Add Openmpi using new libfabric to stackinator options

docs/recipes.md

@@ -144,10 +144,12 @@ For example, in the recipe below, only `netcdf-fortran` will be built with the `
 
 ### MPI
 
-Stackinator can configure cray-mpich (CUDA, ROCM, or non-GPU aware) on a per-environment basis, by setting the `mpi` field in an environment.
+Stackinator can configure cray-mpich (CUDA, ROCM, or non-GPU aware) or OpenMPI (with or without CUDA) on a per-environment basis, by setting the `mpi` field in an environment.
 
 !!! note
-    Future versions of Stackinator will support OpenMPI, MPICH and MVAPICH when (and if) they develop robust support for HPE SlingShot 11 interconnect.
+    Future versions of Stackinator will fully support OpenMPI, MPICH and MVAPICH when (and if) they develop robust support for HPE SlingShot 11 interconnect.
+
+    Current OpenMPI support has been tested lightly and is not guaranteed to be production ready - only OpenMPI@5.x.x is supported (default is @5.0.6 at the time of writing - 2025.03.04) - CUDA is supported, ROCM has not yet been tested.
 
 If the `mpi` field is not set, or is set to `null`, MPI will not be configured in an environment:
 ```yaml title="environments.yaml: no MPI"
@@ -157,12 +159,18 @@ serial-env:
 ```
 
 To configure MPI without GPU support, set the `spec` field with an optional version:
-```yaml title="environments.yaml: MPI without GPU support"
+```yaml title="environments.yaml: Cray MPICH without GPU support"
 host-env:
   mpi:
     spec: cray-mpich@8.1.23
   # ...
 ```
+```yaml title="environments.yaml: OpenMPI without GPU support"
+host-env:
+  mpi:
+    spec: openmpi
+  # ...
+```
 
 GPU-aware MPI can be configured by setting the optional `gpu` field to specify whether to support `cuda` or `rocm` GPUs:
 ```yaml title="environments.yaml: GPU aware MPI"
@@ -176,7 +184,66 @@ rocm-env:
     spec: cray-mpich
     gpu: rocm
   # ...
+ompi-cuda-env:
+  mpi:
+    spec: openmpi
+    gpu: cuda
+  # ...
 ```
+#### Experimental libfabric 2.x support with cray-mpich
+
+HPE has open-sourced the libfabric/cxi provider (and related drivers) and these can be built into cray-mpich by adding a dependency to a newer version of libfabric.
+The system default is libfabric@1.15.2 - which can be changed by adding a `depends` field to the yaml. A non default version (newer than 1.15.2) will trigger a build of libfabric using libcxi, cxi-driver, cassini-headers).
+
+This syntax appplies to both mpich and openmpi builds.
+```yaml title="environments1.yaml: cray-mpich using new libfabric/cxi stack"
+mpich-cxi-env:
+  mpi:
+    spec: cray-mpich
+    gpu: cuda
+    depends: [libfabric@main]
+    # on release of libfabric@2, we recommended using @2 in preference to @main ...
+```
+```yaml title="environments2.yaml: openmpi using new libfabric/cxi stack"
+openmpi-cxi-env:
+  mpi:
+    spec: openmpi
+    gpu: cuda
+    depends: [libfabric@main]
+    # on release of libfabric@2, we recommended using @2 in preference to @main ...
+```
+!!! note
+    Currently the performance of OpenMPI on Alps clusters might not be optimal and work is ongoing to fine tune it especially for intra-node performance.
+
+#### Custom MPI builds
+
+If an experimental version of OpenMPI is required, the yaml syntax supports additional options to enable this. the `xspec` tag may be used to supply extra spack `spec` options.
+To illustrate usage, consider this example: a build of openmpi using a particular git branch named `mpi-continue-5.0.6` which in this case has a variant `+continuations` that is not available on the `main` or released branches.
+The `xspec` tag allows the user to supply arbitrary spack variants and options that replace the defaults used by stackinator in its absence.
+```yaml title="custom-openmpi.yaml: custom build of openmpi"
+openmpi-custom-env:
+  mpi:
+    spec: openmpi@git.mpi-continue-5.0.6=main
+    xspec: +continuations +internal-pmix fabrics=cma,ofi,xpmem schedulers=slurm +cray-xpmem
+    gpu: cuda
+    depends: ["libfabric@main"]
+```
+In this example, we must tell spack to fetch our custom git branch from a repo that is different from the default openmpi github version, by adding the following to our recipe `packages.yaml`
+```yaml title="custom-packages.yaml: custom repo for openmpi"
+  # mpi-continue-5.0.6 branch available from here
+  openmpi:
+    package_attributes:
+      git: https://github.com/your-username/ompi
+```
+
+!!! note
+    To build using a specific git commit, use the syntax
+    ```
+    spec: openmpi@git.9a4079916dd13d4190fe224102b57757789c13da=main
+    ```
+It is therefore possible to build arbitrary versions of MPI using custom options/branches etc using these combinations of settings.
+
+### Version info
 
 !!! alps
 
@@ -207,7 +274,7 @@ The list of software packages to install is configured in the `spec:` field of a
 The `deprecated: ` field controls if Spack should consider versions marked as deprecated, and can be set to `true` or `false` (for considering or not considering deprecated versions, respectively).
 
 The `unify:` field controls the Spack concretiser, and can be set to three values `true`, `false` or `when_possible`.
-The 
+The
 
 ```yaml
 cuda-env:

stackinator/builder.py

@@ -353,21 +353,77 @@ def generate(self, recipe):
             with dst.open("w") as fid:
                 fid.write(cache.generate_mirrors_yaml(recipe.mirror))
 
+        packages_data = {}
+        # append network packages to packages.yaml
+        network_config_path = system_config_path / "network.yaml"
+        if not network_config_path.is_file():
+            raise FileNotFoundError(f"The network configuration file '{network_config_path}' does not exist")
+        with network_config_path.open() as fid:
+            network_config = yaml.load(fid, Loader=yaml.Loader)
+            packages_data.update(network_config.get("packages", {}))
+
         # append recipe packages to packages.yaml
         if recipe.packages:
             system_packages = system_config_path / "packages.yaml"
-            packages_data = {}
             if system_packages.is_file():
                 # load system yaml
                 with system_packages.open() as fid:
                     raw = yaml.load(fid, Loader=yaml.Loader)
-                    packages_data = raw["packages"]
+                    packages_data.update(raw["packages"])
             packages_data.update(recipe.packages["packages"])
             packages_yaml = yaml.dump({"packages": packages_data})
             packages_path = config_path / "packages.yaml"
             with packages_path.open("w") as fid:
                 fid.write(packages_yaml)
 
+        # validate the recipe mpi selection
+        for name, config in recipe.environments.items():
+            # config[mpi] holds the user specified settings in environment.yaml
+            if config["mpi"]:
+                mpi = config["mpi"]
+                user_mpi_spec = mpi["spec"]
+                user_mpi_gpu = mpi["gpu"]
+                user_mpi_xspec = mpi["xspec"] if "xspec" in mpi else None
+                user_mpi_deps = mpi["depends"] if "depends" in mpi else None
+                self._logger.debug(
+                    f"User MPI selection: spec={user_mpi_spec}, gpu={user_mpi_gpu}, "
+                    f"xspec={user_mpi_xspec}, deps={user_mpi_deps}"
+                )
+
+                if user_mpi_spec:
+                    try:
+                        mpi_impl, mpi_ver = user_mpi_spec.strip().split(sep="@", maxsplit=1)
+                    except ValueError:
+                        mpi_impl = user_mpi_spec.strip()
+                        mpi_ver = None
+
+                    # network_config holds the system specified settings in cluster config / network.yaml
+                    if mpi_impl in network_config["mpi_supported"]:
+                        default_ver = network_config[mpi_impl]["version"]
+                        default_spec = network_config[mpi_impl]["spec"] if "spec" in network_config[mpi_impl] else ""
+                        default_deps = (
+                            network_config[mpi_impl]["depends"] if "depends" in network_config[mpi_impl] else ""
+                        )
+                        self._logger.debug(
+                            f"System MPI selection: spec={mpi_impl}@{default_ver} {default_spec}, deps={default_deps}"
+                        )
+
+                        # select users versions or the default versions if user did not specify
+                        version_opt = f"@{mpi_ver or default_ver}"
+                        # create full spec based on user provided spec or default spec
+                        spec_opt = f"{mpi_impl}{version_opt} {user_mpi_xspec or default_spec}"
+                        if user_mpi_gpu and user_mpi_gpu not in spec_opt:
+                            spec_opt = f"{spec_opt} +{user_mpi_gpu}"
+                        deps_opt = user_mpi_deps or default_deps
+                        for dep in deps_opt:
+                            spec_opt = f"{spec_opt} ^{dep}"
+                        self._logger.debug(f"Final  MPI selection: spec={spec_opt}")
+
+                        recipe.environments[name]["specs"].append(spec_opt)
+                    else:
+                        # TODO: Create a custom exception type
+                        raise Exception(f"Unsupported mpi: {mpi_impl}")
+
         # Add custom spack package recipes, configured via Spack repos.
         # Step 1: copy Spack repos to store_path where they will be used to
         #         build the stack, and then be part of the upstream provided

stackinator/recipe.py

@@ -8,15 +8,6 @@
 
 
 class Recipe:
-    valid_mpi_specs = {
-        "cray-mpich": (None, None),
-        "mpich": ("4.1", "device=ch4 netmod=ofi +slurm"),
-        "mvapich2": (
-            "3.0a",
-            "+xpmem fabrics=ch4ofi ch4_max_vcis=4 process_managers=slurm",
-        ),
-        "openmpi": ("5", "+internal-pmix +legacylaunchers +orterunprefix fabrics=cma,ofi,xpmem schedulers=slurm"),
-    }
 
     @property
     def path(self):
@@ -280,43 +271,19 @@ def generate_environment_specs(self, raw):
         for _, config in environments.items():
             config["exclude_from_cache"] = ["cuda", "nvhpc", "perl"]
 
-        # check the environment descriptions and ammend where features are missing
+        # check the environment descriptions and amend where features are missing
         for name, config in environments.items():
             if ("specs" not in config) or (config["specs"] is None):
                 environments[name]["specs"] = []
 
             if "mpi" not in config:
-                environments[name]["mpi"] = {"spec": None, "gpu": None}
+                environments[name]["mpi"] = {"spec": None, "gpu": None, "network": {"spec": None}}
 
-        # complete configuration of MPI in each environment
-        for name, config in environments.items():
-            if config["mpi"]:
-                mpi = config["mpi"]
-                mpi_spec = mpi["spec"]
-                mpi_gpu = mpi["gpu"]
-                if mpi_spec:
-                    try:
-                        mpi_impl, mpi_ver = mpi_spec.strip().split(sep="@", maxsplit=1)
-                    except ValueError:
-                        mpi_impl = mpi_spec.strip()
-                        mpi_ver = None
-
-                    if mpi_impl in Recipe.valid_mpi_specs:
-                        default_ver, options = Recipe.valid_mpi_specs[mpi_impl]
-                        if mpi_ver:
-                            version_opt = f"@{mpi_ver}"
-                        else:
-                            version_opt = f"@{default_ver}" if default_ver else ""
-
-                        spec = f"{mpi_impl}{version_opt} {options or ''}".strip()
-
-                        if mpi_gpu:
-                            spec = f"{spec} +{mpi_gpu}"
-
-                        environments[name]["specs"].append(spec)
-                    else:
-                        # TODO: Create a custom exception type
-                        raise Exception(f"Unsupported mpi: {mpi_impl}")
+            if "network" not in config["mpi"]:
+                environments[name]["mpi"]["network"] = {"spec": ""}
+
+        # we have not loaded the system configs yet, so mpi information will be generated
+        # during the builder phase. We will validate the mpi information then.
 
         # set constraints that ensure the the main compiler is always used to build packages
         # that do not explicitly request a compiler.

stackinator/schema/environments.json

@@ -47,8 +47,14 @@
                                  "gpu":  {
                                      "enum": ["cuda", "rocm", null, false],
                                      "default": null
-                                 }
-                             }
+                                 },
+                                 "xspec": {"type": "string"},
+                                 "depends": {
+                                    "type": "array",
+                                    "items": {"type": "string"},
+                                    "default": []
+                                }
+                            }
                          },
                          {"enum": [null, false]}
                      ],
@@ -87,7 +93,7 @@
                                 },
                                 "uenv": {
                                     "type": "object",
-                                    "additionalPropertis": false,
+                                    "additionalProperties": false,
                                     "properties": {
                                         "add_compilers": { "type": "boolean" },
                                         "prefix_paths": {
@@ -109,4 +115,3 @@
          }
      }
 }
-

unittests/test_schema.py

@@ -133,7 +133,7 @@ def test_environments_yaml(yaml_path):
         # test defaults were set correctly
         assert env["unify"] == "when_possible"
         assert env["packages"] == ["perl", "git"]
-        assert env["mpi"] == {"spec": "cray-mpich", "gpu": "cuda"}
+        assert env["mpi"] == {"depends": [], "spec": "cray-mpich", "gpu": "cuda"}
         assert env["variants"] == ["+mpi", "+cuda"]
         assert env["views"] == {"default": None}