H100: Add benchmark stage and test more features (#4870)

ci/cscs_daint_spack_psmp.yaml

@@ -89,8 +89,8 @@ regression test cp2k daint:
     SLURM_DEBUG: 1
     SLURM_HINT: nomultithread
     SLURM_JOB_NUM_NODES: 1
-    SLURM_NTASKS: 228
-    SLURM_TIMELIMIT: 30
+    SLURM_NTASKS: 288
+    SLURM_TIMELIMIT: 40
     USE_MPI: YES
 
 benchmark cp2k 1 daint node:

ci/cscs_daint_spack_psmp_h100.yaml

@@ -5,6 +5,7 @@ stages:
   - deps
   - cp2k
   - test
+  - benchmark
 
 variables:
   REPOSITORY: cp2k_daint
@@ -35,7 +36,7 @@ build deps daint:
     DOCKERFILE: ci/docker/build_${CI_JOB_STAGE}_${BUILDER}.Dockerfile
     BASE_IMAGE: ${UBUNTU_IMAGE_NAME}
     CP2K_VERSION: ${VERSION}
-    FEATURE_FLAGS: "-gpu H100 -df all"
+    FEATURE_FLAGS: "-gpu H100"
 
 build cp2k daint:
   extends: .container-builder-cscs-gh200
@@ -56,7 +57,7 @@ build cp2k daint:
     DOCKERFILE: ci/docker/build_${CI_JOB_STAGE}_${BUILDER}.Dockerfile
     BASE_IMAGE: ${UBUNTU_IMAGE_NAME}
     DEPS_IMAGE: ${DEPS_IMAGE_NAME}
-    FEATURE_FLAGS: "-gpu H100 -df all"
+    FEATURE_FLAGS: "-gpu H100"
     PERSIST_IMAGE_NAME: ${CSCS_IMAGE_NAME}
 
 test cp2k daint:
@@ -88,6 +89,29 @@ regression test cp2k daint:
     SLURM_DEBUG: 1
     SLURM_HINT: nomultithread
     SLURM_JOB_NUM_NODES: 1
-    SLURM_NTASKS: 228
+    SLURM_NTASKS: 224
+    SLURM_TIMELIMIT: 40
+    USE_MPI: YES
+
+benchmark cp2k 1 daint node:
+  extends: .container-runner-daint-gh200
+  stage: benchmark
+  needs: ["build cp2k daint","test cp2k daint"]
+  timeout: 1h
+  image: ${CSCS_IMAGE_NAME}
+  before_script:
+    - export CP2K_DATA_DIR=/opt/cp2k/install/share/cp2k/data
+    - export OMP_NUM_THREADS=3
+  script:
+    - cp2k /opt/cp2k/benchmarks/CI/H2O-512_md.inp
+  variables:
+    SLURM_CPU_BIND: cores
+    SLURM_CPUS_PER_TASK: 3
+    SLURM_DEBUG: 1
+    SLURM_HINT: nomultithread
+    SLURM_JOB_NUM_NODES: 1
+    SLURM_MPI_TYPE: pmi2
+    SLURM_NTASKS: 72
+    SLURM_NTASKS_PER_NODE: 72
     SLURM_TIMELIMIT: 30
     USE_MPI: YES