qstone

QStone

What

An utility to benchmark the quality of an HPC and Quantum Computers integration. The benchmark allows the definition of a set of users for which a set of configurable quantum applications will be randomly selected. Example of currently supported quantum applications: VQE, PyMatching, RB. Users can create custom applications and use them together with the core applications. QStone generates different portable files (.tar.gz), each supporting a different user and a different scheduler (currently supporting: slurm, lsf, bare_metal). During execution of the benchmark, QStone gathers information on each steps of the applications, allowing investigations on bottlenecks and/or resource constraints at the interface between Quantum and HPC. The benchmark under the following assumptions.

Why

Building an appropriate hardware/software infrastructure to support HPCQC requires loads of work. We believe we shoud use a data driven approach in which we measure, fix, measure again with every new version of Quantum Computers, software and HPC hardware.

Where

Currently supported platforms/architectures:

• MacOS: M1/M2 (Sequoia)
• Intel: x86 (Ubuntu)
• PowerPC: Power9 (RedHat)

Tested on Python [3.9-3.12]

How

Installation

pip install QStone

Execution

Run QStone using Command Line Interface

• Run the generator command

  qstone generate -i conf.json [--atomic/-a] [--scheduler/-s "slurm"/"jsrun"/"bare_metal"]

  Generates tar.gz files that contains all the scripts to run scheduler for each user. Currently supported schedulers: [baremetal, altair/fnc, slurm/schedmd]. QStone expects an input configuration describing the users to want to generate jobs for as well as the description of the quantum computer you are generating jobs for. The optional --atomic flag forces the generation of single step jobs, instead of the default repartition in three jobs (pre, run, post). The -s flag allows selecting the output scheduler, default is bare_metal.

  With config.json:

{
  "environment": { 
     "project_name": "proj_name",
     "connector": "NO_LINK",
     "qpu_ip_address": "0.0.0.0",
     "qpu_port": "55",
     "qpu_management": "LOCK",
     "timeouts" : {
         "http": 5,
         "lock": 4
     } 
  },
  "jobs": [
    {
      "type": "VQE",
      "qubit_min": 2,
      "qubit_max": 4,
      "walltime" : 10,
      "nthreads" : 4,
      "lsf/jsrun_opt": "-nnodes=4 "
    },
    {
      "type": "RB",
      "qubit_min": 2,
      "qubit_max": 4,
      "walltime" : 10,
      "nthreads" : 2,
      "slurm/schedmd_opt": "--cpus-per-task=4"
    }
  ],
  "users": [
    {
      "user": "user0",
      "weight": 1,
      "computations": {
        "VQE": 0.05,
        "RB": 0.94,
        "PyMatching": 0.01
      }
    }
  ]
}

For more information on the config.json format refer to the associated json schema. Among the other things, the config file allows setting different polling/querying/scheduling policies for handling shared access to the QPU. Only SLURM currently supports the high-performance (lowest-latency) "SCHEDULER" mode" please refer to SLURM for more information.

• Alternatively call the generator in script:

from qstone.generators import generator

def main():
    generator.generate_suite(config="config.json",
        num_calls=100,output_folder=".",atomic=False, scheduler="bare_metal")

if __name__ == "__main__":
     main()

• Run the run command to execute chosen scheduler/workload selecting an optional output folder

  qstone run -i scheduler.qstone.tar.gz [-o folder]

The optional -o allows selecting the output folder in which to run the benchmark instance.

• Alternatively may untar on your machine of choice and run as the selected user.

  • Run the jobs by executing sh qstone_suite/qstone.sh

  • Run the profiling tool to extract metrics of interest.

• Run the profile command providing the initial input configuration and output folder to run profiling tool on run information

  qstone profile --cfg conf.json --folder qstone_profile

• Run the profile command providing the initial input configuration and multiple output folders (in case of multi-user run) to run profiling tool on run information

  qstone profile --cfg conf.json --folder qstone_profile --folder qstone_profile2

Supported backend connectivities

• Local no-link runner
• gRPC
• Http
• Rigetti

Examples

• Getting started notebook
• How to add a new type of computation
• How to create a simple gateway

Contributing

Guidance on how to contribute and change logs