Asynchronous Stochastic Gradient Descent on GPU: Is It ...faculty.ucmerced.edu/frusu/Talks/2017-06-google-gd-gpu.pdf · Asynchronous Stochastic Gradient Descent on GPU: Is It Really

Asynchronous Stochastic Gradient Descent on GPU: Is It Really Better than CPU?

Florin Rusu

Yujing Ma, Martin Torres (Ph.D. students)

University of California Merced

Machine Learning (ML) Boom

• Two SIGMOD 2017 tutorials

ML Systems

General purpose (databases) • BIDMach • Bismarck • Cumulon • DeepDive • DimmWitted • GLADE • GraphLab • MADlib • Mahout • MLlib (MLbase) • SimSQL (BUDS) • SystemML • Vowpal Wabbit • …

Deep learning

• Caffe (con Troll)

• CNTK

• DL4J

• Keras

• MXNet

• SINGA

• TensorFlow

• Theano

• Torch

• …

ML Hardware Accelerators

ML Systems with GPU Acceleration

General purpose • BIDMach • Bismarck • Cumulon • DeepDive • DimmWitted • GLADE • GraphLab • MADlib • Mahout • MLlib (MLbase) • SimSQL (BUDS) • SystemML • Vowpal Wabbit • …

Deep learning

• Caffe

• CNTK

• DL4J

• Keras

• MXNet

• SINGA

• TensorFlow

• Theano

• Torch

• …

ML in Databases

• It is not so much about deep learning – Regression (linear, logistic) – Classification (SVM) – Recommendation (LMF)

• Mostly about training – Inside DB, close to data – Over joins or factorized databases – Compressed data, (compressed) large models

• Selection of optimization algorithm and hyper-parameters – BGD vs. SGD vs. SCD

Classification Tasks

• Logistic regression (LR)

• Support Vector Machines (SVM)

Datasets and Platforms

S. Sallinen et al: “High Performance Parallel Stochastic Gradient Descent in Shared Memory” in IPDPS 2016.

• CPU: Intel Xeon E5-2660 (14 cores, 28 threads)

• GPU: Tesla K80 (use only one multiprocessor)

Experiments • Stochastic gradient descent (SGD) optimizer: mini-batch with 4096 batch size • Average time per iteration over 100 iterations (measure only the iteration time) • TensorFlow and MXNet support only dense data: covtype and w8a are “densified”;

others do not fit in GPU memory

LR

SVM

Research Questions

• Why is GPU not significantly better than CPU on LR and SVM models?

– The gain in deep nets seems to come mostly from convolutions, not gradient computations

– SparseMatrix-Vector (SpMV) and SparseMatrix-Matrix (SpMM) are harder to optimize

• Can we improve the GPU performance?

Gradient Descent

(Mini-)Batch Gradient Descent (BGD)

Parallel BGD • Parallel execution on CPU and GPU

• Synchronous execution on CPU

Stochastic Gradient Descent (SGD)

Parallel SGD (Hogwild)

• No synchronization or locks

BGD vs. SGD

GPU Architecture Tesla K80 (GK210) • # MP = 13 • # cores/MP = 192 • # warps/MP = 64 • # blocks/MP = 16 • # threads/MP = 2048 • # threads/warp (SIMD) = 32 • # threads/block = 1024

• # registers/MP = 217

• # registers/block = 216 • # registers/thread = 255

• Shared mem/MP = 112KB • Shared mem/block = 48KB

• L1 cache = 48KB • Read-only texture = 48KB • L2 cache = 1.5MB • Global mem = 12GB

Map Hogwild to GPU

Algorithm 1. Copy data and model to GPU 2. While not converge do

1. Execute kernel update_model that implements Hogwild

3. End while

Design Space

Data access • Storage scheme

• Row-store • Column-store

• Partitioning • Round-robin • Chunking

Data replication • Number of threads

accessing an example • 1-way • K-way

Model replication • Where is model stored on GPU

memory hierarchy • Per thread (registers) • Per block (shared memory) • Per kernel (global memory)

Evaluation Metrics

• DimmWitted by Zhang and Re in PVLDB 2014

• Hardware efficiency

– Time to convergence

• Statistical efficiency

– Number of iterations to convergence

Data Access – Storage Scheme

Data Access – Partitioning

Evaluation – Dense Data (covtype)

Hardware efficiency Statistical efficiency

Evaluation – Sparse Data (news)

Hardware efficiency Statistical efficiency

Data Replication

Evaluation – Dense Data (covtype)

Hardware efficiency Statistical efficiency

Evaluation – Sparse Data (news)

Hardware efficiency Statistical efficiency

Model Replication – PerThread

Model Replication – PerBlock

Model Replication – PerKernel

Evaluation – Dense Data (covtype)

Hardware efficiency Statistical efficiency

Evaluation – Sparse Data (news)

Hardware efficiency Statistical efficiency

Comparison with Synchronous SGD

CPU vs. GPU

Conclusions

• Synchronous mini-batch in deep learning systems is rarely faster in convergence on GPU than on CPU

• Asynchronous SGD on GPU is always faster in time per iteration than synchronous mini-batch on GPU

• Asynchronous SGD on GPU is sometimes faster in convergence than asynchronous SGD on CPU

Thank you.

Questions ???