Running an application under Chimbuko¶

Online Analysis¶

In this section we detail how to run Chimbuko both offline (Chimbuko analysis performed after application has completed) or online (Chimbuko analysis performed in real-time). In the first sections below we will describe how to run Chimbuko assuming it has been installed directly onto the system and is available on all nodes. The same procedure can be used to run Chimbuko inside a (single) Docker container, which can be convenient for testing or for offline analysis. For online analysis it is often more convenient to run Chimbuko through Singularity containers, and we will describe how to do so in a separate section below.

Chimbuko is designed primarily for online analysis. In this mode an instance of the AD module is spawned for each instance of the application (e.g. for each MPI rank); a centralized parameter server aggregates global statistics and distributes information to the visualization module; and a centralized provenance database maintains detailed information regarding each anomaly.

It is expected that the user create a run script that performs three basic functions:

Launch Chimbuko services (pserver, provDB, visualization).
Launch the Anomaly Detection (AD) component
Launch the application

Because launching the AD and application requires explicit invocations of mpirun or the system equivalent (e.g. jsrun on Summit) tailored to the specific setup desired by the user, we are unfortunately unable to automate this entire process. In this section we will walk through the creation of a typical run script.

The first step is to load Chimbuko. If installed via Spack this can be accomplished simply by:

spack load chimbuko

(It may be necessary to source the setup_env.sh script to setup spack first, which by default installs in the spack/share/spack/ subdirectory of Spack’s install location.)

A configuration file is used to set various configuration variables that are used for launching Chimbuko services and used by the Anomaly detection algorithm in Chimbuko. A template is provided in the Chimbuko install directory $(spack location -i chimbuko-performance-analysis)/scripts/launch/chimbuko_config.sh, which the user should copy to their working directory and modify as necessary.

A number of variables in chimbuko_config.sh are marked <------------ ***SET ME*** and must be set appropriately:

viz_root : The root path of the Chimbuko visualization installation. For spack users this can be set to $(spack location -i chimbuko-visualization2) (default)
export C_FORCE_ROOT=1 : This variable is necessary when using Chimbuko in a Docker image, otherwise set to 0.
service_node_iface : The network interface upon which communication between the AD and the parameter server is performed. There are multiple ways to list the available interfaces, for example using ip link show. On Summit this command must be executed on a job node and not the login node. The interface ib0 should remain applicable for this machine.
provdb_engine : The libfabric provider used for the provenance database. Available fabrics can be found using fi_info (on Summit this must be executed on a job node, not the login node). For more details, cf. here.
provdb_domain : The network domain, used by verbs provider. It can be found using fi_info and looking for a verbs;ofi_rxm provider that has the FI_EP_RDM type. On Summit this must be performed on a compute node, but the default, mlx5_0 should remain applicable for this machine.
TAU_EXEC : This specifies how to execute tau_exec.
TAU_PYTHON : This specifies how to execute tau_python.
TAU_MAKEFILE : The Tau Makefile. For spack users this variable is set by Spack when loading Tau and this line can be commented out.
export EXE_NAME=<name> : This specifies the name of the executable (without full path). Replace <name> with an actual name of the application executable.
export FI_UNIVERSE_SIZE=<number> : Libfabric (used by the provDB) requires knowledge of how many clients are to be expected. For optimal performance this should be set equal or larger than the number of ranks.

A full list of variables along with their description is provided in the Appendix Section, and more guidance is also provided in the template script.

Next, in the run script, export the config script location and also source its contents into the script environment as follows:

export CHIMBUKO_CONFIG=chimbuko_config.sh
source ${CHIMBUKO_CONFIG}

In order to avoid having the Chimbuko services interfere with the running of the application, we typically run the services on a dedicated node. It is also necessary to place the ranks of the AD module on the same node as the corresponding rank of the application to avoid having to pipe the traces over the network. Unfortunately the means of setting this up will vary from system to system depending on the job scheduler (e.g. using a hostfile with mpirun).

We launch the services (provenance database, the visualization module, and the parameter server) using the run_services.sh script provided in the PerformanceAnalysis install path. The location of this script is contained in the chimbuko_services variable set by chimbuko_config.sh. By default this location is inferred automatically but it can be set manually in the chimbuko_config.sh. A description of commands used in the services script is provided in Appendix.

The services are run as a background task such that they continue to run throughout the job. The service script writes out several useful files in the chimbuko/vars subdirectory, including a file containing the full run command for the AD module, chimbuko/vars/chimbuko_ad_cmdline.var, assuming a basic (single application) workflow. We will use the existence of this file in a wait condition to ensure the services are ready before launching the online AD module and application instances:

<LAUNCH ON HEAD NODE> ${chimbuko_services} &
while [ ! -f chimbuko/vars/chimbuko_ad_cmdline.var ]; do sleep 1; done

Here <LAUNCH ON HEAD NODE> is the appropriate command to launch a process on the head node of the job.

We next launch the AD modules:

ad_cmd=$(cat chimbuko/vars/chimbuko_ad_cmdline.var)
eval "<LAUNCH N RANKS OF AD ON BODY NODES> ${ad_cmd} &"

Where <LAUNCH N RANKS OF AD ON BODY NODES> is the appropriate command to launch N ranks of the online AD module on the nodes other than the head node, where N is the same as the number of application ranks. Note that this command must ensure that AD rank i is launched on the same physical node as application rank i

For more complicated workflows the AD will need to be invoked differently. To aid the user the services script writes a second file, chimbuko/vars/chimbuko_ad_opts.var, which contains just the initial command line options for the AD. Examples of various setups can be found among the benchmark applications.

Finally, the application is instantiated using the following command:

<LAUNCH N RANKS OF APP ON BODY NODES> ${TAU_EXEC} ${EXE} ${EXE_CMDS}

Where <LAUNCH N RANKS OF APP ON BODY NODES> is the appropriate command to launch N ranks of the application on the nodes other than the head node, ${TAU_EXEC} is defined in chimbuko config file as described above, ${EXE} is the full path to the application’s executable and ${EXE_CMDS} specifies all input parameters that are required by the application executable.

Chimbuko can be run to perform offline analysis of the application by changing configuration for Tau’s ADIOS plugin as described here.

Running on Summit¶

In this section we provide specifics on launching on the Summit machine.

The following chimbuko_config.sh setup provides optimal network performance for the Chimbuko services:

service_node_iface : ib0
provdb_engine : verbs
provdb_domain : mlx5_0

Summit job components are started using IBM’s custom jsrun command, which supports two methods for completely specifying resource sets that we can use to perform the placement of the services and the AD and application ranks. Note that jsrun does not allow different resource sets to share the same hardware resources, hence we are forced to dedicate cores to the AD instances.

ERF files¶

(WARNING: As of 12/8/21 this feature is broken and the user should use the secondary URS method documented below until a fix is made available)

The default method for completely specifying resource sets is explicit resource files (ERF) which are supplied using the following command: jsrun --erf_input=${erf_file}. The file format is documented here.

For convenience we provide a script here to generate the ERF files, which is executed as follows:

./gen_erf_summit.pl ${n_nodes_total} ${n_mpi_ranks_per_node} ${n_cores_per_rank_main} ${n_gpus_per_rank_main} ${ncores_per_host_ad}

where

${n_nodes_total} is the total number of nodes used, including the one node that is dedicated to run the services.
${n_mpi_ranks_per_node} is the number of MPI ranks of the application (and AD) that will run on each node (must be a multiple of 2).
${n_cores_per_rank_main} and ${n_gpus_per_rank_main} specify the number of cores and GPUs, respectively, given to each rank of the application.
${ncores_per_host_ad} is the number of cores dedicated to the Chimbuko AD modules (must be a multiple of 2), with the application running on the remaining cores. Note that the total number of cores allocated per node must not exceed 42.

The script writes out three files: services.erf, ad.erf and main.erf. This allows us to fully specify the various <LAUNCH ...> commands from the previous section:

which can be used as follows:

<LAUNCH ON HEAD NODE> = jsrun --erf_input=services.erf
<LAUNCH N RANKS OF AD ON BODY NODES> = jsrun --erf_input=ad.erf
<LAUNCH N RANKS OF APP ON BODY NODES> = jsrun --erf_input=main.erf

URS files¶

The jsrun command also supports specifying resource sets using jsrun --use_resource=${urf_file} or jsrun -U ${urf_file} where documentation of the format of these “URS” files can be found here <https://www.ibm.com/docs/en/spectrum-lsf/10.1.0?topic=SSWRJV_10.1.0/jsm/jsrun.html>. Note that, unlike the ERF files, the URS files do not allow specification of resource sets at the level of hardware threads, only at the level of cores.

For convenience we provide a script here to generate the URS files, which is executed as follows:

./gen_urs_summit.pl ${n_nodes_total} ${n_mpi_ranks_per_node} ${n_cores_per_rank_main} ${n_gpus_per_rank_main}

where

${n_nodes_total} is the total number of nodes used, including the one node that is dedicated to run the services.
${n_mpi_ranks_per_node} is the number of MPI ranks of the application (and AD) that will run on each node (must be a multiple of 2).
${n_cores_per_rank_main} and ${n_gpus_per_rank_main} specify the number of cores and GPUs, respectively, given to each rank of the application.

Note that 1 core is assigned per rank of the AD, and so ${n_mpi_ranks_per_node} * (${n_cores_per_rank_main} + 1) should not exceed 42, the number of cores per node.

The script writes out three files: services.urs, ad.urs and main.urs. This allows us to fully specify the various <LAUNCH ...> commands from the previous section:

which can be used as follows:

<LAUNCH ON HEAD NODE> = jsrun -U services.urs
<LAUNCH N RANKS OF AD ON BODY NODES> = jsrun -U ad.urs
<LAUNCH N RANKS OF APP ON BODY NODES> = jsrun -U main.urs

Running on Spock¶

In this section we provide specifics on launching on the Spock machine.

Spock uses the slurm job management system. To control the explicit placement of the ranks we will use the --nodelist (-w) slurm option to specify the nodes associated with a resource set, the --nodes (-N) option to specify the number of nodes and the --overlap option to allow the AD and application resource sets to coexist on the same node. These options are documented here.

The --nodelist option requires the range of full hostnames of the nodes to be provided. In order to simplify the generation of this list we provide a script here that parses the SLURM_JOB_NODELIST environment variable and generates the nodelist for the services and application. To use:

service_node=$(./get_nodes.pl HEAD)
body_nodelist=$(./get_nodes.pl BODY)

We can now set the various <LAUNCH ..> commands in the section above:

<LAUNCH ON HEAD NODE> = srun -n 1 -c 64 --threads-per-core=1 -N 1-1 --ntasks-per-node=1 -w ${service_node}
<LAUNCH N RANKS OF AD ON BODY NODES> = srun -n ${N} -c 1 -N ${bodynodes}-${bodynodes} --ntasks-per-node=${n_mpi_ranks_per_node} -w ${body_nodelist} --overlap
<LAUNCH N RANKS OF APP ON BODY NODES> = srun -n ${N} -c ${ncores_per_rank_main} -N ${bodynodes}-${bodynodes} --ntasks-per-node=${n_mpi_ranks_per_node} -w ${body_nodelist} --gpus-per-task=${n_gpus_per_rank_main} --gpu-bind=closest --overlap

Where

${n_nodes_total} is the total number of nodes used, including the one node that is dedicated to run the services.
${bodynodes} is the number of nodes dedicated to the application and AD ranks (i.e. n_nodes_total-1)
${n_mpi_ranks_per_node} is the number of MPI ranks of the application (and AD) that will run on each node (must be a multiple of 2).
${n_cores_per_rank_main} and ${n_gpus_per_rank_main} specify the number of cores and GPUs, respectively, given to each rank of the application.

Note that we have assigned 1 core to each rank of the AD, and so ${n_mpi_ranks_per_node} * (${n_cores_per_rank_main} + 1) should not exceed 64, the number of available cores.

Scaling to large job sizes¶

Chimbuko supports runs with many thousands of MPI ranks. However achieving optimal performance of Chimbuko in this context can require some tuning of parameters in the chimbuko_config.sh. Firstly, ensure

FI_UNIVERSE_SIZE is set larger than the number of ranks.
Communication with the provDB (provdb_engine in the config) should be performed over the optimal OpenFabrics transport, i.e. verbs for Summit.

If the provenance database is taking a long time to drain its input buffers at the end of the job it typically means the database was overloaded and was not able to keep up with the volume of data. The provDB can be scaled in two ways:

provdb_nshards increases the number of independent database shards that can be written to in parallel.
provdb_ninstances controls the number of independent instances of the server exist

Increasing the number of shards should be the first option that is attempted. Each shard is managed by a separate Argobots execution stream and will run in parallel providing enough hardware threads are available to the services.

If increasing the number of shards is not sufficient, more provDB server instances can be run on further nodes, allowing indefinite scaling. However at present the built-in Chimbuko run_services.sh script can only support launching multiple provDB instances in the same resource set; for running servers on different resource sets the user must launch them manually with an appropriate job script. The provdb_ninstances variable must also be set to inform the other services components to coordinate with multiple server instances.

An example of running two different server instances on different nodes of Summit, for a run of our benchmark with 4032 ranks can be found in the scripts/summit/provdb_multiinstance subdirectory of the PerformanceAnalysis. The benchmark source can be found in the benchmark_suite/benchmark_provdb subdirectory.

Online analysis of an MPI application with a non-MPI installation of Chimbuko (advanced)¶

It is possible to use a non-MPI build of Chimbuko to analyze an MPI application. Indeed this is the only option for systems with job managers that do not allow tasks launched using different calls to mpirun (or equivalent) to occupy the same node.

There are two aspects to this that differ from a normal run of Chimbuko:

The instances of the online AD ‘driver’ must be launched alongside the ranks of the application. This can be achieved by creating a wrapper script that instantiates both the driver and the application, and launching this script using mpirun.
The driver instances must be manually provided with the application rank index to which they are to attach.

The assignment of a rank can be achieved using the -rank <rank> command line option of the driver component. Unfortunately this prevents the usage of the auto-generated AD run command that is output by the services script; instead the user must launch the driver manually in the wrapper script:

driver ${TAU_ADIOS2_ENGINE} ${TAU_ADIOS2_PATH} ${TAU_ADIOS2_FILE_PREFIX}-${EXE_NAME} ${ad_opts} -rank ${rank} 2>&1 | tee chimbuko/logs/ad.${rank}.log

Here the first four variables are set by sourcing the chimbuko_config.sh script that the user provides. The variable ad_opts should be assigned to the contents of the chimbuko/vars/chimbuko_ad_opts.var file that is generated by the services script (this variable contains the various commands required for the driver to attach to the services). Finally the rank must be obtained from the appropriate environment variable set by the mpirun variant, for example

rank=${OMPI_COMM_WORLD_RANK}

An example is provided for the func_multimodal mini-app in the Chimbuko PerformanceAnalysis repository:

benchmark_suite/func_multimodal/run_nompi.sh
benchmark_suite/func_multimodal/wrap_nompi.sh

Online analysis of a non-MPI application with a non-MPI installation of Chimbuko (advanced)¶

In the context of a non-MPI application, instances of the application must still be associated with an index within Chimbuko that allows for their discrimination. This proceeds much as in the previous section, but with a catch: by default Chimbuko assumes that the instance index passed in by the -rank <rank> option matches the rank index reflected by the trace data and the ADIOS trace filename produced by Tau. However for a non-MPI application, Tau assigns rank 0 to all instances. In order to communicate this to Chimbuko a second command line option must be used: -override_rank 0. Here the 0 tells Chimbuko that the input data is labeled as 0 in both the filename and the trace data. Chimbuko will then overwrite the rank index in the trace data to match that of its internal rank index to ensure that this new label is passed through the analysis. Note that the user must make sure that each application instance is assigned either a different TAU_ADIOS2_PATH or TAU_ADIOS2_FILE_PREFIX otherwise the trace data files will overwrite each other.

Examples¶

The “benchmark_suite” subdirectory of the source repository contains a number of examples of using Chimbuko including Makefiles and run scripts designed to allow them to be run in our Docker environments. Examples for CPU-only workflows include:

c_from_python (Python/C): A function with artificial anomalies that is part of a C library called from Python.
func_multimodal (C++): A function with multiple “modes” with different runtimes, and artificial anomalies introduced periodically.
mpi_comm_outlier (C++): An MPI application with anomalies introduced in the communication between two specific ranks.
mpi_comm_outlier (C++): An MPI application with anomalies introduced in the communication on a specific thread.
multiinstance_nompi (C++): An application with artificial anomalies for which multiple instances are run simultaneously without MPI. This example demonstrates how to manually specify the “rank” index to allow the data from the different instances to be correctly tagged.
python_hello (Python): An example of running a simple Python application with Chimbuko.
simple_workflow (C++): An example of a workflow with multiple components. This example demonstrates to how specify the “program index” to allow the data from different workflow components to be correctly tagged.

For GPU workflows we presently have examples only for Nvidia GPUS:

cupti_gpu_kernel_outlier (C++/CUDA): An example with an artificial anomaly introduced into a CUDA kernel. This example demonstrates how to compile and run with C++/CUDA.
cupti_gpu_kernel_outlier_multistream (C++/CUDA): A variant of the above but with the kernel executed simultaneously on multiple streams.

For convenience we provide docker images in which these examples can be run alongside the full Chimbuko stack. The CPU examples can be run as:

docker pull chimbuko/run_examples:latest
docker run --rm -it -p 5002:5002 --cap-add=SYS_PTRACE --security-opt seccomp=unconfined chimbuko/run_examples:latest

And connect to this visualization server at localhost:5002.

For the GPU examples the user must have access to a system with an installation of the NVidia CUDA driver and runtime compatible with CUDA 10.1 as well as a Docker installation configured to support the GPU. Internally we use the nvidia-docker tool to start the Docker images. To run,

docker pull chimbuko/run_examples:latest-gpu
nvidia-docker run -p 5002:5002 --cap-add=SYS_PTRACE --security-opt seccomp=unconfined chimbuko/run_examples:latest-gpu

And connect to this visualization server at localhost:5002.

We also provide DockerFiles and run scripts for two real-world scientific applications described below:

NWChem¶

NWChem (Northwest Computational Chemistry Package) is the US DOE’s premier massively parallel computational chemistry package, largely written in Fortran. We provide a Docker image demonstrating the coupling of an NWChem molecular dynamics simulation of the ethanol molecule with Chimbuko. To run the image:

docker pull chimbuko/run_nwchem:latest
docker run -p 5002:5002 --cap-add=SYS_PTRACE --security-opt seccomp=unconfined chimbuko/run_nwchem:latest

And connect to this visualization server at localhost:5002.

MOCU (ExaLearn)¶

The MOCU application is part of the ExaLearn project, a US DOE-funded organization whose role is to develop machine learning techniques for HPC environments. The MOCU (Mean Objective Cost of Uncertainty) code is a PyCuda GPU application for NVidia GPUs that computes uncertainty quantification values of the Kuramoto model of coupled oscillators, which is often used to model the behavior of chemical and biological systems as well as in neuroscience.

To run the image the user must have access to a system with an installation of the NVidia CUDA driver and runtime compatible with CUDA 10.1 as well as a Docker installation configured to support the GPU. Internally we use the nvidia-docker tool to start the Docker images. To run:

docker pull chimbuko/run_mocu:latest
nvidia-docker run -p 5002:5002 --cap-add=SYS_PTRACE --security-opt seccomp=unconfined chimbuko/run_mocu:latest

And connect to this visualization server at localhost:5002.

Interacting with the Provenance Database¶

The provenance database is stored in a single file, provdb.${SHARD}.unqlite in the job’s run directory. From this directory the user can interact with the provenance database via the visualization module. A more general command line interface to the database is also provided via the provdb_query tool that allows the user to execute arbitrary jx9 queries on the database.

The provdb_query tool has two modes of operation: filter and execute.

filter mode allows the user to provide a jx9 filter function that is applied to filter out entries in a particular collection. The result is displayed in JSON format and can be piped to disk. It can be used as follows:

provdb_query filter ${COLLECTION} ${QUERY}

Where the variables are as follows:

COLLECTION : One of the three collections in the database, anomalies, normalexecs, metadata (cf Provenance Database).
QUERY: The query, format described below.

The QUERY argument should be a jx9 function returning a bool and enclosed in quotation marks. It should be of the format

QUERY="function(\$entry){ return \$entry['some_field'] == ${SOME_VALUE}; }"

Alternatively the query can be set to “DUMP”, which will output all entries.

The function is applied sequentially to each element of the collection. Inside the function the entry is described by the variable $entry. Note that the backslash-dollar (\$) is necessary to prevent the shell from trying to expand the variable. Fields of $entry can be queried using the square-bracket notation with the field name inside. In the sketch above the field “some_field” is compared to a value ${SOME_VALUE} (here representing a numerical value or a value expanded by the shell, not a jx9 variable!).

Some examples:

Find every anomaly whose function contains the substring “Kokkos”:

provdb_query filter anomalies "function(\$a){ return substr_count(\$a['func'],'Kokkos') > 0; }"

Find all events that occured on a GPU:

provdb_query filter anomalies "function(\$a){ return \$a['is_gpu_event']; }"

If the pserver is connected to the provenance database, at the end of the run the aggregated function profile data and global averages of counters will be stored in a “global” database “provdb.global.unqlite”. This database can be queried using the filter-global mode, which like the above allows the user to provide a jx9 filter function that is applied to filter out entries in a particular collection. The result is displayed in JSON format and can be piped to disk. It can be used as follows:

provdb_query filter-global ${COLLECTION} ${QUERY}

Where the variables are as follows:

COLLECTION : One of the two collections in the database, func_stats, counter_stats.
QUERY: The query, format described below.

The formatting of the QUERY argument is described above.

execute mode allows running a complete jx9 script on the database as a whole, allowing for more complex queries that collect different outputs and span collections.

provdb_query execute ${CODE} ${VARIABLES} ${OPTIONS}

Where the variables are as follows:

CODE : The jx9 script
VARIABLES : a comma-separated list (without spaces) of the variables assigned by the script

The CODE argument is a complete jx9 script. As above, backslashes (’') must be placed before internal ‘$’ and ‘”’ characters to prevent shell expansion.

If the option -from_file is specified the ${CODE} variable above will be treated as a filename from which to obtain the script. Note that in this case the backslashes before the special characters are not necessary.