def main():
assert workspace.IsNUMAEnabled() and workspace.GetNumNUMANodes() >= 2
single_init, single_net = build_net("single_net", False)
cross_init, cross_net = build_net("cross_net", True)
workspace.CreateNet(single_init)
workspace.RunNet(single_init.Name())
workspace.CreateNet(cross_init)
workspace.RunNet(cross_init.Name())
workspace.CreateNet(single_net)
workspace.CreateNet(cross_net)
for _ in range(4):
t = time.time()
workspace.RunNet(single_net.Name(), NUM_ITER)
dt = time.time() - t
print("Single socket time:", dt)
single_bw = 4 * SHAPE_LEN * SHAPE_LEN * NUM_REPLICAS * NUM_ITER / dt / GB
print("Single socket BW: {} GB/s".format(single_bw))
t = time.time()
workspace.RunNet(cross_net.Name(), NUM_ITER)
dt = time.time() - t
print("Cross socket time:", dt)
cross_bw = 4 * SHAPE_LEN * SHAPE_LEN * NUM_REPLICAS * NUM_ITER / dt / GB
print("Cross socket BW: {} GB/s".format(cross_bw))
print("Single BW / Cross BW: {}".format(single_bw / cross_bw))
def main():
assert workspace.IsNUMAEnabled() and workspace.GetNumNUMANodes() >= 2
single_net = build_net("single_net", False)
cross_net = build_net("cross_net", True)
workspace.CreateNet(single_net)
workspace.CreateNet(cross_net)
for _ in range(4):
t = time.time()
workspace.RunNet(single_net.Name(), 5000)
print("Single socket time:", time.time() - t)
t = time.time()
workspace.RunNet(cross_net.Name(), 5000)
print("Cross socket time:", time.time() - t)
gpu_device_option = caffe2_pb2.DeviceOption()
gpu_device_option.device_type = caffe2_pb2.CUDA
gpu_device_option.device_id = 0
net.CopyCPUToGPU("output_blob_0",
"output_blob_0_gpu",
device_option=gpu_device_option)
net.CopyCPUToGPU("output_blob_1",
"output_blob_1_gpu",
device_option=gpu_device_option)
return net
@unittest.skipIf(not workspace.IsNUMAEnabled(), "NUMA is not enabled")
@unittest.skipIf(workspace.GetNumNUMANodes() < 2, "Not enough NUMA nodes")
@unittest.skipIf(not workspace.has_gpu_support, "No GPU support")
class NUMATest(TestCase):
def test_numa(self):
net = build_test_net("test_numa")
workspace.RunNetOnce(net)
self.assertEqual(workspace.GetBlobNUMANode("output_blob_0"), 0)
self.assertEqual(workspace.GetBlobNUMANode("output_blob_1"), 1)
if __name__ == '__main__':
unittest.main()
"--num-workers",
type=int,
default=1,
help="the number of worker per numa node in the thread pool",
)
parser.add_argument("--num-numa-nodes", type=int, default=1)
args, extra_args = parser.parse_known_args()
### some basic setup ###
np.random.seed(args.numpy_rand_seed)
np.set_printoptions(precision=args.print_precision)
global_options = [
"caffe2", "--caffe2_log_level=2", "--caffe2_cpu_numa_enabled=1"
] + extra_args
workspace.GlobalInit(global_options)
assert workspace.IsNUMAEnabled()
ln = np.fromstring(args.arch, dtype=int, sep="-")
# test prints
print("mlp arch (" + str(ln.size - 1) +
" layers, with input to output dimensions):")
print(ln)
### prepare training data ###
nbatches = int(np.ceil((args.data_size * 1.0) / args.mini_batch_size))
# inputs
m0 = ln[0]
lX = []
# targets
ml = ln[ln.size - 1]
lT = []
for j in range(0, nbatches):