def test_normalize(self):
# cpurequest absent, memorylimit absent
self.params.update({
"gpu_limit": 0,
"cpulimit": "1000m",
"memoryrequest": "2048Mi"
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("1000m", job_params.cpu_limit)
self.assertEqual("2048Mi", job_params.memory_request)
self.assertEqual("2048Mi", job_params.memory_limit)
# cpurequest > cpulimit
self.params["cpurequest"] = "2000m"
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("1000m", job_params.cpu_limit)
self.assertEqual("2048Mi", job_params.memory_request)
self.assertEqual("2048Mi", job_params.memory_limit)
def test_backward_compatibility_cpu_job(self):
# Cpu job on cpu node
self.params["resourcegpu"] = 0
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("2000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("8192Mi", job_params.memory_limit)
# Cpu job on gpu node
self.quota["cpu"].pop("Standard_D2s_v3", None)
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
def test_cpu_job_on_gpu_node(self):
# Cpu job on gpu node will have default value
self.params.update({
"gpu_limit": 0,
"sku": "Standard_ND24rs",
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
# Gpu proportional should show the same
self.config["job_resource_policy"] = "gpu_proportional"
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
def test_gpu_job(self):
# gpu_limit precedes resourcegpu
self.params["resourcegpu"] = 0
self.params["gpu_limit"] = 3
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(3, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
# Gpu proportional
self.config["job_resource_policy"] = "gpu_proportional"
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(3, job_params.gpu_limit)
self.assertEqual("5000m", job_params.cpu_request)
self.assertEqual("6000m", job_params.cpu_limit)
self.assertEqual("103219Mi", job_params.memory_request)
self.assertEqual("114688Mi", job_params.memory_limit)
# Request override
self.params.update({
"cpurequest": "1200m",
"memoryrequest": "4096Mi",
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(3, job_params.gpu_limit)
self.assertEqual("1200m", job_params.cpu_request)
self.assertEqual("1200m", job_params.cpu_limit)
self.assertEqual("4096Mi", job_params.memory_request)
self.assertEqual("4096Mi", job_params.memory_limit)
# Limit override
self.params.update({
"cpulimit": "1500m",
"memorylimit": "4608Mi",
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(3, job_params.gpu_limit)
self.assertEqual("1200m", job_params.cpu_request)
self.assertEqual("1500m", job_params.cpu_limit)
self.assertEqual("4096Mi", job_params.memory_request)
self.assertEqual("4608Mi", job_params.memory_limit)
def test_cpu_job_on_cpu_node(self):
self.params["gpu_limit"] = 0
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("2000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("8192Mi", job_params.memory_limit)
# Gpu proportional should show the same
self.config["job_resource_policy"] = "gpu_proportional"
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("2000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("8192Mi", job_params.memory_limit)
# Request override
self.params.update({
"cpurequest": "1200m",
"memoryrequest": "4096Mi",
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1200m", job_params.cpu_request)
self.assertEqual("1200m", job_params.cpu_limit)
self.assertEqual("4096Mi", job_params.memory_request)
self.assertEqual("4096Mi", job_params.memory_limit)
# Limit override
self.params.update({
"cpulimit": "1500m",
"memorylimit": "4608Mi",
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1200m", job_params.cpu_request)
self.assertEqual("1500m", job_params.cpu_limit)
self.assertEqual("4096Mi", job_params.memory_request)
self.assertEqual("4608Mi", job_params.memory_limit)
def test_cpu_job_on_cpu_node(self):
self.params["gpu_limit"] = 0
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(None, job_params.gpu_type)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("2000m", job_params.cpu_limit)
self.assertEqual("7372Mi", job_params.memory_request)
self.assertEqual("8192Mi", job_params.memory_limit)
# Request override
self.params.update({
"cpurequest": "1200m",
"memoryrequest": "4096Mi",
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(None, job_params.gpu_type)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1200m", job_params.cpu_request)
self.assertEqual("1200m", job_params.cpu_limit)
self.assertEqual("4096Mi", job_params.memory_request)
self.assertEqual("4096Mi", job_params.memory_limit)
# Limit override
self.params.update({
"cpulimit": "1500m",
"memorylimit": "4608Mi",
})
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(None, job_params.gpu_type)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1200m", job_params.cpu_request)
self.assertEqual("1500m", job_params.cpu_limit)
self.assertEqual("4096Mi", job_params.memory_request)
self.assertEqual("4608Mi", job_params.memory_limit)
def test_gpu_job(self):
# gpu_limit precedes resourcegpu
self.params["resourcegpu"] = 0
self.params["gpu_limit"] = 3
self.params["_allow_partial_node"] = True
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual("P40", job_params.gpu_type)
self.assertEqual(4, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
# Admin is allowed to submit partial nodes
job_params = make_job_params(self.params,
self.quota,
self.metadata,
self.config,
is_admin=True)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual("P40", job_params.gpu_type)
self.assertEqual(3, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
# Gpu proportional
self.config["job_resource_policy"] = "gpu_proportional"
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual("P40", job_params.gpu_type)
self.assertEqual(4, job_params.gpu_limit)
self.assertEqual("21000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("412876Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
def test_backward_compatibility_gpu_job(self):
# Gpu job with proportionally assigned cpu and memory
self.params["resourcegpu"] = 3
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(3, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
def test_backward_compatibility_gpu_job(self):
# 1 gpu per pod for workers on gpu nodes, total 2 gpus
self.params["resourcegpu"] = 2
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(2, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
def test_backward_compatibility_gpu_job(self):
# Gpu jobs asking for all gpus on nodes
self.params["resourcegpu"] = 4
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_ND24rs", job_params.sku)
self.assertEqual(4, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("24000m", job_params.cpu_limit)
self.assertEqual("0Mi", job_params.memory_request)
self.assertEqual("458752Mi", job_params.memory_limit)
def test_backward_compatibility_cpu_job(self):
# Cpu job running on cpu nodes occupy entire nodes
self.params["resourcegpu"] = 0
job_params = make_job_params(self.params, self.quota, self.metadata,
self.config)
self.assertIsNotNone(job_params)
self.assertTrue(job_params.is_valid())
self.assertEqual("Standard_D2s_v3", job_params.sku)
self.assertEqual(0, job_params.gpu_limit)
self.assertEqual("1000m", job_params.cpu_request)
self.assertEqual("2000m", job_params.cpu_limit)
self.assertEqual("7372Mi", job_params.memory_request)
self.assertEqual("8192Mi", job_params.memory_limit)