Building Google Cloud ML Engine From Scratch on AWS with PipelineAI - ODSC London 2017 - Oct 13,...

BUILDING GOOGLE CLOUD ML ENGINE FROM SCRATCH WITH PIPELINE.AI

ODSC CONFERENCELONDON, ENGLANDOCTOBER 13, 2017

CHRIS FREGLY, FOUNDER @ PIPELINE.AI

INTRODUCTIONS: ME§ Chris Fregly, Research Engineer @ PipelineAI§ Formerly Netflix, Databricks, IBM Spark Center

§ Advanced Spark and TensorFlow MeetupPlease Join Our 40,000+ Members Globally!

Contact [email protected]

@cfregly

*San Francisco*Chicago*Austin*Washington DC*London

INTRODUCTIONS: YOU

§ Software Engineer or Data Scientist interested in optimizing and deploying TensorFlow models to production

§ Assume you have a working knowledge of TensorFlow

CONTENT BREAKDOWN

§ PipelineAI Features§ 50% Training Optimizations (GPUs, Pipeline, XLA+JIT)§ 50% Prediction Optimizations (XLA+AOT, TF Serving)

§ Why Heavy Focus on Predicting?§ Training: boring batch O(num_data_scientists)§ Inference: exciting real-time O(num_users_of_app)

100% OPEN SOURCE CODE§ https://github.com/PipelineAI/pipeline/

§ Please 🌟 this GitHub Repo!

§ All slides, code, notebooks, and Docker images here:https://github.com/PipelineAI/pipeline/tree/master/gpu

AGENDA§ PipelineAI Features

§ Experiment Safely in Production§ Tune Both Model + Runtime Parameters§ Compare Models Both Offline + Online§ Shift Traffic (Across Clouds) to Winning Model

§ Optimize TensorFlow Training§ GPUs + Ingestion + Training Pipeline§ XLA JIT Compiler

§ Optimize TensorFlow Inference§ XLA AOT Compiler + Graph Transform Tool (GTT)§ TensorFlow Serving

PIPELINE.AI

EXPERIMENT SAFELY IN PRODUCTION

§ Setup Experiments Directly from Jupyter Notebooks

§ Deploy to 1% Prod Traffic

§ Or Deploy in Shadow Mode

§ Tear-Down Experiments Quickly





MODEL + RUNTIME PACKAGING

§ Package Model + Runtime into Immutable Docker Image

§ Same Package: Local, Dev, and Prod

§ No Dependency Surprises in Production

OPTIMIZE MODEL + RUNTIME AS ONE

§ Tune Model Params + Runtime Configs Together

§ Generate Native CPU + GPU Code

§ Quantize Model Weights + Activations

§ Swap Runtimes: TF Serving, TensorRT, CPU, GPU, TPU

NVIDIA TENSORRT RUNTIME

§ Performs Post-Training Optimizations

§ GPU-Optimized Prediction Runtime

§ Alternative to TensorFlow Serving





COMPARE MODELS OFFLINE + ONLINE

§ Offline Metrics§ Training Accuracy§ Validation Accuracy

§ Online / Real-Time Metrics§ Prediction Precision§ Latency + Throughput

PREDICTION PROFILING + TUNING§ Pinpoint Performance Bottlenecks

§ Fine-Grained Prediction Metrics

§ 3 Logic Steps in a Prediction1.transform_request()2.predict()3.transform_response()





SHIFT TRAFFIC TO MAXIMIZE REVENUE§ Shift Traffic to Winning Model using Bandit AI Algorithms

SHIFT TRAFFIC TO MINIMIZE COST

§ Real-Time Cost Per Prediction

§ Across Clouds + On-Premise

§ Bandit-Based Explore/Exploit

VIEW LIVE PREDICTION STREAMS§ Visually Compare Real-Time Predictions

CONTINUOUS MODEL TRAINING§ Identify and Fix Borderline Predictions (50-50% Confidence)

§ Fix Along Class Boundaries

§ Retrain on New Labeled Data

§ Enables Crowd Sourcing

§ Game-ify Labeling Process





SETTING UP TENSORFLOW WITH GPUS

§ Very Painful!

§ Especially inside Docker§ Use nvidia-docker

§ Especially on Kubernetes!§ Use Kubernetes 1.7+

§ http://pipeline.ai for GitHub + DockerHub Links

GPU HALF-PRECISION SUPPORT§ FP32 is “Full Precision”, FP16 is “Half Precision”§ Supported by Pascal P100 (2016) and Volta V100 (2017)§ Flexible FP32 GPU Cores Can Fit 2 FP16’s for 2x Throughput!§ Half-Precision is OK for Approximate Deep Learning Use Cases

VOLTA V100 RECENTLY ANNOUNCED§ 84 Streaming Multiprocessors (SM’s)§ 5,376 GPU Cores§ 672 Tensor Cores (ie. Google TPU)

§ Mixed FP16/FP32 Precision § More Shared Memory§ New L0 Instruction Cache§ Faster L1 Data Cache§ V100 vs. P100 Performance

§ 12x TFLOPS @ Peak Training§ 6x Inference Throughput

V100 AND CUDA 9§ Independent Thread Scheduling - Finally!!

§ Similar to CPU fine-grained thread synchronization semantics§ Allows GPU to yield execution of any thread

§ Still Optimized for SIMT (Same Instruction Multiple Thread)§ SIMT units automatically scheduled together

§ Explicit Thread Synchronization

P100 V100

GPU CUDA PROGRAMMING

§ Barbaric, But Fun Barbaric

§ Must Know Hardware Very Well

§ Hardware Changes are Painful

§ Many Great Debuggers Exist

CUDA STREAMS

§ Asynchronous I/O Transfer

§ Overlap Compute and I/O

§ Keeps GPUs Saturated

§ Fundamental to Queue Framework in TensorFlow

TRAINING TERMINOLOGY§ Tensors: N-Dimensional Arrays§ ie. Scalar, Vector, Matrix

§ Operations: MatMul, Add, SummaryLog,…§ Graph: Graph of Operations (DAG)§ Session: Contains Graph(s)§ Feeds: Feed inputs into Placeholder§ Fetches: Fetch output from Operation§ Variables: What we learn through training§ aka “weights”, “parameters”

§ Devices: Hardware device on which we train

-TensorFlow-Trains

Variables

-User-FetchesOutputs

-User-FeedsInputs

-TensorFlow-Performs

Operations

-TensorFlow-Flows

Tensors

with tf.device(“/gpu:0,/gpu:1”)

TENSORFLOW MODEL§ MetaGraph

§ Combines GraphDef and Metadata§ GraphDef

§ Architecture of your model (nodes, edges)

§ Metadata§ Asset: Accompanying assets to your model§ SignatureDef: Maps external : internal tensors

§ Variables§ Stored separately during training (checkpoint)§ Allows training to continue from any checkpoint§ Variables are “frozen” into Constants when deployed for inference

GraphDef

x

W

mul add

b

MetaGraphMetadata

AssetsSignatureDef

TagsVersion

Variables:“W” : 0.328“b” : -1.407

TENSORFLOW SESSION

Session

graph: GraphDef

Variables:“W” : 0.328“b” : -1.407

Variables arePeriodically

Checkpointed

GraphDefis Static

EXTEND EXISTING DATA PIPELINES§ Data Processing

§ HDFS/Hadoop§ Spark

§ Containers§ Docker§ Google Container

§ Container Orchestrators§ Kubernetes§ Mesos

<dependency> <groupId>org.tensorflow</groupId> <artifactId>tensorflow-hadoop</artifactId>

</dependency>

https://github.com/tensorflow/ecosystem

DON’T USE FEED_DICT

§ Not Optimized for Production Pipelines§ feed_dict Requires Python <-> C++ Serialization§ Single-threaded, Synchronous, SLOW!§ Can’t Retrieve Until Current Batch is Complete§ CPUs/GPUs Not Fully Utilized!§ Use Queue or Dataset API

QUEUES

§ More than just a traditional Queue§ Perform I/O, pre-processing, cropping, shuffling§ Pulls from HDFS, S3, Google Storage, Kafka, ...§ Combine many small files into large TFRecord files§ Use CPUs to free GPUs for compute§ Uses CUDA Streams§ Helps saturate CPUs and GPUs

QUEUE CAPACITY PLANNING§ batch_size

§ # examples / batch (ie. 64 jpg)§ Limited by GPU RAM

§ num_processing_threads§ CPU threads pull and pre-process batches of data§ Limited by CPU Cores

§ queue_capacity§ Limited by CPU RAM (ie. 5 * batch_size)

DETECT UNDERUTILIZED CPUS, GPUS

§ Instrument training code to generate “timelines”

§ Analyze with Google Web Tracing Framework (WTF)

§ Monitor CPU with `top`, GPU with `nvidia-smi`

http://google.github.io/tracing-framework/

from tensorflow.python.client import timeline

trace = timeline.Timeline(step_stats=run_metadata.step_stats)

with open('timeline.json', 'w') as trace_file:trace_file.write(trace.generate_chrome_trace_format(show_memory=True))

SINGLE NODE, MULTI-GPU TRAINING§ cpu:0

§ By default, all CPUs§ Requires extra config to target a CPU

§ gpu:0..n§ Each GPU has a unique id§ TF usually prefers a single GPU

§ xla_cpu:0, xla_gpu:0..n§ “JIT Compiler Device”§ Hints TensorFlow to attempt JIT Compile

with tf.device(“/cpu:0”):

with tf.device(“/gpu:0”):

with tf.device(“/gpu:1”):

GPU 0 GPU 1

MULTI-NODE DISTRIBUTED TRAINING§ TensorFlow Automatically Inserts Send and Receive Ops into Graph§ Parameter Server Synchronously Aggregates Updates to Variables§ Nodes with Multiple GPUs will Pre-Aggregate Before Sending to PS

Worker0 Worker0

Worker1

Worker0 Worker1 Worker2

gpu0 gpu1

gpu2 gpu3

gpu0 gpu1

gpu2 gpu3

gpu0 gpu1

gpu2 gpu3

gpu0

gpu1

gpu0

gpu0

SYNCHRONOUS VS. ASYNCHRONOUS§ Synchronous

§ Nodes compute gradients§ Nodes update Parameter Server (PS)§ Nodes sync on PS for latest gradients

§ Asynchronous§ Some nodes delay in computing gradients§ Nodes don’t update PS§ Nodes get stale gradients from PS§ May not converge due to stale reads!

BATCH NORMALIZATION

§ Each mini-batch may have wildly different distributions§ Normalize per batch (and layer)§ Speeds up training!!§ Weights are learned quicker§ Final model is more accurate§ Final mean and variance will be folded into Graph later

-- Pretty Much Always Use Batch Normalization! --

z = tf.matmul(a_prev, W)a = tf.nn.relu(z)

a_mean, a_var = tf.nn.moments(a, [0])

scale = tf.Variable(tf.ones([depth/channels]))beta = tf.Variable(tf.zeros ([depth/channels]))

bn = tf.nn.batch_normalizaton(a, a_mean, a_var, beta, scale, 0.001)

OPTIMIZE GRAPH EXECUTION ORDER

§ https://github.com/yaroslavvb/stuff

Linearize to minimize graphmemory usage

SEPARATE TRAINING + VALIDATION

§ Separate Training and Validation Clusters

§ Validate Upon Checkpoint

§ Avoids Resource Contention

TrainingCluster

ValidationCluster

Parameter ServerCluster





XLA FRAMEWORK§ Accelerated Linear Algebra (XLA)§ Goals:

§ Reduce reliance on custom operators§ Improve execution speed§ Improve memory usage§ Reduce mobile footprint§ Improve portability

§ Helps TensorFlow Stay Both Flexible and Performant

XLA HIGH LEVEL OPTIMIZER (HLO)

§ Compiler Intermediate Representation (IR)§ Independent of Source and Target Language§ Define Graphs using HLO Operations§ XLA Step 1 Emits Target-Independent HLO § XLA Step 2 Emits Target-Dependent LLVM§ LLVM Emits Native Code Specific to Target § Supports x86-64, ARM64 (CPU), and NVPTX (GPU)

JIT COMPILER§ Just-In-Time Compiler§ Built on XLA Framework§ Goals:

§ Reduce memory movement – especially useful on GPUs§ Reduce overhead of multiple function calls

§ Similar to Spark Operator Fusing in Spark 2.0§ Unroll Loops, Fuse Operators, Fold Constants, …§ Scope to session, device, or `with jit_scope():`

VISUALIZING JIT COMPILER IN ACTION

Before After

Google Web Tracing Framework:http://google.github.io/tracing-framework/

from tensorflow.python.client import timelinetrace = timeline.Timeline(step_stats=run_metadata.step_stats)with open('timeline.json', 'w') as trace_file:trace_file.write(

trace.generate_chrome_trace_format(show_memory=True))

VISUALIZING FUSING OPERATORS

pip install graphviz

dot -Tpng \/tmp/hlo_graph_1.w5LcGs.dot \-o hlo_graph_1.png

GraphViz:http://www.graphviz.org

hlo_*.dot files generated by XLA





AOT COMPILER§ Standalone, Ahead-Of-Time (AOT) Compiler§ Built on XLA framework§ tfcompile§ Creates executable with minimal TensorFlow Runtime needed

§ Includes only dependencies needed by subgraph computation§ Creates functions with feeds (inputs) and fetches (outputs)

§ Packaged as cc_libary header and object files to link into your app§ Commonly used for mobile device inference graph

§ Currently, only CPU x86-64 and ARM are supported - no GPU

GRAPH TRANSFORM TOOL (GTT)

§ Optimize Trained Models for Inference§ Remove training-only Ops (checkpoint, drop out, logs)§ Remove unreachable nodes between given feed -> fetch§ Fuse adjacent operators to improve memory bandwidth§ Fold final batch norm mean and variance into variables§ Round weights/variables improves compression (ie. 70%)§ Quantize (FP32 -> INT8) to speed up math operations

BEFORE OPTIMIZATIONS

GRAPH TRANSFORM TOOLtransform_graph \--in_graph=tensorflow_inception_graph.pb \ ß Original Graph--out_graph=optimized_inception_graph.pb \ ß Transformed Graph--inputs='Mul' \ ß Feed (Input)--outputs='softmax' \ ß Fetch (Output) --transforms=' ß List of Transforms strip_unused_nodesremove_nodes(op=Identity, op=CheckNumerics) fold_constants(ignore_errors=true) fold_batch_normsfold_old_batch_normsquantize_weightsquantize_nodes'

AFTER STRIPPING UNUSED NODES

§ Optimizations§ strip_unused_nodes

§ Results§ Graph much simpler§ File size much smaller

AFTER REMOVING UNUSED NODES

§ Optimizations§ strip_unused_nodes§ remove_nodes

§ Results§ Pesky nodes removed§ File size a bit smaller

AFTER FOLDING CONSTANTS

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants

§ Results§ Placeholders (feeds) -> Variables*

(*Why Variables and not Constants?)

AFTER FOLDING BATCH NORMS

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms

§ Results§ Graph remains the same§ File size approximately the same

WEIGHT QUANTIZATION

§ FP16 and INT8 Are Computationally Simpler and Faster§ Weights/Variables are Constants§ Easy to Linearly Quantize

AFTER QUANTIZING WEIGHTS

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms§ quantize_weights

§ Results§ Graph is same, file size is smaller, compute is faster

BUT WAIT, THERE’S MORE!

ACTIVATION QUANTIZATION§ Activations Not Known Ahead of Time

§ Depends on input, not easy to quantize§ Requires Additional Calibration Step

§ Use a “representative” dataset§ Per Neural Network Layer…

§ Collect histogram of activation values§ Generate many quantized distributions with different saturation thresholds§ Choose threshold to minimize…

KL_divergence(ref_distribution, quant_distribution)§ Not Much Time or Data is Required (Minutes on Commodity Hardware)

AFTER ACTIVATION QUANTIZATION

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms§ quantize_weights§ quantize_nodes (activations)

§ Results§ Larger graph, needs calibration!

Requires additional freeze_requantization_ranges

FREEZING MODEL FOR DEPLOYMENT§ Optimizations

§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms§ quantize_weights§ quantize_nodes§ freeze_graph

§ Results§ Variables -> Constants

Finally!We’re Ready to Deploy!!





TENSORFLOW SERVING OVERVIEW§ Inference

§ Only Forward Propagation through Network§ Predict, Classify, Regress, …

§ Bundle§ GraphDef, Variables, Metadata, …

§ Assets§ ie. Map of ClassificationID -> String§ {9283: “penguin”, 9284: “bridge”}

§ Version§ Every Model Has a Version Number (Integer)

§ Version Policy§ ie. Serve Only Latest (Highest), Serve Both Latest and Previous, …

MULTI-HEADED INFERENCE

§ Multiple “heads” (aka “responses”) from 1 model prediction§ Optimizes bandwidth, CPU, latency, memory, coolness§ Response includes both class and scores § Inputs sent only once§ Feed scores into ensemble models§ Use model for feature engineering

REQUEST BATCHING§ max_batch_size

§ Enables throughput/latency tradeoff§ Bounded by RAM

§ batch_timeout_micros§ Defines batch time window, latency upper-bound§ Bounded by RAM

§ num_batch_threads§ Defines parallelism§ Bounded by CPU cores

§ max_enqueued_batches§ Defines queue upper bound, throttling§ Bounded by RAM

Reaching either thresholdwill trigger a batch

YOU JUST LEARNED…§ PipelineAI Features




THANKS! ANY QUESTIONS?§ https://github.com/PipelineAI/pipeline/

§ Please 🌟 this GitHub Repo!

§ All slides, code, notebooks, and Docker images here:https://github.com/PipelineAI/pipeline/tree/master/gpu

Contact [email protected]

@cfregly

Date post:	22-Jan-2018
Category:	Data & Analytics
Upload:	chris-fregly
View:	682 times
Download:	7 times

Building Google Cloud ML Engine From Scratch on AWS with PipelineAI - ODSC London 2017 - Oct 13,...

Data & Analytics