Installation and Usage

Installation

For x86 systems we provide pre-built docker images users can quickly start with their own TAU instrumented applications (See Chimbuko docker). Otherwise, we recommend that Chimbuko be installed via the Spack package manager. Below we provide instructions for installing Chimbuko on a typical Ubuntu desktop and also on the Summit computer. Some details on installing Chimbuko in absence of Spack can be found in the Appendix.

In all cases, the first step is to download and install Spack following the instructions here . Note that installing Spack requires Python.

We require Spack repositories for Chimbuko and for the Mochi stack:

git clone https://github.com/mochi-hpc/mochi-spack-packages.git
spack repo add mochi-spack-packages
git clone https://github.com/CODARcode/PerformanceAnalysis.git
spack repo add PerformanceAnalysis/spack/repo/chimbuko

Then follow the build instructions below.

Basic installation

A basic installation of Chimbuko can be achieved very easily:

spack install chimbuko^py-setuptools-scm+toml

Note that the dependency on py-setuptools-scm+toml resolves a dependency conflict likely resulting from a bug in Spack’s current dependency resolution.

A Dockerfile (instructions for building a Docker image) that installs Chimbuko on top of a basic Ubuntu 18.04 image following the above steps can be found here .

Once installed, the unit and integration tests can be run as:

cd $(spack location -i chimbuko-performance-analysis)/test
./run_all.sh

A note on libfabric providers

The Mercury library used for the provenance database requires a libfabric provider that supports the FI_EP_RDM endpoint. By default spack installs libfabric with the sockets, tcp and udp providers, of which only sockets supports this endpoint. However sockets is being deprecated as its performance is not as good as other dedicated providers. We recommend installing the rxm utility provider alongside tcp for most purposes, by appending the spack spec with ^libfabric fabrics=sockets,tcp,rxm.

For network hardware supporting the Linux Verbs API (such as Infiniband) the verbs provider (with rxm) may provide better performance. This can be added to the spec as, for example, ^libfabric fabrics=sockets,tcp,rxm,verbs.

Details of how to choose the libfabrics provider used by Mercury can be found here. For further information consider the Mercury documentation .

Integrating with system-installed MPI

Chimbuko requires an installation of MPI. While Spack can install MPI automatically as a dependency of Chimbuko, in most cases it is desirable to utilize the system installation. Instructions on configuring Spack to use external dependencies can be found here . The simplest approach in general is to edit (create) a packages.yaml in one of Spack’s search paths, e.g. ~/.spack/packages.yaml, with the following content:

packages:
  openmpi:
    buildable: false
    externals:
      - spec: openmpi@4.0.4
        prefix: /opt/openmpi4.0.4

Modified as necessary to point to your installation.

Summit

While the above instructions are sufficient for building Chimbuko on Summit, it is advantageous to take advantage of the pre-existing modules for many of the dependencies. For convenience we provide a Spack environment which can be used to install in a self-contained environment Chimbuko using various system libraries. To install, first download the Chimbuko and Mochi repositories:

git clone https://github.com/mochi-hpc/mochi-spack-packages.git
git clone https://github.com/CODARcode/PerformanceAnalysis.git

Copy the file spack/environments/summit.yaml from the PerformanceAnalysis git repository to a convenient location and edit the paths in the repos section to point to the paths at which you downloaded the repositories:

repos:
  - /autofs/nccs-svm1_home1/ckelly/install/mochi-spack-packages
  - /autofs/nccs-svm1_home1/ckelly/src/AD/PerformanceAnalysis/spack/repo/chimbuko

This environment uses the gcc/9.1.0 and cuda/11.1.0 modules, which must be loaded prior to installation and running:

module load gcc/9.1.0 cuda/11.2.0

Then simply create a new environment and install:

spack env create my_chimbuko_env summit.yaml
spack env activate my_chimbuko_env
spack install

Once installed, simply

spack env activate my_chimbuko_env
spack load tau chimbuko-performance-analysis chimbuko-visualization2

after loading the modules above.

Instrumenting an application with Tau

In this section we briefly describe how to instrument an application with Tau. For more details refer to Tau’s documentation.

The communication between Tau and Chimbuko’s AD module is performed using Tau’s ADIOS2 plugin. The communication is performed in batches known as IO steps or IO frames. During an IO step Tau collects data which is communicated to Chimbuko at the end of the step.

In order to enable the ADIOS2 plugin Tau must be compiled with the -adios compile option, pointing to the ADIOS2 install directory. The user must set a few environment variables as described here.

How Tau is used depends on the language in which the application is written. Below we describe how to instrument applications written in several common languages.

C++

In order to generate the full trace output needed by Chimbuko, the C++ application must be instrumented at compile time. The most reliable method is using compiler instrumentation:

  • Set environment variable TAU_OPTIONS=”-optShared -optRevert -optVerbose -optCompInst”

  • Replace C++ compiler with tau_cxx.sh

Note that this method of instrumentation typically introduces significant overheads into the running of the application as entry and exit events are generated for every function.

Tau also offers instrumentation via source-to-source transcription, which applies heuristics to determine which functions to instrument. This should be the preferred option, but unfortunately the support for many modern C++11 and later features is limited. To use source instrumentation:

  • Set environment variable TAU_OPTIONS=”-optShared -optRevert -optVerbose”

  • Replace C++ compiler with tau_cxx.sh

CUDA/C++

Calls to the GPU via CUDA are instrumented through CUDA’s CUPTI performance API at runtime. To enable this the user must execute their application under the tau_exec wrapper with the -cupti option. The user can optionally include the -um and -env options which track unified memory and add more environment counters (GPU temperature, fan speed, etc), respectively, which are associated with anomalies by Chimbuko and the information added to the provenance database entries. The Tau configuration is specified with the -T flag followed by a comma-separated list that should match the components that comprise the name of the corresponding Makefile, e.g. -T papi,mpi,pthread,cupti,pdt,adios2 corresponds to Makefile.tau-papi-mpi-pthread-cupti-pdt-adios2.

For example

tau_exec -cupti -env -um -T papi,mpi,pthread,cupti,pdt,adios2 ${APPLICATION} ${APPLICATION_OPTS}

The C++ components of the application should be compiled as in the previous section. In the special case of mixed C++/CUDA code, for which the user desires to instrument also the C++ component, the CUDA compiler first separates the CUDA and C++ code and passes the components to their corresponding compilers. We must therefore specify to the CUDA compiler that it should use the tau_cxx.sh compiler wrapper as its C++ compiler, thus:

For example

nvcc -ccbin tau_cxx.sh -x cu code.cc

Here the -x cu option ensures that the compiler treats the file as CUDA/C++ and ignores the extension. Note that options passed to the C++ compiler should be prefixed with -Xcompiler (for more information see the CUDA compiler documentation).

Python

As an interpreted language, Python applications must be wrapped by Tau’s Python wrapper, tau_python, in order to be instrumented. The -adios2_trace option enables the tracing required by Chimbuko. Python support defaults to python2 unless the interpreter is specified.

For example, for a python3 application:

tau_python -vv -tau-python-interpreter=python3 -adios2_trace ${APPLICATION}.py

Note that the ADIOS2 filename required by Chimbuko will be set not to the application name but to the name of the python interpreter, e.g. for TAU_ADIOS2_FILENAME=tau-metrics and using python3.6, the filename will be “tau-metrics-python3.6”.

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.

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 as follows:

export CHIMBUKO_CONFIG=chimbuko_config.sh

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).

In this section we will concentrate on the Summit supercomputer, where the process is made more difficult by the restrictions on sharing resources between resource sets which forces us to dedicate cores to the AD instances. To achieve the job placement the first step is to generate explicit resource files (ERF) for the head node, AD, and main programs. For convenience we provide a script here to generate the ERF files. It generates three ERF files (main.erf, ad.erf, services.erf) which are later used as input ERF files to instantiate Chimbuko services using jsrun command. This can be achieved by running the following script

./gen_erf.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.

More details on ERF can be found here.

In the next step the Chimbuko services are launched by running the run_services.sh script using jsrun command as following:

jsrun --erf_input=services.erf ${chimbuko_services} &
while [ ! -f chimbuko/vars/chimbuko_ad_cmdline.var ]; do sleep 1; done

Here –erf_input=services.erf launches the services using the the services ERF generated in the previous step. ${chimbuko_services} is the path to a script which specifies commands to launch the provenance database, the visualization module, and the parameter server. This variable is set by the configuration script. A description of commands used in the services script is provided in Appendix.

The while loop after the jsrun command is used to wait until the services have progressed to a stage at which connection is possible. At this point the script generates a command file (chimbuko/vars/chimbuko_ad_cmdline.var) which provides the to launch the Chimbuko’s anomaly detection driver program, assuming a basic (single component) workflow. This can be invoked as follows:

ad_cmd=$(cat chimbuko/vars/chimbuko_ad_cmdline.var)
eval "jsrun --erf_input=ad.erf -e prepended ${ad_cmd} &"

Here the ad.erf file is used as –erf_input for the jsrun command.

For more complicated workflows the AD will need to be invoked differently. To aid the user we write 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 instantiated using the following command:

jsrun --erf_input=main.erf -e prepended ${TAU_EXEC} ${EXE} ${EXE_CMDS}

Here we use the third ERF (main.erf) which was generated in the previous step. ${TAU_EXEC} is defined in chimbuko config file as described in previous step. ${EXE} is the full path to the application’s executable. ${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.

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.