Tensorflow Serving, TensorRT Inference Server (Triton), Multi Model Server (MXNet)

benchmark.md

Environments

• CPU: Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz
• GPU: NVIDIA V100
• Memory: 251GiB
• OS: Ubuntu 16.04.6 LTS (Xenial Xerus)

Docker Images:

• tensorflow/tensorflow:latest-gpu
• tensorflow/serving:latest-gpu
• nvcr.io/nvidia/tensorrtserver:19.10-py3

Framework	Model	Model Type	Images	Batch size	Time(s)
Tensorflow	ResNet50	TF Savedmodel	32000	32	83.189
Tensorflow	ResNet50	TF Savedmodel	32000	10	86.897
Tensorflow Serving	ResNet50	TF Savedmodel	32000	32	120.496
Tensorflow Serving	ResNet50	TF Savedmodel	32000	10	116.887
Triton (TensorRT Inference Server)	ResNet50	TF Savedmodel	32000	32	201.855
Triton (TensorRT Inference Server)	ResNet50	TF Savedmodel	32000	10	171.056
Falcon + msgpack + Tensorflow	ResNet50	TF Savedmodel	32000	32	115.686
Falcon + msgpack + Tensorflow	ResNet50	TF Savedmodel	32000	10	115.572

falcon_app.py

import falcon
from wsgiref import simple_server
import tensorflow as tf
import numpy as np
from time import time
from falcon import media


# setup msgpack handler in falcon
handlers = media.Handlers({
    'application/msgpack': media.MessagePackHandler(),
})

api = falcon.API(media_type='application/msgpack')

api.req_options.media_handlers = handlers
api.resp_options.media_handlers = handlers

gpus = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_visible_devices(gpus[0], 'GPU')
# tf.config.experimental.set_memory_growth(gpus[0], True)


class Inference():
    def __init__(self):
        self.model = tf.saved_model.load('./')

    def on_get(self, req, resp):
        resp.media = {'msg': 'working'}

    def on_post(self, req, resp):
        batch = req.media['batch']
        data = np.frombuffer(req.media['data'], dtype=np.float32).reshape((batch, 224, 224, 3))
        t0 = time()
        result = self.model(inputs=data).numpy()
        resp.media = {
            'result': result.tobytes(),
        }


api.add_route('/api', Inference())
httpd = simple_server.make_server('0.0.0.0', 8000, api)
httpd.serve_forever()

falcon_client.py

import requests
from time import time
import numpy as np
import msgpack
from tqdm import tqdm


if __name__ == "__main__":
    url = 'http://localhost:8000/api'
    session = requests.Session()
    session.headers.update({'Content-Type': 'application/msgpack'})
    packer = msgpack.Packer(
        autoreset=True,
        use_bin_type=True,
    )

    batch = 10
    epoch = 3200
    results = []

    print(msgpack.unpackb(session.get(url).content))
    resp = session.post(
        url,
        data=packer.pack({
            'data': np.random.random((batch, 224, 224, 3)).astype(np.float32).tobytes(),
            'batch': batch,
        })
    )
    print(msgpack.unpackb(resp.content))

    t0 = time()
    for i in tqdm(range(epoch)):
        resp = session.post(
            url,
            data=packer.pack({
                'data': np.random.random((batch, 224, 224, 3)).astype(np.float32).tobytes(),
                'batch': batch,
            })
        )
        results.append(msgpack.unpackb(resp.content)['result'])

    print(np.frombuffer(results[-1], dtype=np.float32))
    print(f'{epoch} epochs with batch {batch} ({epoch * batch}) in {time() - t0}')

multimodels.md

Model: ResNet50

• GPU: V100
• Test multiple models on one GPU card processing 32000 images (batch 10, epoch 3200)

Framework	GPU RAM	GPU Utilization	Time
PyTorch	2G	100%	35s
PyTorch x 2	4G	100%	76s
TensorFlow	8G	52%	81s
TensorFlow x 2	16G	67-97%	92s
ONNX Runtime GPU	1G	52%	72s
ONNX Runtime GPU x 2	2G	68-98%	103s

Model: SeResNeXt50

• GPU: V100
• Python: 3.8

Framework	Batch Size	Epoch	GPU RAM	GPU Utilization	Time
PyTorch (1.6.0)	1	32000	1.4G	61%	406s
TensorFlow (2.3.0)	1	32000	8G	70%	860s
ONNX Runtime GPU (1.4.0)	1	32000	0.9G	77%	264s
PyTorch (1.6.0)	10	3200	2.6G	93%	60s
TensorFlow (2.3.0)	10	3200	8G	70%	153s
ONNX Runtime GPU (1.4.0)	10	3200	1.2G	60%	97s
PyTorch (1.6.0)	32	1000	5.6G	97%	48s
TensorFlow (2.3.0)	32	1000	9G	70%	114s
ONNX Runtime GPU (1.4.0)	32	1000	1.9G	57%	91s

serving_model.md

Tensorflow Serving

https://www.tensorflow.org/tfx/serving

• coupled with Tensorflow ecosystem (also support other format, not out-of-box)
• A/B testing
• provide both gRPC and HTTP RESTful API
• prometheus integration
• batching
• multiple models
• preprocessing & postprocessing can be implemented with signatures

Triton Inference Server

https://github.com/NVIDIA/triton-inference-server/

• support multiply backends: ONNX, PyTorch, TensorFlow, Caffe2, TensorRT
• both gRPC and HTTP with SDK
• internal health check and prometheus metrics
• batching
• concurrent model execution
• preprocessing & postprocessing can be done with ensemble models
• shm-size, memlock, stack configurations are not available for Kubernetes

Multi Model Server

https://github.com/awslabs/multi-model-server

• require Java 8
• provide HTTP
• Java layer communicates with Python workers through Unix Domain Socket or TCP
• batching
• multiple models
• log4j
• management API
• need to write model loading and inference code (means can use any runtime you want)
• easy to add preprocessing and postprocessing to the service

Torch Serve

https://github.com/pytorch/serve

• forked from AWS Multi Model Server
• only for PyTorch models
• doesn't have specific optimization for PyTorch models

GraphPipe

https://oracle.github.io/graphpipe

• use flatbuffer which is more efficient
• 2 years ago...

tf_runtime.py

import tensorflow as tf
import numpy as np
from time import time
from tqdm import tqdm


gpus = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_visible_devices(gpus[1], 'GPU')
# tf.config.experimental.set_memory_growth(gpus[0], True)


if __name__ == "__main__":
    # model
    model = tf.saved_model.load('/tmp/resnet50_tf/1/')

    # init
    epoch = 3200
    batch = 10
    data = np.random.random((batch, 224, 224, 3)).astype(np.float32)
    print(model(inputs=data).numpy())

    # bench
    t0 = time()
    results = []
    for _ in tqdm(range(epoch)):
        data = np.random.random((batch, 224, 224, 3)).astype(np.float32)
        result = model(inputs=data).numpy()
        results.append(result)

    print(f'{epoch} epochs with batch {batch} ({batch * epoch}) in {time() - t0} seconds')

tf_serving.py

import grpc
import tensorflow as tf
import numpy as np
from time import time
from tqdm import tqdm

from tensorflow_serving.apis import predict_pb2
from tensorflow_serving.apis import prediction_service_pb2_grpc


if __name__ == "__main__":
    channel = grpc.insecure_channel('localhost:8500')
    stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
    # Send request
    # See prediction_service.proto for gRPC request/response details.
    request = predict_pb2.PredictRequest()
    request.model_spec.name = 'resnet50_tf'
    request.model_spec.signature_name = 'serving_default'

    batch = 10
    epoch = 3200
    results = []
    t0 = time()
    for _ in tqdm(range(epoch)):
        data = np.random.random((batch, 224, 224, 3)).astype(np.float32)
        request.inputs['input_1'].CopyFrom(
            tf.make_tensor_proto(data, shape=[batch, 224, 224, 3]))
        result = stub.Predict(request, 10.0)  # 10 secs timeout
        results.append(result)

    print(f'run {epoch} epochs with batch {batch} ({batch * epoch}) in {time() - t0} s')

triton_client.py

from time import time
from tqdm import tqdm
import numpy as np
import grpc
from tensorrtserver.api import api_pb2
from tensorrtserver.api import grpc_service_pb2
from tensorrtserver.api import grpc_service_pb2_grpc


if __name__ == '__main__':
    batch = 32
    channel = grpc.insecure_channel('localhost:8001')
    stub = grpc_service_pb2_grpc.GRPCServiceStub(channel)

    req = grpc_service_pb2.InferRequest()
    req.model_name = 'resnet50'
    req.model_version = -1
    req.meta_data.batch_size = batch
    output = api_pb2.InferRequestHeader.Output()
    output.name = 'predictions'
    output.cls.count = 1000
    req.meta_data.output.append(output)
    req.meta_data.input.add(name='input_1')

    # init
    del req.raw_input[:]
    data = b''
    for _ in range(batch):
        data += np.random.random((224, 224, 3)).astype(np.float32).tobytes()
    req.raw_input.append(data)
    print(stub.Infer(req))

    # bench
    t0 = time()
    epoch = 1000
    results = []
    for _ in tqdm(range(epoch)):
        del req.raw_input[:]
        data = b''
        for b in range(batch):
            data += np.random.random((224, 224, 3)).astype(np.float32).tobytes()
        req.raw_input.append(data)
        results.append(stub.Infer(req))

    print(f'{epoch} epochs with batch {batch} ({batch * epoch}) in {time() - t0} seconds')

This shows TF serving has a better web server, have you considered to use async non-blocking inference request?

This shows TF serving has a better web server, have you considered to use async non-blocking inference request?

By the way, this disabled the dynamic batch. For the async part, I personally think this is CPU and GPU bound tasks, so it won't benefit from async. But I haven't try the async frameworks like starlette. If you have such benchmark results, I'd like to know the details.

I recommend using https://github.com/triton-inference-server/server/blob/master/docs/perf_analyzer.md to continue your study, and put the result csv following the instruction here https://docs.nvidia.com/deeplearning/triton-inference-server/master-user-guide/docs/optimization.html#visualizing-latency-vs-throughput to understand the triton performance. Recently I did some perf analysis on Triton with TRT optimization, we are looking at RN50 about 3k images/s at ~20ms latency.

I recommend using https://github.com/triton-inference-server/server/blob/master/docs/perf_analyzer.md to continue your study, and put the result csv following the instruction here https://docs.nvidia.com/deeplearning/triton-inference-server/master-user-guide/docs/optimization.html#visualizing-latency-vs-throughput to understand the triton performance.

Thanks for your advice. I'll try this profiling tool.

Recently I did some perf analysis on Triton with TRT optimization, we are looking at RN50 about 3k images/s at ~50ms latency.

That's impressive. May I know your test environments? [GPU type + num, quantization?, distillation?]

1 T4, TRT INT8, no distillation