GNN4ITk as a Service

GNN4ITk as a Service#

The server#

Last update: 2024-09-04.

Reference: repo

THe following are instructions of running a Trition server for GNN4ITk using Perlmutter at NERSC.

Clone the tracking as a service repo and download the model.

git clone https://github.com/xju2/tracking-as-a-service
cd tracking-as-a-service
python scripts/download_model.py

Request a GPU node and write down the server ID, like nid00888, which will the server URL for the client.

srun -C "gpu&hbm80g" -q interactive -N 1 -G 1 -c 32 -t 4:00:00 -A m3443 --pty /bin/bash -l

Launch the server.

./scripts/start-tritonserver.sh

The client#

Last update: 2024-09-04.

We developed a Trition tool in Athena to facilitate the use of Trition. The tool is called TritonTool link to code, which implements the inference interface, IAthInferenceTool.

We created an example to show how to use the TritonTool in Athena. The example is located in InnerDetector/InDetGNNTracking/src/GNNTrackFinderTritonTool.h link to code. Following are the steps to compile the code, setup the environment, and run the example at Perlmutter. Please adjust the steps for other platforms.

Because the TritonClient is not yet installed in the Athena release (see the MR), we created a container that contains the TritonClient: docexoty/alma9-atlasos:triton-client-grpc1p62p3. For the same reason, The example is only available at the gnn_aas branch of my fork, https://gitlab.cern.ch/xju/athena.git.

Clone the code and Launch the container. See the ATLAS gittutorial documentation on how to use git. If not limited to disk space, I recommend to use full checkout.

git clone -b gnn_aas https://gitlab.cern.ch/xju/athena.git

shifter --image=docexoty/alma9-atlasos:triton-client-grpc1p62p3 --module=cvmfs bash
ATHENA_PATH="path-to-Athena"
cd $ATHENA_PATH
source /global/cfs/cdirs/atlas/scripts/setupATLAS.sh
setupATLAS

asetup Athena,main,here,latest
which athena

The command which athena should output the path to the athena executable.

Set up the Triton dependency environment (this can be skipped once the Trition Client is installed.)

LCG_VERSION=$(echo $ROOTSYS | awk -F / '{print $10}')
LCG_ROOT=${LCG_RELEASE_BASE}/${LCG_VERSION}
LCG_PATH=/cvmfs/sft.cern.ch/lcg/releases/${LCG_VERSION}


grpc_Install_Dir=${LCG_PATH}/grpc/1.48.0/${LCG_PLATFORM}
c_ares_ROOT=${LCG_PATH}/c_ares/1.17.1/${LCG_PLATFORM}
absl_ROOT=${LCG_PATH}/absl/20230802.1/${LCG_PLATFORM}
re2_ROOT=${LCG_PATH}/re2/2023.11.01/${LCG_PLATFORM}
tritonclient_ROOT=/home/atlas/install/tritonclient

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${tritonclient_ROOT}/lib64:$grpc_Install_Dir/lib:$c_ares_ROOT/lib64:$absl_ROOT/lib64:$re2_ROOT/lib64

Compile the relevant packages. Create a file named package_filters.txt with the following content:

+ InnerDetector/InDetGNNTracking
+ Control/AthOnnx/AthOnnxInterfaces
+ Control/AthOnnx/AthTritonComps
+ Tracking/TrkConfig
- .*

Then compile the packages in a build directory mkdir ../run_athena/build && cd ../run_athena/build:

cmake -DATLAS_PACKAGE_FILTER_FILE=../../athena/package_filters.txt \
  -DCMAKE_PREFIX_PATH="${tritonclient_ROOT}/lib64/cmake;${grpc_Install_Dir}/lib/cmake;${c_ares_ROOT}/lib64/cmake;${absl_ROOT}/lib64/cmake;${re2_ROOT}/lib64/cmake" ../../athena/Projects/WorkDir

make -j20

source x86_64-el9-gcc13-opt/setup.sh

Run the example in the run_athena/run directory. mkdir ../run && cd ../run. You may have to use rucio to download the RDO file.

RDO_FILENAME="inputData/RDO.37737772._000213.pool.root.1"

function clean_up() {
    rm InDetIdDict.xml PoolFileCatalog.xml hostnamelookup.tmp eventLoopHeartBeat.txt
}

function gnn4pixel() {
    clean_up
    export ATHENA_CORE_NUMBER=1
    if [ -z "$1" ]; then
        echo "Please provide the Triton server ID."
        return
    fi
    TritionServer=$1

    Reco_tf.py \
        --CA 'all:True' --autoConfiguration 'everything' \
        --conditionsTag 'all:OFLCOND-MC15c-SDR-14-05' \
        --geometryVersion 'all:ATLAS-P2-RUN4-03-00-00' \
        --multithreaded 'False' \
        --steering 'doRAWtoALL' \
        --digiSteeringConf 'StandardInTimeOnlyTruth' \
        --postInclude 'all:PyJobTransforms.UseFrontier' \
        --preInclude 'all:Campaigns.PhaseIIPileUp200' 'InDetConfig.ConfigurationHelpers.OnlyTrackingPreInclude' 'InDetGNNTracking.InDetGNNTrackingFlags.gnnTritonValidation' \
        --preExec 'flags.Tracking.GNN.usePixelHitsOnly = True; flags.Tracking.ITkGNNPass.doAmbiguityResolutionForGNN = False; flags.Tracking.GNN.Triton.model = "GNN4Pixel"' "flags.Tracking.GNN.Triton.url = \"${TritionServer}\"" \
        --inputRDOFile "${RDO_FILENAME}" \
        --outputAODFile 'test.aod.gnnTriton.root'  \
        --jobNumber '1' \
		--athenaopts='--loglevel=INFO' \
        --maxEvents -1 2>&1 | tee log.gnnTrition.txt
}

gnn4pixel nid00888

To evaluate the performance, we can run the IDPVM package see the link for more info.

runIDPVM.py --filesInput test.aod.gnnreader.debug.root --outputFile physval.root --doTightPrimary 

The output file physval.root contains the performance of the GNN tracking.

GNN4ITk as a Service

Contents

GNN4ITk as a Service#

The server#

The client#