def run(cloudburst: CloudburstConnection, num_requests: int, batch_size: int,
gpu: bool):
with open('imagenet_classes.txt', 'r') as f:
classes = [line.strip() for line in f.readlines()]
cloudburst.put_object('imagenet-classes', classes)
def resnet_init_gpu(self, cloudburst):
import os
import torch
import torchvision
from torchvision import transforms
tpath = os.path.join(os.getenv('TORCH_HOME'), 'checkpoints')
self.resnet = torch.load(os.path.join(tpath, 'resnet101.model')).cuda()
self.resnet.eval()
self.transforms = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
self.classes = cloudburst.get('imagenet-classes')
def resnet_model_gpu(self, table: Table) -> str:
"""
AlexNet for image classification on ImageNet
"""
import torch
inputs = []
for row in table.get():
img = self.transforms(row['img'])
inputs.append(img)
inputs = torch.stack(inputs, dim=0).cuda()
output = self.resnet(inputs)
_, indices = torch.sort(output, descending=True)
indices = indices.cpu().detach().numpy()
result = []
for idx_set in indices:
index = idx_set[0]
result.append(self.classes[index])
return result
def resnet_init_cpu(self, cloudburst):
import os
import torch
import torchvision
from torchvision import transforms
tpath = os.path.join(os.getenv('TORCH_HOME'), 'checkpoints')
self.resnet = torch.load(os.path.join(tpath, 'resnet101.model'))
self.resnet.eval()
self.transforms = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
self.classes = cloudburst.get('imagenet-classes')
def resnet_model_cpu(self, table: Table) -> str:
"""
AlexNet for image classification on ImageNet
"""
import torch
inputs = []
for row in table.get():
img = self.transforms(row['img'])
inputs.append(img)
inputs = torch.stack(inputs, dim=0)
output = self.resnet(inputs)
_, indices = torch.sort(output, descending=True)
indices = indices.detach().numpy()
result = []
for idx_set in indices:
index = idx_set[0]
result.append(self.classes[index])
return result
print(f'Creating flow with size {batch_size} batches.')
flow = Flow('batching-benchmark', FlowType.PUSH, cloudburst)
if gpu:
flow.map(resnet_model_gpu,
init=resnet_init_gpu,
names=['class'],
gpu=True,
batching=True)
else:
flow.map(resnet_model_cpu,
init=resnet_init_cpu,
names=['class'],
batching=True)
flow.deploy()
print('Flow successfully deployed!')
latencies = []
inp = Table([('img', NumpyType)])
img = np.array(Image.open('panda.jpg').convert('RGB').resize((224, 224)))
inp.insert([img])
kvs = cloudburst.kvs_client
if gpu:
print('Starting GPU warmup...')
for _ in range(50):
flow.run(inp).get()
print('Finished warmup...')
print('Starting benchmark...')
for i in range(num_requests):
if i % 100 == 0 and i > 0:
print(f'On request {i}...')
futs = []
for _ in range(batch_size):
futs.append(flow.run(inp))
pending = set([fut.obj_id for fut in futs])
# Break these apart to batch the KVS get requests.
start = time.time()
while len(pending) > 0:
get_start = time.time()
response = kvs.get(list(pending))
for key in response:
if response[key] is not None:
pending.discard(key)
end = time.time()
latencies.append(end - start)
compute_time = np.mean(latencies) * num_requests
tput = (batch_size * num_requests) / (compute_time)
print('THROUGHPUT: %.2f' % (tput))
print_latency_stats(latencies, 'E2E')
resnet = resnet_model_gpu
resnet_cons = resnet_init_gpu
incept = inceptionv3_model_gpu
incept_cons = inceptionv3_init_gpu
trans = transform_batch
else:
resnet = resnet_model
resnet_cons = resnet_init
incept = inceptionv3_model
incept_cons = inceptionv3_init
trans = transform
with open('imagenet_classes.txt', 'r') as f:
classes = [line.strip() for line in f.readlines()]
cloudburst.put_object('imagenet-classes', classes)
flow = Flow('cascade-flow', FlowType.PUSH, cloudburst)
rnet = flow.map(trans,
init=transform_init,
names=['img'],
batching=gpu) \
.map(resnet,
init=resnet_cons,
names=['img', 'resnet_index', 'resnet_max_prob'],
gpu=gpu,
batching=gpu)
incept = rnet.filter(low_prob) \
.map(incept,
init=incept_cons,
def run(cloudburst: CloudburstConnection,
num_requests: int,
data_size: str,
breakpoint: bool,
do_optimize: bool):
print('Creating data...')
size = DATA_SIZES[data_size]
for i in range(1, NUM_DATA_POINTS+1):
arr = np.random.rand(size)
cloudburst.put_object('data-' + str(i), arr)
def stage1(self, row: Row) -> (int, str):
idx = int(row['req_num'] / 10) + 1
key = 'data-%d' % (idx)
return idx, key
def stage2(self, row: Row) -> str:
import numpy as np
arr = row[row['key']]
return float(np.sum(arr))
print(f'Creating flow with {data_size} ({DATA_SIZES[data_size]}) inputs.')
flow = Flow('locality-benchmark', FlowType.PUSH, cloudburst)
flow.map(stage1, names=['index', 'key']) \
.lookup('key', dynamic=True) \
.map(stage2, names=['sum'])
optimize_rules['breakpoint'] = breakpoint
if do_optimize:
flow = optimize(flow, rules=optimize_rules)
print('Flow has been optimized...')
flow.deploy()
print('Flow successfully deployed!')
latencies = []
inp = Table([('req_num', IntType)])
if breakpoint:
print('Starting warmup...')
for i in range(NUM_DATA_POINTS):
inp = Table([('req_num', IntType)])
inp.insert([i * 10])
res = flow.run(inp).get()
print('Pausing to let cache metadata propagate...')
time.sleep(15)
print('Starting benchmark...')
for i in range(num_requests):
if i % 100 == 0 and i > 0:
print(f'On request {i}...')
inp = Table([('req_num', IntType)])
inp.insert([i])
start = time.time()
res = flow.run(inp).get()
end = time.time()
latencies.append(end - start)
with open('data.bts', 'wb') as f:
from cloudburst.shared.serializer import Serializer
ser = Serializer()
bts = ser.dump(latencies)
f.write(bts)
print_latency_stats(latencies, 'E2E')