def main(num_replicas, batch_size, method, filename):
for node_id in ["up", "down"]:
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(
max_batch_size=batch_size, num_replicas=num_replicas
)
serve_benchmark.create_backend(noop, node_id, backend_config=config)
serve_benchmark.link(node_id, node_id)
with serve_benchmark.using_router("up"):
up_handle = serve_benchmark.get_handle("up")
with serve_benchmark.using_router("down"):
down_handle = serve_benchmark.get_handle("down")
# start = time.perf_counter()
oids = []
if method == "chain":
# closed loop latencies
for i in range(num_queries):
r = up_handle.options(tracing_metadata={"pipeline-id": i}).remote(
data=image_data
)
oid = down_handle.options(
tracing_metadata={"pipeline-id": i}
).remote(
data=r # torch tensor
)
ray.wait([oid], 1)
elif method == "group":
res = [
up_handle.options(tracing_metadata={"pipeline-id": i}).remote(
sleep_time=0.01, data=image_data
)
for i in range(num_queries)
]
oids = [
down_handle.options(tracing_metadata={"pipeline-id": i}).remote(
sleep_time=0.02, data=d # torch tensor
)
for i, d in enumerate(res)
]
else:
raise RuntimeError("Unreachable")
# print(f"Submission time {time.perf_counter() - start}")
# end = time.perf_counter()
# duration = end - start
# qps = num_queries / duration
# print(f"Throughput {qps}")
trace_file = filename or tempfile.mkstemp(suffix=".json")[1]
with open(trace_file, "w") as f:
json.dump(serve_benchmark.get_trace(), f)
ray.timeline(f"ray_trace_{batch_size}-{method}.json")
print(f"Serve trace file written to {trace_file}")
def __init__(self, vertex_config, model_name):
super().__init__()
handle_list = list()
for node_id in vertex_config.keys():
backend_config = vertex_config[node_id]
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(**backend_config)
if node_id == "prepoc":
min_img_size = 224
transform = transforms.Compose([
transforms.Resize(min_img_size),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
])
serve_benchmark.create_backend(Transform,
node_id,
transform,
backend_config=config)
elif node_id == "model":
serve_benchmark.create_backend(
PredictModelPytorch,
node_id,
model_name,
True,
backend_config=config,
)
serve_benchmark.link(node_id, node_id)
handle_list.append(serve_benchmark.get_handle(node_id))
self.chain_handle = ChainHandle(handle_list)
def __init__(self, max_batch_size, pipeline_length):
self.plength = pipeline_length
self.handles = list()
for index in range(self.plength):
node_id = f"service-{index}"
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(
max_batch_size=max_batch_size, num_replicas=1)
serve_benchmark.create_backend(noop,
node_id,
backend_config=config)
serve_benchmark.link(node_id, node_id)
self.handles.append(serve_benchmark.get_handle(node_id))
def get_backend_config(self, backend_tag=None):
with serve_benchmark.using_router(self.endpoint_name):
backend_tag = self._ensure_backend_unique(backend_tag)
return serve_benchmark.get_backend_config(backend_tag)
def set_max_batch_size(self, new_max_batch_size, backend_tag=None):
with serve_benchmark.using_router(self.endpoint_name):
backend_tag = self._ensure_backend_unique(backend_tag)
config = serve_benchmark.get_backend_config(backend_tag)
config.max_batch_size = new_max_batch_size
serve_benchmark.set_backend_config(backend_tag, config)
def scale(self, new_num_replicas, backend_tag=None):
with serve_benchmark.using_router(self.endpoint_name):
backend_tag = self._ensure_backend_unique(backend_tag)
config = serve_benchmark.get_backend_config(backend_tag)
config.num_replicas = new_num_replicas
serve_benchmark.set_backend_config(backend_tag, config)
def main(num_replicas, method):
for node_id in ["up", "down"]:
with serve_benchmark.using_router(node_id):
serve_benchmark.create_endpoint(node_id)
config = serve_benchmark.BackendConfig(max_batch_size=1,
num_replicas=num_replicas)
serve_benchmark.create_backend(noop,
node_id,
backend_config=config)
serve_benchmark.link(node_id, node_id)
with serve_benchmark.using_router("up"):
up_handle = serve_benchmark.get_handle("up")
with serve_benchmark.using_router("down"):
down_handle = serve_benchmark.get_handle("down")
start = time.perf_counter()
oids = []
if method == "chain":
for i in range(num_queries):
r = up_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(sleep_time=0.01, data=image_data)
r = down_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(
sleep_time=0.02,
data=r # torch tensor
)
oids.append(r)
elif method == "group":
res = [
up_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(sleep_time=0.01, data=image_data)
for i in range(num_queries)
]
oids = [
down_handle.options(tracing_metadata={
"pipeline-id": i
}).remote(
sleep_time=0.02,
data=d # torch tensor
) for i, d in enumerate(res)
]
else:
raise RuntimeError("Unreachable")
print(f"Submission time {time.perf_counter() - start}")
ray.wait(oids, len(oids))
end = time.perf_counter()
duration = end - start
qps = num_queries / duration
print(f"Throughput {qps}")
_, trace_file = tempfile.mkstemp(suffix=".json")
with open(trace_file, "w") as f:
json.dump(serve_benchmark.get_trace(), f)
print(f"trace file written to {trace_file}")