Adding a System

This guide is intended for those wanting to run a benchmark on a new system, such as vendors, system administrators, or application developers. It assumes a system specification does not already exist.

System specifications include details like:

  • How many CPUs are there per node on the system

  • What pre-installed MPI/GPU libraries are available

A system description is a system.py file, where Benchpark provides the API where you can represent a systems as an object and customize the description with command line arguments.

Identifying a Similar System

The easiest place to start when configuring a new system is to find the closest similar one that has an existing configuration already. Existing system configurations are listed in the table in System Specifications.

If you are running on a system with an accelerator, find an existing system with the same accelerator vendor, and then secondarily, if you can, match the actual accererator.

  1. accelerator.vendor

  2. accelerator.name

Once you have found an existing system with a similar accelerator or if you do not have an accelerator, match the following processor specs as closely as you can.

  1. processor.name

  2. processor.ISA

  3. processor.uArch

For example, if your system has an NVIDIA A100 GPU and an Intel x86 Icelake CPUs, a similar config would share the A100 GPU, and CPU architecture may or may not match. Or, if I do not have GPUs and instead have SapphireRapids CPUs, the closest match would be another system with x86_64, Xeon Platinum, SapphireRapids.

If there is not an exact match that is okay, steps for customizing are provided below.

Editing an Existing System to Match

If you want to add support for a new system you can add a class definition for that system in a separate directory in systems/. The best way is to copy the system.py for the most similar system identified above, and then paste it in a new directory and update it. For example the genericx86 system is defined in:

$benchpark
├── systems
   ├── genericx86
      ├── system.py

The System base class defined in /lib/benchpark/system.py is shown below, some or all of the functions can be overridden to define custom system behavior.

# Copyright 2023 Lawrence Livermore National Security, LLC and other
# Benchpark Project Developers. See the top-level COPYRIGHT file for details.
#
# SPDX-License-Identifier: Apache-2.0

import hashlib
import importlib.util
import os
import pathlib
import sys

import benchpark.paths
from benchpark.directives import ExperimentSystemBase
import benchpark.repo
from benchpark.runtime import RuntimeResources

from typing import Dict, Tuple
import benchpark.spec
import benchpark.variant

bootstrapper = RuntimeResources(benchpark.paths.benchpark_home)  # noqa
bootstrapper.bootstrap()  # noqa

import ramble.config as cfg  # noqa
import ramble.language.language_helpers  # noqa
import ramble.language.shared_language  # noqa
import spack.util.spack_yaml as syaml  # noqa

# We cannot import this the normal way because it from modern Spack
# and mixing modern Spack modules with ramble modules that depend on
# ancient Spack will cause errors. This module is safe to load as an
# individual because it is not used by Ramble
# The following code block implements the line
# import spack.schema.packages as packages_schema
schemas = {
    "spack.schema.packages": f"{bootstrapper.spack_location}/lib/spack/spack/schema/packages.py",
    "spack.schema.compilers": f"{bootstrapper.spack_location}/lib/spack/spack/schema/compilers.py",
}


def load_schema(schema_id, schema_path):
    schema_spec = importlib.util.spec_from_file_location(schema_id, schema_path)
    schema = importlib.util.module_from_spec(schema_spec)
    sys.modules[schema_id] = schema
    schema_spec.loader.exec_module(schema)
    return schema


packages_schema = load_schema(
    "spack.schema.packages",
    f"{bootstrapper.spack_location}/lib/spack/spack/schema/packages.py",
)
compilers_schema = load_schema(
    "spack.schema.compilers",
    f"{bootstrapper.spack_location}/lib/spack/spack/schema/compilers.py",
)


_repo_path = benchpark.repo.paths[benchpark.repo.ObjectTypes.systems]


def _hash_id(content_list):
    sha256_hash = hashlib.sha256()
    for x in content_list:
        sha256_hash.update(x.encode("utf-8"))
    return sha256_hash.hexdigest()


class System(ExperimentSystemBase):
    variants: Dict[
        str,
        Tuple["benchpark.variant.Variant", "benchpark.spec.ConcreteSystemSpec"],
    ]

    def __init__(self, spec):
        self.spec: "benchpark.spec.ConcreteSystemSpec" = spec
        super().__init__()

    def initialize(self):
        self.external_resources = None

        self.sys_cores_per_node = None
        self.sys_gpus_per_node = None
        self.sys_mem_per_node = None
        self.scheduler = None
        self.timeout = "120"
        self.queue = None

        self.required = ["sys_cores_per_node", "scheduler", "timeout"]

    def generate_description(self, output_dir):
        self.initialize()
        output_dir = pathlib.Path(output_dir)

        variables_yaml = output_dir / "variables.yaml"
        with open(variables_yaml, "w") as f:
            f.write(self.variables_yaml())

        self.external_packages(output_dir)
        self.compiler_description(output_dir)

        spec_hash = self.system_uid()

        system_id_path = output_dir / "system_id.yaml"
        with open(system_id_path, "w") as f:
            f.write(
                f"""\
system:
  name: {self.__class__.__name__}
  spec: {str(self.spec)}
  config-hash: {spec_hash}
"""
            )

    def system_uid(self):
        return _hash_id([str(self.spec)])

    def _merge_config_files(self, schema, selections, dst_path):
        data = cfg.read_config_file(selections[0], schema)
        for selection in selections[1:]:
            cfg.merge_yaml(data, cfg.read_config_file(selection, schema))

        with open(dst_path, "w") as outstream:
            syaml.dump_config(data, outstream)

    def external_pkg_configs(self):
        return None

    def compiler_configs(self):
        return None

    def external_packages(self, output_dir):
        selections = self.external_pkg_configs()
        if not selections:
            return

        aux = output_dir / "auxiliary_software_files"
        os.makedirs(aux, exist_ok=True)
        aux_packages = aux / "packages.yaml"

        self._merge_config_files(packages_schema.schema, selections, aux_packages)

    def compiler_description(self, output_dir):
        selections = self.compiler_configs()
        if not selections:
            return

        aux = output_dir / "auxiliary_software_files"
        os.makedirs(aux, exist_ok=True)
        aux_compilers = aux / "compilers.yaml"

        self._merge_config_files(compilers_schema.schema, selections, aux_compilers)

    def system_specific_variables(self):
        return {}

    def variables_yaml(self):
        for attr in self.required:
            if not getattr(self, attr, None):
                raise ValueError(f"Missing required info: {attr}")

        optionals = list()
        for opt in ["sys_gpus_per_node", "sys_mem_per_node", "queue"]:
            if getattr(self, opt, None):
                optionals.append(f"{opt}: {getattr(self, opt)}")

        system_specific = list()
        for k, v in self.system_specific_variables().items():
            system_specific.append(f"{k}: {v}")

        extra_variables = optionals + system_specific
        indent = " " * 2
        extras_as_cfg = ""
        if extra_variables:
            extras_as_cfg = f"\n{indent}".join(extra_variables)

        return f"""\
# SPDX-License-Identifier: Apache-2.0

variables:
  timeout: "{self.timeout}"
  scheduler: "{self.scheduler}"
  sys_cores_per_node: "{self.sys_cores_per_node}"
  {extras_as_cfg}
  max_request: "1000"  # n_ranks/n_nodes cannot exceed this
  n_ranks: '1000001'  # placeholder value
  n_nodes: '1000001'  # placeholder value
  batch_submit: "placeholder"
  mpi_command: "placeholder"
"""

The main driver for configuring a system is done by defining a subclass for that system in a systems/{SYSTEM}/system.py file, which inherits from the System base class.

As is, the generic_x86 system subclass should run on most x86_64 systems, but we mostly provide it as a starting point for modifying or testing. Potential common changes might be to edit the scheduler or number of cores per node, adding a GPU configuration, or adding other external compilers or packages.

To make these changes, we provided an example below, where we start with the generic_x86 system.py, and make a system called Modifiedx86.

1. First, make a copy of the system.py file in generic_x86 folder and move it into a new folder, e.g., systems/modified_x86/system.py. Then, update the class name to Modifiedx86.:

class Modifiedx86(System):

  1. Next, to match our new system, we change the scheduler to slurm and the number of cores per node to 48, and number of GPUs per node to 2.:

    # this sets basic attributes of our system
def initialize(self):
    super().initialize()
    self.scheduler = "slurm"
    self.sys_cores_per_node = "48"
    self.sys_gpus_per_node = "2"

3. Let’s say the new system’s GPUs are NVIDIA, we can add a variant that allows us to specify the version of CUDA we want to use, and the location of those CUDA installations on our system. We then add the spack package configuration for our CUDA installations into the systems/modified_x86/externals/cuda directory (examples in Siera and Tioga systems).

# import the variant feature at the top of your system.py
from benchpark.directives import variant

# this allows us to specify which cuda version we want as a command line parameter
variant(
    "cuda",
    default="11-8-0",
    values=("11-8-0", "10-1-243"),
    description="CUDA version",
)

# set this to pass to spack
def system_specific_variables(self):
    return {"cuda_arch": "70"}

# define the external package locations
def external_pkg_configs(self):
    externals = Modifiedx86.resource_location / "externals"

    cuda_ver = self.spec.variants["cuda"][0]

    selections = []
    if cuda_ver == "10-1-243":
        selections.append(externals / "cuda" / "00-version-10-1-243-packages.yaml")
    elif cuda_ver == "11-8-0":
        selections.append(externals / "cuda" / "01-version-11-8-0-packages.yaml")

    return selections

4. Next, add any of the packages that can be managed by spack, such as blas/cublas pointing to the correct version, this will generate the software configurations for spack (software.yaml). The actual version will be rendered by Ramble when it is built.

def sw_description(self):
      return """\
software:
  packages:
    default-compiler:
      pkg_spec: gcc
    compiler-gcc:
      pkg_spec: gcc
    default-mpi:
      pkg_spec: openmpi
    blas:
      pkg_spec: cublas@{default_cuda_version}
    cublas-cuda:
      pkg_spec: cublas@{default_cuda_version}
"""

5. The full system.py class for the modified_x86 system should now look like:

import pathlib

from benchpark.directives import variant
from benchpark.system import System

class Modifiedx86(System):

    variant(
        "cuda",
        default="11-8-0",
        values=("11-8-0", "10-1-243"),
        description="CUDA version",
    )

    def initialize(self):
        super().initialize()

        self.scheduler = "slurm"
        setattr(self, "sys_cores_per_node", 48)
        self.sys_gpus_per_node = "2"

    def generate_description(self, output_dir):
        super().generate_description(output_dir)

        sw_description = pathlib.Path(output_dir) / "software.yaml"

        with open(sw_description, "w") as f:
            f.write(self.sw_description())

    def system_specific_variables(self):
        return {"cuda_arch": "70"}

    def external_pkg_configs(self):
        externals = Modifiedx86.resource_location / "externals"

        cuda_ver = self.spec.variants["cuda"][0]

        selections = []
        if cuda_ver == "10-1-243":
            selections.append(externals / "cuda" / "00-version-10-1-243-packages.yaml")
        elif cuda_ver == "11-8-0":
            selections.append(externals / "cuda" / "01-version-11-8-0-packages.yaml")

        return selections

    def sw_description(self):
        """This is somewhat vestigial, and maybe deleted later. The experiments
        will fail if these variables are not defined though, so for now
        they are still generated (but with more-generic values).
        """
        return """\
  software:
    packages:
      default-compiler:
        pkg_spec: gcc
      compiler-gcc:
        pkg_spec: gcc
      default-mpi:
        pkg_spec: openmpi
      blas:
        pkg_spec: cublas@{default_cuda_version}
      cublas-cuda:
        pkg_spec: cublas@{default_cuda_version}
"""

Once the modified system subclass is written, run: benchpark system init --dest=modifiedx86-system modifiedx86

This will generate the required yaml configurations for your system and you can validate it works with a static experiment test.

Validating the System

To manually validate your new system, you should initialize it and run an existing experiment such as saxpy. For example:

benchpark system init --dest=modifiedx86-system modifiedx86
benchpark experiment init --dest=saxpy saxpy openmp=oui
benchpark setup ./saxpy ./modifiedx86-system workspace/

Then you can run the commands provided by the output, the experiments should be built and run successfully without any errors.

The following yaml files are examples of what is generated for the modified_x86 system from the example after it is initialized:

  1. system_id.yaml describes the system hardware, including the integrator (and the name of the product node or cluster type), the processor, (optionally) the accelerator, and the network; the information included here is what you will typically see recorded about the system on Top500.org. We intend to make the system definitions in Benchpark searchable, and will add a schema to enforce consistency; until then, please copy the file and fill out all of the fields without changing the keys. Also listed is the specific system the config was developed and tested on, as well as the known systems with the same hardware so that the users of those systems can find this system specification.

system:
  name: Modifiedx86
  spec: sysbuiltin.modifiedx86 cuda=11-8-0
  config-hash: 5310ebe8b2c841108e5da854c75dab931f5397a7fb41726902bb8a51ffb84a36

2. software.yaml defines default compiler and package names your package manager (Spack) should use to build the benchmarks on this system. software.yaml becomes the spack section in the Ramble configuration file.

software:
  packages:
    default-compiler:
      pkg_spec: 'gcc'
    compiler-gcc:
      pkg_spec: 'gcc'
    default-mpi:
      pkg_spec: 'openmpi'
    blas:
      pkg_spec: cublas@{default_cuda_version}
    cublas-cuda:
      pkg_spec: cublas@{default_cuda_version}

  1. variables.yaml defines system-specific launcher and job scheduler.

variables:
  timeout: "120"
  scheduler: "slurm"
  sys_cores_per_node: "48"
  sys_gpus_per_node: 2
  cuda_arch: 70
  max_request: "1000"  # n_ranks/n_nodes cannot exceed this
  n_ranks: '1000001'  # placeholder value
  n_nodes: '1000001'  # placeholder value
  batch_submit: "placeholder"
  mpi_command: "placeholder"

Once you can run an experiment successfully, and the yaml looks correct the new system has been validated and you can continue your Benchpark Workflow.