|
| 1 | +import reframe as rfm |
| 2 | +import reframe.utility.sanity as sn |
| 3 | + |
| 4 | + |
| 5 | +@rfm.required_version('>=2.16-dev0') |
| 6 | +@rfm.simple_test |
| 7 | +class HPCGCheckRef(rfm.RegressionTest): |
| 8 | + def __init__(self): |
| 9 | + super().__init__() |
| 10 | + |
| 11 | + self.descr = 'HPCG reference benchmark' |
| 12 | + self.valid_systems = ['daint:mc', 'daint:gpu', 'dom:gpu', 'dom:mc'] |
| 13 | + self.valid_prog_environs = ['PrgEnv-gnu'] |
| 14 | + self.modules = ['craype-hugepages8M'] |
| 15 | + self.build_system = 'Make' |
| 16 | + self.build_system.options = ['arch=MPI_GCC_OMP'] |
| 17 | + self.sourcesdir = 'https://github.com/hpcg-benchmark/hpcg.git' |
| 18 | + |
| 19 | + self.executable = 'bin/xhpcg' |
| 20 | + self.executable_opts = ['--nx=104', '--ny=104', '--nz=104', '-t2'] |
| 21 | + # use glob to catch the output file suffix dependent on execution time |
| 22 | + self.output_file = sn.getitem(sn.glob('HPCG*.txt'), 0) |
| 23 | + |
| 24 | + self.num_tasks = 0 |
| 25 | + self.num_cpus_per_task = 1 |
| 26 | + self.system_num_tasks = { |
| 27 | + 'daint:mc': 36, |
| 28 | + 'daint:gpu': 12, |
| 29 | + 'dom:mc': 36, |
| 30 | + 'dom:gpu': 12, |
| 31 | + } |
| 32 | + |
| 33 | + self.reference = { |
| 34 | + 'daint:gpu': { |
| 35 | + 'gflops': (7.6, -0.1, None, 'Gflop/s') |
| 36 | + }, |
| 37 | + 'daint:mc': { |
| 38 | + 'gflops': (13.4, -0.1, None, 'Gflop/s') |
| 39 | + }, |
| 40 | + 'dom:gpu': { |
| 41 | + 'gflops': (7.6, -0.1, None, 'Gflop/s') |
| 42 | + }, |
| 43 | + 'dom:mc': { |
| 44 | + 'gflops': (13.4, -0.1, None, 'Gflop/s') |
| 45 | + }, |
| 46 | + } |
| 47 | + |
| 48 | + self.maintainers = ['SK'] |
| 49 | + self.tags = {'diagnostic'} |
| 50 | + |
| 51 | + @property |
| 52 | + @sn.sanity_function |
| 53 | + def num_tasks_assigned(self): |
| 54 | + return self.job.num_tasks |
| 55 | + |
| 56 | + def setup(self, partition, environ, **job_opts): |
| 57 | + self.num_tasks_per_node = self.system_num_tasks[partition.fullname] |
| 58 | + |
| 59 | + num_nodes = self.num_tasks_assigned / self.num_tasks_per_node |
| 60 | + self.perf_patterns = { |
| 61 | + 'gflops': sn.extractsingle( |
| 62 | + r'HPCG result is VALID with a GFLOP\/s rating of=\s*' |
| 63 | + r'(?P<perf>\S+)', |
| 64 | + self.output_file, 'perf', float) / num_nodes |
| 65 | + } |
| 66 | + |
| 67 | + self.sanity_patterns = sn.all([ |
| 68 | + sn.assert_eq(4, sn.count( |
| 69 | + sn.findall(r'PASSED', self.output_file))), |
| 70 | + sn.assert_eq(0, self.num_tasks_assigned % self.num_tasks_per_node) |
| 71 | + ]) |
| 72 | + |
| 73 | + super().setup(partition, environ, **job_opts) |
| 74 | + |
| 75 | + |
| 76 | +@rfm.required_version('>=2.16-dev0') |
| 77 | +@rfm.simple_test |
| 78 | +class HPCGCheckMKL(rfm.RegressionTest): |
| 79 | + def __init__(self): |
| 80 | + super().__init__() |
| 81 | + |
| 82 | + self.descr = 'HPCG benchmark Intel MKL implementation' |
| 83 | + self.valid_systems = ['daint:mc', 'dom:mc', 'daint:gpu', 'dom:gpu'] |
| 84 | + self.valid_prog_environs = ['PrgEnv-intel'] |
| 85 | + self.modules = ['craype-hugepages8M'] |
| 86 | + self.build_system = 'Make' |
| 87 | + self.prebuild_cmd = ['cp -r ${MKLROOT}/benchmarks/hpcg/* .', |
| 88 | + 'mv Make.CrayXC setup', |
| 89 | + './configure CrayXC'] |
| 90 | + |
| 91 | + self.num_tasks = 0 |
| 92 | + self.num_tasks_per_core = 2 |
| 93 | + self.problem_size = 104 |
| 94 | + |
| 95 | + self.variables = { |
| 96 | + 'HUGETLB_VERBOSE': '0', |
| 97 | + 'MPICH_MAX_THREAD_SAFETY': 'multiple', |
| 98 | + 'MPICH_USE_DMAPP_COLL': '1', |
| 99 | + 'PMI_NO_FORK': '1', |
| 100 | + 'KMP_HW_SUBSET': '9c,2t', |
| 101 | + 'KMP_AFFINITY': 'granularity=fine,compact' |
| 102 | + } |
| 103 | + |
| 104 | + self.executable = 'bin/xhpcg_avx2' |
| 105 | + self.executable_opts = ['--nx=%d' % self.problem_size, |
| 106 | + '--ny=%d' % self.problem_size, |
| 107 | + '--nz=%d' % self.problem_size, '-t2'] |
| 108 | + |
| 109 | + self.reference = { |
| 110 | + 'dom:mc': { |
| 111 | + 'gflops': (22, -0.1, None, 'Gflop/s') |
| 112 | + }, |
| 113 | + 'daint:mc': { |
| 114 | + 'gflops': (22, -0.1, None, 'Gflop/s') |
| 115 | + }, |
| 116 | + 'dom:gpu': { |
| 117 | + 'gflops': (10.7, -0.1, None, 'Gflop/s') |
| 118 | + }, |
| 119 | + 'daint:gpu': { |
| 120 | + 'gflops': (10.7, -0.1, None, 'Gflop/s') |
| 121 | + }, |
| 122 | + } |
| 123 | + |
| 124 | + self.maintainers = ['SK'] |
| 125 | + self.tags = {'diagnostic'} |
| 126 | + |
| 127 | + @property |
| 128 | + @sn.sanity_function |
| 129 | + def num_tasks_assigned(self): |
| 130 | + return self.job.num_tasks |
| 131 | + |
| 132 | + @property |
| 133 | + @sn.sanity_function |
| 134 | + def outfile_lazy(self): |
| 135 | + pattern = 'n%d-%dp-%dt-*.yaml' % (self.problem_size, |
| 136 | + self.job.num_tasks, |
| 137 | + self.num_cpus_per_task) |
| 138 | + return sn.getitem(sn.glob(pattern), 0) |
| 139 | + |
| 140 | + def setup(self, partition, environ, **job_opts): |
| 141 | + if partition.fullname in ['daint:gpu', 'dom:gpu']: |
| 142 | + self.num_tasks_per_node = 2 |
| 143 | + self.num_cpus_per_task = 12 |
| 144 | + else: |
| 145 | + self.num_tasks_per_node = 4 |
| 146 | + self.num_cpus_per_task = 18 |
| 147 | + |
| 148 | + # since this is a flexible test, we divide the extracted |
| 149 | + # performance by the number of nodes and compare |
| 150 | + # against a single reference |
| 151 | + num_nodes = self.num_tasks_assigned / self.num_tasks_per_node |
| 152 | + self.perf_patterns = { |
| 153 | + 'gflops': sn.extractsingle( |
| 154 | + r'HPCG result is VALID with a GFLOP\/s rating of:\s*' |
| 155 | + r'(?P<perf>\S+)', |
| 156 | + self.outfile_lazy, 'perf', float) / num_nodes |
| 157 | + } |
| 158 | + |
| 159 | + self.sanity_patterns = sn.all([ |
| 160 | + sn.assert_eq(4, sn.count( |
| 161 | + sn.findall(r'PASSED', self.outfile_lazy))), |
| 162 | + sn.assert_eq(0, self.num_tasks_assigned % self.num_tasks_per_node) |
| 163 | + ]) |
| 164 | + |
| 165 | + super().setup(partition, environ, **job_opts) |
0 commit comments