+dat/test-1/ch_mesh-64_ranks-2_tpn-16_trial-1.csv
  

• Inside Makefile sets of targets can be defined using the target-set function so that they may be invoked by the name of the set (by default, all targets will be executed within one job):

    $ make dat/test-1/ch_test-2
  

• If you want to invoke a set of targets with one job per target, suffix with ^split:

    $ make dat/test-1/ch_test-2^split
  

For the commands that result in more than one job, the jobs are submitted concurrently to the job manager, unless QUEUE=debug is specified (which supports only one job at a time). If you want to force multiple jobs to run in sequence use make -j1. Currently there is not yet a way to force sequential execution pruns invoke bia a target set (see above).

This infrastructure is implemented in GNU Make by makefiles in prefix-tools/make/Makefile.job and prefix-tools/clusters/*/make/Makefile. Note that these makefiles are installed into the prefix by the app-portage/prefix-tools* packages, so if you want to change them, the process is to commit the changes, bump the package versions, and re-emerge the package (you could also temporarily edit the installed copies in $EPREFIX/usr/lib/prefix-tools, but be careful to not lose your edits upon package re-builds). The makefiles installed in the prefix are included by per-experiment makefiles that exist outside the prefix and define actual app commands (e.g. exp/cahn-hilliard/Makefile).

Tips for maintaining the Prefix

On a cluster, Portage (the package manager) tools (emerge, equery, ebuild eix) can be used from login machine if the prefix was built unoptimized for a particular architecture (i.e. the *-amd64 profile), and if some care is taken as described below.

On USC Discovery cluster, invoke portage tools through the wrapper p4port on the login machine and pport on the worker nodes, like so:

p4port emerge app-portage/prefix-tools

The wrapper takes care of setting the build directory to a temporary directory in tmpfs (for speed and for working around inability to build on BeegFS due to lacking hard link support), and of setting number of parallel processes to use appropriately.

Fetching sources

Whether the package manager is allowed to access the Internet to download tarballs or update VCS repositories (for packages built directly from repos), is controlled by EVCS_OFFLINE variable in$EPREFIX/etc/portage/make.conf.

By default, when prefix is built from scratch, online fetching starts disabled and is kept disabled until the very end of the build, at which point it is automatically turned on. It is turned on, because otherwise updating VCS packages, won't work intutitively. In offline mode, the VCS packages end up updated only up to the version stored in the package's repo clone in distfiles/, regardless of the snapshot timestamp in the package version.

For example, if you have version (snapshot) 20210101 of a VCS package A installed, and the recipe for A in the casper repo gets updated to snapshot 20210102, and you pull that update, and rebuild in offline mode, you'll end up rebuilding the old 20210101 snapshot, despite the version you'll see being the new one 20210102. This is because the way snapshot is implemented is using git rev-list --before=TIMESTAMP, so if the repo clone in distfiles/ is out of date, and emerge is disallowed to check online, it won't ever see new commits, and the before will evaluate to the same top commit, despite TIMESTAMP having increased.

Unfetchable tarballs

Online build hosts will automatically fetch tarballs from upstream (subject to broken links or server downtime), but some tarballs (listed below) cannot be fetched from upstream at all. So, even for online hosts, you have to obtain the tarball directory as described above.

For reference, the special source tarballs that cannot be fetched from upstream:

• portage-DATE.tar.bz2

• gentoo-DATE.tar.bz2

  The snapshot date is indicated in the job script. Cannot always be fetched from online, because upstream hosts only about 1 month.

• gentoo-headers{,-base}-KERNEL_VERSION.tar.gz

  Archives for the kernel version running on the host. The archives for 3.10 are available in distfiles/ (see below). To make an archive for other kernel versions, see the comments in the ebuild.

• tetgen (manually fill out form to get download link)

• ampi (manually fill out form to get download link)

• pyop2 (due to checksum changes in tarball autogenerated by GitHub?)

Note: It is important to use .tar.gz archive format (not .tar.bz2) for tarballs used during the stage 1 of bootstrap, because otherwise the build host needs to have bzip2 installed (an extra dependency). In the provided distfiles archive and directory, the format is tar.gz.

Updating

Not that each Prefix directory is standalone and does not reference the casper-utils directory that was used to build the prefix. Updating an existing Prefix and pulling new commits to casper-utilsrepo are unrelated operations.

First, you need to decide what kind of update you need: A. "passively" track changes made by others to the prefix: you want to pull latest changes to the package recipes in the casper repo, B. like A, but you know that the snapshot on top of which the overlay is based has been updated (you can tell by reading the git log of the casper overlay repo, and by checking the SNAPSHOT_DATE variable in casper-utils/jobs/gpref-sys.sh) C. "actively" change the prefix for others: you want to "rebase" the casper overlay repo onto a new version of gentoo repo.

(For Cases B and C only) Update the gentoo repo to a new snapshot

1. Update the tarballs in the distfiles/ directory (usually, the prefix will point to casper-utils/distfiles, but check DISTDIR in PREFIX/etc/portage/make.conf to make sure). See earlier sections in this document for where to get distfiles content from, then rsync from that location to get any files that may have been added:

    rsync -aP PATH_TO_DISTFILES/ distfiles/
   

2. Update the Gentoo repo to a known tested snapshot (not any upstream snapshot! upstream moves quickly and we only take updates very rarely). The latest tested snapshot is in jobs/gpref-sys.sh in the SNAPSHOT_DATE variable. Note: it is important to do step 1 above, because upstream servers don't store snapshots indefinitely. Update to the snapshot given by YYYYMMDD date with command:

    emerge-webrsync -v --keep --revert=YYYYMMDD
   

3. Rebuild the packages in casper overlay, by following the subsection below (this will usually require a lot of fixing and very careful attention to forked packages, which may need to be re-forked (re-based) if they had been updated upstream, or at least their version needs to be pinned)

4. Edit casper-utils/jobs/gpref-sys.sh and set SNAPSHOT_DATE variable to the new date, so that when a prefix is built from scratch, it is built on the new snapshot that you just updated to.

Update casper overlay repo

To update the casper overlay, which involves rebuilding any packages that have changed (and their downstream dependencies requering a rebuild), enter the overlay directory:

cd $EPREFIX/var/db/repos/casper

The first time, add the remote for convenience (note: the remote pointing into casper-utils/ebuilds is needed only for boostrap, and if you want to commit a new snapshot of ebuilds to casper-utils):

git remote add up [email protected]:ISI-apex/casper-ebuilds.git

Pull the changes:

git pull --ff-only up master

Generic Linux host

Run emerge to rebuild what needs to be rebuild given the updated ebuilds:

emerge --ask --update --changed-use --changed-slot --newuse --newrepo \
  --deep --with-bdeps=y @world

Don't continue if there are errors; it may be difficult to diagnose the problems, but it has to be done, otherwise the prefix will end up in a chaotic state. The only known errors that are ok, are:

• dev-python/randomgen: versions >=1.18 are masked, continue with installing 1.16.x

After the emerge finishes, exit and re-enter the prefix, for any updated config files that set environment variables to be re-applied.

On Summit

Follow the instructions for Generic Linux host above, on a login machine.

On USC Discovery

On a login machine, first enable online fetching (see subsection above), then tell portage to figure out what needs to be rebuilt and fetch the sources, replace ... with the arguments given for the generic Linux host):

p4port emerge ...

On USC Discovery, the build should be done on a worker node, not on the shared login machine where CPU usage limits are enforced. To schedule a job:

$EPREFIX/ptools/psbatch discovery ARCH 1 16 02:00:00 pport emerge ...

where ARCH is the architecture for which the prefix was built (or any for an unoptimized prefix). Partition may be specified by appending a suffix to the first argument like discovery:oneweek.

The job will remain in SLURM queue even if the psbatch script is killed (including even if the login node is rebooted).

If your prefix is not optimized for a particular architecture, and if when you do the first emerge -f on the login machine, you see that the packages that are being updated are quick to build (e.g. python packages without a compilation step at all, or small packages quick to compile), then you can just run the second emerge on the login machine too:

$EPREFIX/ptools/pstart p4port emerge ...

After the emerge finishes, exit and re-enter the prefix, for any updated config files that set environment variables to be re-applied.

Live package versions

Package recipes (.ebuilds) usually have versions corresponding to releases from upstrea. However, there are special recipes called "live" of the form package-VER.9999.ebuild, where VER is some components of the version. These build the tip of the master branch in the upstream repository.

To enable "live" versions a package (master tip from VCS, which will change every time you re-emerge the package), which may be useful for some packages provided by the casper overlay, add to $EPREFIX/etc/portage/package.accept_keywords packages in the following format:

app-portage/prefix-tools **

There are two (solved) issues with live ebuilds:

1. once installed it's not possible to tell which commit was installed (without having saved the build log, and then it is painful anyway)
2. once you tested a commit that you bulit via a live ebuild, there's no way to pin it into the package repository as the latest version to be installed by default.

Both of these are solved by the custom snapshot.eclass. To use it, ebuilds needs to inherit it and invoke it explicitly, so not all packages use it yet.

This eclass installs a file as part of the installation of the live ebuild into PREFIX/etc/snapshot/category/package with the commit hash in the file.

Also, this eclass supports easy way of pinning by creating a symlink to the live ebuild with a name that identifies the commit (the commit can only be identified by a date or date+time, in order for the version to be valid and well-ordered):

$ ln -s package-1.9999.ebuild package-1.0.YYYYMMDD.hhmmss_p.ebuild

The .hhmmss can be ommitted, in which case it is interpreted as time 00:00:00+0000 (UTC). The convension we use is _p suffix means that the version includes everything that is in release 1.0 and commits up to date YYYYMMDD, and the _pre suffix means that the version is the not yet released 2.0 with commits up to YYYYMMDD (i.e. when upstream has bumped the version in the repository but doesn't yet realeased a new tarball).

Note that using the timestamp uniquely identifies a commit:

git rev-list --first-parent --max-count 1 --before=YYYYMMDDThh:mm:ss+0000 HEAD

If you want to see the commit timestamps with a date that you'd use to set this snapshot timestamp, then use:

TZ=UTC git log --date=iso-local

Note that this snapshot functionality needs extra care in offline mode (EVCS_OFFLINE=1 in $PREFIX/etc/portage/make.conf): updating the timestamp in the package version is not enough in offline mode, you also need to make sure that the clone of the package's repo in distfiles/git-r3/ is update to date (the way to pull latest commits into that clone is to turn online mode and simply emerge the package; regardless of version). See section on Fetching sources in this README for more details. You normally don't have to worry about this because the prefix build script switches the prefix into online mode at the end of the build.

Forking package recipes

Whenever you need to make a change to the build recipe for a package you need to fork it from the gentoo repo into the casper repo. The procedure is:

1. cd $EPREFIX/var/db/repos

2. Check the the package hasn't already been forked:

    $ ls casper/category/package
   

3. Copy the package directory:

    $ cp -a gentoo/category/package casper/category/
   

4. Commit the unmodified copy, so that your later diffs are easy to see, recording the latest version in the commit message:

    $ git add casper/category/package
    $ git commit -m `category/package: LATEST VERSION (gentoo)`
   

5. Pin the package version, so that when the package gets updated in Gentoo, and you update the Gentoo snapshot in your prefix, the package manager doesn't replace your forked version with the updated version. Add to casper/profiles/casper/package.mask:

    >category/package-VERSION
   

However, pinning is not necessary in some cases. For example, if you change something about dependencies that enables the package to be pulled in successfully by the package manager, then your forked version will be the only one that would work regardless of updates available in gentoo repo, or if those updates are also fixed to start working then you want to pull them in. An example where pinning is not needed is when your change is just to relax the minimum version of the kernel in the dependency list.

Also, note that this pinning guideline has not been followed in the past, so a lot of currently forked packages are not pinned, but should be ideally.

Maintaining the fork

Once a package is forked into casper repo, you need to keep an eye on the updates that may happen to it in the gentoo repo, and whenever possible either unfork (by simply removing the package from the casper repo), or rebase your changes onto the updated version. Upstream moves quickly, but the gentoo repo in your prefix is only updated manually (see the update section earlier in this README), so the fork maintainance is something that needs attention only every time after you update the gentoo snapshot.

The rebase is a manual process:

1. Look at the list of patches that you've made to your fork, relative to the time when the copy was made from the gentoo:

   $ cd casper $ git log category/package

2. Then, copy the new version from gentoo repo (you may also need to copy patches in files/):

   $ cp gentoo/category/package/package-UPDATED_VERSION.ebuild
   casper/category/package/

3. Commit the unmodified updated version:

   $ git add casper/category/package/package-UPDATED_VERSION.ebuild $ git commit -m 'category/package: UPDATED_VERSION (gentoo)

4. Apply the patches you've noted in Step 1. This can be done either by manually editting the new .ebuild file, or exporting the patch with git format-patch HASH^..HEAD and applying it manually with patch -p1 package-UPDATED_VERSION.ebuild patch_file.patch.

Troubleshooting: errors about manifests

If emerge complains about a manifest file (the warning is non-fatal, but do not ignore it, do not procede with the operation if get warning), this means that whoever committed a change to a package ebuild recipe did not regenerate the manifest (bad). To regenerate the manifest, run a command like so for the respective package:

p4port ebuild app-portage/prefix-tools/prefix-tools-9999.ebuild manifest

Troubleshooting: emerge/ebuild hangs

If emerge is waiting for a lock to be released (it will say so), then it might be due to a previous emerge/ebuild not having completed cleanly, so try manually cleanup the lock file (after checking for any portage processes that might still be running):

rm $EPREFIX/var/db/.pkg.portage_lockfile

If emerge or ebuild hangs at the very end of a merge without any relevant output, it may be due to file system issues (especially when running on networked filesystems of various kinds). Try disable the sync that is called after merging (and has been observed to hang on NFS):

FEATURES="-merge-sync" emerge ...

Troubleshooting: strange errors after prefix was built from scratch

If you built a prefix from scratch, from a release tag (which means the prefix was tested on the systems mentioned in the tag name), and you're seeing strange failures when testing, then the first quick thing to try is to re-build the package that appears in the stack traces, if you get one: emerge category/package.

This is especially relevant if the build job experienced a failure (and you restarted it) -- for example, on Theta, occassional spurious filesystem errors have been observed to cause the build job to fail with a spurious "no space left on device" error. Simply restarting the job would let it continue, but the interruption damaged some package installation, so they cause errors later on. Simply re-emerging the package fixes these kinds of issues.

Evaluate CASPER Auto-tuner

Experiments are available for evaluating CASPER Auto-Tuner by first profiling the application performance and then using these measurements to train a performance prediction model (as a function of tunable parameter values) and evaluate the model on a test subset of the profiling data. The experiments for the following benchmarks are available:

• exp/tune-halide: tune parameters in the Halide schedule for the blur filter Halide pipeline, evaluated for both a CPU target and a GPU target,
• exp/tune-fenics: tune number of ranks and threads in MPI+OpenMP linear solver operator in context of a 2D Lid-Driven Cavity Finite Element problem implemented in FEniCS framework.

All the dependencies (LLVM, Halide, OpenCL, etc) of this experiment are already installed in the Prefix (see first chapter for building a prefix). The following commands can be run inside the prefix: to enter the Prefix see the previous chapter. If your system has all the dependencies of the right versions (with the right patches, etc) installed, then the following commands should work on your system too (without the Prefix), but nasty build issues might arise that were already solved in the Prefix.

See Makefile in each experiment's directory all available make targets, including the fine-grained targets for generated intermediate artifacts and for cleaning; what follows is a brief summary. Multiple targets may be passed to the same make command where appropriate.

Tune Halide schedule

To build and test the binaries for the Halide experiment:

$ cd casper-utils/exp/tune-halide
$ make
$ make test

To perform a quick smoke-test (does not produce a useful model), generate a small profiling dataset and train the respective prediction models:

$ make model-small

To generate the large dataset (takes >24 hours) and train the prediction models:

$ make model-large

See tune-halide/Makefile for fine-grained make targets for generated intermediate artifacts and for cleaning.

Tune MPI+OpenMP params in linear solver in FEniCS FEM program

The FEniCS benchmark application is in the following directory, and is written in Python so does not involve an explicit compile step

$ cd casper-utils/exp/tune-fenics

The perform a quick smoke-test (does not produce a useful model):

$ make model-small

For larger profiling dataset (takes several hours to collect), there are two options depending on how the datapoints along the input size are spread: linearly ({2,4,6,...}), or geometrically ({2,4,16,...}):

$ make model-lin-large
$ make model-log-large

Evaluation of the model

When performance prediction models finish training and testing, a few metrics will be printed:

• MAE: mean absolute error
• MSE: mean squared error
• MAPE(>x s): mean absolute percentage error, restricting to data instances whose runtime is greater than x seconds

Note: the reason to exclude data instances with small runtime is that sometimes MAPE gets extremely large (>1000%) when the actual runtime is small. In addition, we care more about tasks with a large runtime in the scheduling process.

• rho: Spearman's rank correlation coefficient (or Spearman's rho), which indicates if a higher ranked candidate through prediction also has a higher ranked true runtime. In short, how well does our model tell about the order of runtimes.

The speedup from best parameter values over poorest or over average parameter values can be plotted using scripts in exp/tune/, however at the moment, the raw data output by the previous steps needs to be manually post-processed to extract aggregated maximum/minimum/etc values.

[1] https://docs.github.com/en/github/managing-large-files/removing-files-from-git-large-file-storage#git-lfs-objects-in-your-repository [2] https://stackoverflow.com/questions/36626793/disable-git-lfs-for-a-remote