Using BQSKit on a GPU Cluster

This guide explains how to use BQSKit with GPUs by leveraging the bqskit-qfactor-jax package. The bqskit-qfactor-jax package provides GPU implementation support for the QFactor and QFactor-Sample instantiation algorithms. For more detailed information and advanced configurations of the BQSKit runtime, refer to the BQSKit distribution guide.

We will guide you through the installation, setup, and execution process for BQSKit on a GPU cluster.

bqskit-qfactor-jax Package Installation

First, you will need to install bqskit-qfactor-jax. This can easily done by using pip

pip install bqskit-qfactor-jax

This command will also install all the dependencies including BQSKit and JAX with GPU support.

Optimizing a Circuit Using QFactor-Sample and the Gate Deletion Flow

This section explains how to optimize a quantum circuit using QFactor-Sample and the gate deletion flow.

First, we load the circuit to be optimized using the Circuit class.

from bqskit import Circuit

# Load a circuit from QASM
in_circuit = Circuit.from_file("circuit_to_opt.qasm")

Then we create the instantiator instance and set the number of multistarts to 32.

from qfactorjax.qfactor_sample_jax import QFactorSampleJax

num_multistarts = 32

qfactor_sample_gpu_instantiator = QFactorSampleJax()

instantiate_options = {
        'method': qfactor_sample_gpu_instantiator,
        'multistarts': num_multistarts,
    }

Next, generate the optimization flow.

from bqskit.passes import *

# Prepare the compilation passes
passes = [
    # Convert U3s to VU
    ToVariablePass(),

    # Split the circuit into partitions
    QuickPartitioner(partition_size),

    # For each partition perform scanning gate removal using QFactor jax
    ForEachBlockPass([
        ScanningGateRemovalPass(
            instantiate_options=instantiate_options,
        ),
    ]),

    # Combine the partitions back into a circuit
    UnfoldPass(),

    # Convert back the VariablueUnitaires into U3s
    ToU3Pass(),
]

Finally, use a compiler instance to execute the passes, and then print the statistics. If your system has more than a single GPU, then you should initiate a detached server and connect to it. A destailed explanation on how to setup BQSKit runtime is given in the next sections of the this guide.

from bqskit.compiler import Compiler

with Compiler(num_workers=1) as compiler:

    out_circuit = compiler.compile(in_circuit, passes)

    print(
            f'Circuit finished with gates: {out_circuit.gate_counts}, '
            f'while started with {in_circuit.gate_counts}',
        )

QFactor-JAX and QFactor-Sample-JAX Use Examples

For other usage examples, please refer to the examples directory in the bqskit-qfactor-jax package. There, you will find two Toffoli instantiation examples using QFactor and QFactor-Sample, as well as two different synthesis flows that also utilize these algorithms.

Setting Up a Multi-GPU Environment

To run BQSKit with multiple GPUs, you need to set up the BQSKit runtime properly. Each worker should be assigned to a specific GPU by leveraging NVIDIA’s CUDA_VISIBLE_DEVICES environment variable. Several workers can use the same GPU by utilizing NVIDIA’s MPS. You can set up the runtime on a single server ( or interactive node on a cluster) or using SBATCH on several nodes. You can find scripts to help you set up the runtime in this link.

You may configure the number of GPUs to use on each server and also the number of workers on each GPU. If you use too many workers on the same GPU, you will run out of memory and experience an out-of-memory exception. If you are using QFactor, you may use the following table as a starting configuration and adjust the number of workers according to your specific circuit, unitary size, and GPU performance. If you are using QFactor-Sample, start with a single worker and increase if the memory permits it. You can use the nvidia-smi command to check the GPU usage during execution; it specifies the utilization of the memory and the execution units.

Unitary Size

Workers per GPU

3,4

10

5

8

6

4

7

2

8 and more

1

Make sure that in your Python script, you are creating the compiler object with the appropriate IP address. When running on the same node as the server, you can use `localhost` as the IP address.

with Compiler('localhost') as compiler:
    out_circuit = compiler.compile(in_circuit, passes)

Single Server Multiple GPUs Setup

This section of the guide explains the main concepts in the single_server_env.sh script template and how to use it. The script creates a GPU-enabled BQSKit runtime and is easily configured for any system.

After you configure the template (replacing every <> with an appropriate value) run it, and then in a separate shell execute your python script that uses this runtime environment.

The environment script has the following parts:

  1. Variable configuration - choosing the number of GPUs to use, and the number of workers per GPU. Moreover, the scratch dir path is configured and later used for logging.

#!/bin/bash
hostname=$(uname -n)
unique_id=bqskit_${RANDOM}
amount_of_gpus=<Number of GPUS to use in the node>
amount_of_workers_per_gpu=<Number of workers per GPU>
total_amount_of_workers=$(($amount_of_gpus * $amount_of_workers_per_gpu))
scratch_dir=<temp_dir>

  1. Log file monitoring functions to monitor the startup of BQSKit managers and server.

wait_for_outgoing_thread_in_manager_log() {
    while [[ ! -f "$manager_log_file" ]]
    do
            sleep 0.5
    done

    while ! grep -q "Started outgoing thread." $manager_log_file; do
            sleep 1
    done
}

wait_for_server_to_connect(){
    while [[ ! -f "$server_log_file" ]]
    do
            sleep 0.5
    done

    while ! grep -q "Connected to manager" $server_log_file; do
            sleep 1
    done
}

  1. Creating the log directory, and deleting any old log files that conflict with the current run logs.

mkdir -p $scratch_dir/bqskit_logs

manager_log_file=$scratch_dir/bqskit_logs/manager_${unique_id}.log
server_log_file=$scratch_dir/bqskit_logs/server_${unique_id}.log

echo "Will start bqskit runtime with id $unique_id gpus = $amount_of_gpus and workers per gpu = $amount_of_workers_per_gpu"

# Clean old server and manager logs, if exists
rm -f $manager_log_file
rm -f $server_log_file

  1. Starting NVIDA MPS to allow efficient execution of multiple works on a single GPU.

echo "Starting MPS server"
nvidia-cuda-mps-control -d

  1. Starting the BQSKit manager, and indicating to wait for workers to connect to it. Waiting for the manager to start listening for a connection from a server. This is important as the server might timeout if the manager isn’t ready for the connection.

echo "starting BQSKit managers"

bqskit-manager -x -n$total_amount_of_workers -vvv &> $manager_log_file &
manager_pid=$!
wait_for_outgoing_thread_in_manager_log

  1. Starting the BQSKit server indicating that there is a single manager in the current server. Waiting until the server connects to the manager before continuing to start the workers.

echo "starting BQSKit server"
bqskit-server $hostname -vvv &>> $server_log_file &
server_pid=$!

wait_for_server_to_connect

  1. Starting the workers, each seeing only a specific GPU.

echo "Starting $total_amount_of_workers workers on $amount_of_gpus gpus"
for (( gpu_id=0; gpu_id<$amount_of_gpus; gpu_id++ ))
do
    XLA_PYTHON_CLIENT_PREALLOCATE=false CUDA_VISIBLE_DEVICES=$gpu_id bqskit-worker $amount_of_workers_per_gpu > $scratch_dir/bqskit_logs/workers_${SLURM_JOB_ID}_${hostname}_${gpu_id}.log &
done

  1. After all the processes have finished, stop the MPS server.

wait

echo "Stop MPS on $hostname"
echo quit | nvidia-cuda-mps-control

Multis-Server Multi-GPU Environment Setup

This section of the guide explains the main concepts in the init_multi_node_multi_gpu_slurm_run.sh run_workers_and_managers.sh scripts and how to use them. After configuring the scripts (updating every <>), place both of them in the same directory and initiate an SBATCH command. These scripts assume a SLURM environment but can be easily ported to other distribution systems.

sbatch init_multi_node_multi_gpu_slurm_run.sh

The rest of this section explains both of the scripts in detail.

init_multi_node_multi_gpu_slurm_run

This is a SLURM batch script for running a multi-node BQSKit task across multiple GPUs. It manages job submission, environment setup, launching the BQSKit server and workers on different nodes, and the execution of the main application.

  1. Job configuration and logging - this is a standard SLURM SBATCH header.

#!/bin/bash
#SBATCH --job-name=<job_name>
#SBATCH -C gpu
#SBATCH -q regular
#SBATCH -t <time_to_run>
#SBATCH -n <number_of_nodes>
#SBATCH --gpus=<total number of GPUs, not nodes>
#SBATCH --output=<full_path_to_log_file>

scratch_dir=<temp_dir>

  1. Shell environment setup - Please consult with your HPC system admin to choose the appropriate modules to load that will enable you to JAX on NVDIA’s GPUs. You may use NERSC’s Perlmutter documentation as a reference.

### load any modules needed and activate the conda environment
module load <module1>
module load <module2>
conda activate <conda-env-name>

  1. Starting the managers on all of the nodes using SLURM’s srun command, initiating the run_workers_and_managers.sh script across all nodes. The former handles starting managers and workers on each node.

echo "starting BQSKit managers on all nodes"
srun run_workers_and_managers.sh <number_of_gpus_per_node> <number_of_workers_per_gpu> &
managers_pid=$!

managers_started_file=$scratch_dir/managers_${SLURM_JOB_ID}_started
n=<number_of_nodes>

  1. Waiting for all managers to start, by tracking the number of lines in the log file, one created by each manager.

while [[ ! -f "$managers_started_file" ]]
do
        sleep 0.5
done

while [ "$(cat "$managers_started_file" | wc -l)" -lt "$n" ]; do
    sleep 1
done

  1. Starting the BQSKit server on the main node, and using SLURM’s SLURM_JOB_NODELIST environment variable to indicate the BQSKit server the hostnames of the managers.

echo "starting BQSKit server on the main node"
bqskit-server $(scontrol show hostnames "$SLURM_JOB_NODELIST" | tr '\n' ' ') &> $scratch_dir/bqskit_logs/server_${SLURM_JOB_ID}.log &
server_pid=$!

uname -a >> $scratch_dir/server_${SLURM_JOB_ID}_started

  1. Executing the main application that will connect to the BQSKit runtime

python <Your command>

  1. After the run is over, closing the BQSKit server.

echo "Killing the server"
kill -2 $server_pid

run_workers_and_managers.sh

This script is executed by each node to start the workers and managers on that specific node. It interacts with init_multi_node_multi_gpu_slurm_run.sh, the SBATCH script. If GPUs are required, the workers are spawnd seperatly from the manager, allowing for better configuratio of each worker.

The script starts with argument parsing and some variable configuration

#!/bin/bash

node_id=$(uname -n)
amount_of_gpus=$1
amount_of_workers_per_gpu=$2
total_amount_of_workers=$(($amount_of_gpus * $amount_of_workers_per_gpu))

scratch_dir=<temp_dir>
manager_log_file="$scratch_dir/bqskit_logs/manager_${SLURM_JOB_ID}_${node_id}.log"
server_started_file="$scratch_dir/server_${SLURM_JOB_ID}_started"
managers_started_file="$scratch_dir/managers_${SLURM_JOB_ID}_started"

touch $managers_started_file

Then the script declares a few utility methods.

wait_for_outgoing_thread_in_manager_log() {
    while ! grep -q "Started outgoing thread." $manager_log_file; do
        sleep 1
    done
    uname -a >> $managers_started_file
}

start_mps_servers() {
    echo "Starting MPS servers on node $node_id with CUDA $CUDA_VISIBLE_DEVICES"
    nvidia-cuda-mps-control -d
}

wait_for_bqskit_server() {
    i=0
    while [[ ! -f $server_started_file && $i -lt 10 ]]; do
        sleep 1
        i=$((i+1))
    done
}

start_workers() {
    echo "Starting $total_amount_of_workers workers on $amount_of_gpus gpus"
    for (( gpu_id=0; gpu_id<$amount_of_gpus; gpu_id++ )); do
        XLA_PYTHON_CLIENT_PREALLOCATE=false CUDA_VISIBLE_DEVICES=$gpu_id bqskit-worker $amount_of_workers_per_gpu &> $scratch_dir/bqskit_logs/workers_${SLURM_JOB_ID}_${node_id}_${gpu_id}.log &
    done
    wait
}

stop_mps_servers() {
    echo "Stop MPS servers on node $node_id"
    echo quit | nvidia-cuda-mps-control
}

Finally, the script checks if GPUs are not needed, it spawns the manager with its default behavior, else using the “-x” argument, it indicates to the manager to wait for connecting workers.

if [ $amount_of_gpus -eq 0 ]; then
    echo "Will run manager on node $node_id with n args of $amount_of_workers_per_gpu"
    bqskit-manager -n $amount_of_workers_per_gpu -v &> $manager_log_file
    echo "Manager finished on node $node_id"
else
    echo "Will run manager on node $node_id"
    bqskit-manager -x -n$total_amount_of_workers -vvv &> $manager_log_file &
    wait_for_outgoing_thread_in_manager_log
    start_mps_servers
    wait_for_bqskit_server
    start_workers
    echo "Manager and workers finished on node $node_id" >> $manager_log_file
    stop_mps_servers
fi