Distributed deep learning platforms can improve productivity for practitioners by enabling faster training through distributed computing. Hopsworks is a new data platform built on HopsFS that provides first-class support for Python, deep learning, machine learning, and strong data governance. It allows for distributed deep learning training and parallel experiments across GPU servers as well as secure streaming analytics using Kafka, Spark, and Flink.

1. End-to-End Platform Support for
Distributed Deep Learning in Finance
Jim Dowling
CEO, Logical Clocks AB
Assoc Prof, KTH Stockholm
Senior Researcher, RISE SICS
jim_dowling

2. Deep Learning in Finance
•Financial modelling problems are typically complex and
non-linear.
•If you’re lucky, you have lots of labelled data
-Deep learning models can learn non-linear relationships and
recurrent structures that generalize beyond the training data.
•Potential areas in finance: pricing, portfolio construction,
risk management and HFT*
2/33
* https://towardsdatascience.com/deep-learning-in-finance-9e088cb17c03

3. More Data means Better Predictions
Prediction
Performance
Traditional ML
Deep Neural Nets
Amount Labelled Data
Hand-crafted
can outperform
1980s1990s2000s 2010s 2020s?
3/33

4. Do we need more Compute?
“Methods that scale with computation
are the future of AI”*
- Rich Sutton (A Founding Father of Reinforcement Learning)
* https://www.youtube.com/watch?v=EeMCEQa85tw
4/33

5. Reduce DNN Training Time
In 2017, Facebook
reduced training
time on ImageNet
for a CNN from 2
weeks to 1 hour
by scaling out to
256 GPUs using
Ring-AllReduce on
Caffe2.
https://arxiv.org/abs/1706.02677
5/33

6. • Hyper-parameter optimization is
parallelizable
• Neural Architecture Search
(Google)
- 450 GPU / 7 days
- 900 TPU / 5 days
- New SOTA on CIFAR10 (2.13% top 1)
- New SOTA on ImageNet (3.8% top 5)
Reduce Experiment Time with Parallel Experiments
https://arxiv.org/abs/1802.01548
6/33

7. Training Time and ML Practitioner Productivity
7
•Distributed Deep Learning
-Interactive analysis!
-Instant gratification!
“My Model’s Training.”
Training
7/33

8. More Compute should mean Faster Training
Training
Performance
Single-Host
Distributed
Available Compute
20152016 2017 2018?
8/33

9. Distributed Training: Theory and Practice
9 9/33
Image from @hardmaru on Twitter.

10. Distributed Training Algorithms not all Equal
Training
Performance
Parameter Servers
AllReduce
Available Compute
10/33

11. Ring-AllReduce vs Parameter Server
GPU 0
GPU 1
GPU 2
GPU 3
send
send
send
send
recv
recv
recv
recv GPU 1 GPU 2 GPU 3 GPU 4
Param Server(s)
Network Bandwidth is the Bottleneck for Distributed Training
11/33

12. AllReduce outperforms Parameter Servers
12/33
*https://github.com/uber/horovod
16 servers with 4 P100 GPUs (64 GPUs) each connected by ROCE-capable 25 Gbit/s network
(synthetic data). Speed below is images processed per second.*
For Bigger Models, Parameter Servers don’t scale

13. Infiniband for Training to overcome N/W Bottleneck
RDMA/Infiniband
Read Input Files, Write Model Checkpoints to Network FS
Aggregate Gradients
Separate Gradient sharing/aggregation network traffic from I/O traffic.
13/33

14. Horovod on Hops
import horovod.tensorflow as hvd
def conv_model(feature, target, mode)
…..
def main(_):
hvd.init()
opt = hvd.DistributedOptimizer(opt)
if hvd.local_rank()==0:
hooks = [hvd.BroadcastGlobalVariablesHook(0), ..]
…..
else:
hooks = [hvd.BroadcastGlobalVariablesHook(0), ..]
…..
from hops import allreduce
allreduce.launch(spark, 'hdfs:///Projects/…/all_reduce.ipynb')
“Pure” TensorFlow code
14/33

15. Parallel Experiments

16. Parallel Experiments on Hops
def model_fn(learning_rate, dropout):
import tensorflow as tf
from hops import tensorboard, hdfs, devices
[TensorFlow Code here]
from hops import experiment
args_dict = {'learning_rate': [0.001, 0.005, 0.01],
'dropout': [0.5, 0.6]}
experiment.launch(spark, model_fn, args_dict)
Launch TF jobs in Spark Executors
17/33
Launches 6 Spark Executors with a different Hyperparameter
combinations. Each Executor can have 1-N GPUs.

17. Parallel Experiments Visualization on TensorBoard
18/33
Parallel Experiment Results Visualization

18. Lots of good GPUs > A few great GPUs
100 x Nvidia 1080Ti (DeepLearning11)
8 x Nvidia V100 (DGX-1)
VS
Both top (100 GPUs) and bottom (8 GPUs) cost the same: $150K (March 2018).
19/33

19. Share GPUs to Maximize Utilization
GPU Resource Management (Hops, Mesos)
20/33
4 GPUs on any host
10 GPUs on 1 host
100 GPUs on 10 hosts with ‘Infiniband’
20 GPUs on 2 hosts with ‘Infiniband_P100’

20. DeepLearning11 Server $15K (10 x 1080Ti)
21/33

21. Economics of GPUs and the Cloud
Time
GPU
Utilization
On-Premise GPU
Cloud
DeepLearning11 (10x1080Tis) will pay for itself in 11 weeks,
compared to using a p3.8xlarge in AWS
22/33

22. Distributed Deep Learning for Finance
•Platform for Hyperscale Data Science
•Controlled* access to datasets
*GDPR-compliance, Sarbanes-Oxley, etc
23/33

23. Hopsworks

24. Hops: Next Generation Hadoop*
16x
Throughput
FasterBigger
*https://www.usenix.org/conference/fast17/technical-sessions/presentation/niazi
37x
Number of files
Scale Challenge Winner (2017)
25
GPUs in
YARN
25/33

25. Hopsworks Data Platform
Develop Train Test Serve
MySQL Cluster
Hive
InfluxDB
ElasticSearch
KafkaProjects,Datasets,Users
HopsFS / YARN
Spark, Flink, Tensorflow
Jupyter, Zeppelin
Jobs, Kibana, Grafana
REST
API
Hopsworks
26/33

26. Proj-42
Projects sandbox Private Data
A Project is a Grouping of Users and Data
Proj-X
Shared TopicTopic /Projs/My/Data
Proj-AllCompanyDB
Ismail et al, Hopsworks: Improving User Experience and Development on Hadoop with Scalable, Strongly Consistent Metadata, ICDCS 2017
27/33

27. How are Projects used?
Engineering
Kafka Topic
FX Project
FX Topic
FX DB
FX Data Stream
Shared Interactive Analytics
FX team
28/33

28. Per-Project Python Envs with Conda
Python libraries are usable by Spark/Tensorflow
29/33

29. HopsFS
YARN
FeatureStore
Tensorflow
Serving
Public Cloud or On-Premise
Tensorboard
TensorFlow in Hopsworks
Experiments
Kafka
Hive
30/33

30. One Click Deployment of TensorFlow Models
31/33

31. Hops API
•Python/Java/Scala library
-Manage tensorboard, Load/save models in HDFS
-Horovod, TensorFlowOnSpark
-Parameter sweeps for parallel experiments
-Neural Architecture Search with Genetic Algorithms
-Secure Streaming Analytics with Kafka/Spark/Flink
• SSL/TLS certs, Avro Schema, Endpoints for Kafka/Hopsworks/etc
32/33

32. Deep Learning Hierarchy of Scale
DDL
AllReduce
on GPU Servers
DDL with GPU Servers
and Parameter Servers
Parallel Experiments on GPU Servers
Single GPU
Many GPUs on a Single GPU Server
33/33

33. Summary
•Distribution can make Deep Learning practitioners more
productive.
https://www.oreilly.com/ideas/distributed-tensorflow
•Hopsworks is a new Data Platform built on HopsFS with
first-class support for Python / Deep Learning / ML /
Strong Data Governance

34. The Team
Jim Dowling, Seif Haridi, Tor Björn Minde, Gautier Berthou, Salman
Niazi, Mahmoud Ismail, Theofilos Kakantousis, Ermias
Gebremeskel, Antonios Kouzoupis, Alex Ormenisan, Fabio Buso,
Robin Andersson, August Bonds, Filotas Siskos, Mahmoud Hamed.
Active:
Alumni:
Vasileios Giannokostas, Johan Svedlund Nordström,Rizvi Hasan, Paul Mälzer, Bram
Leenders, Juan Roca, Misganu Dessalegn, K “Sri” Srijeyanthan, Jude D’Souza, Alberto
Lorente, Andre Moré, Ali Gholami, Davis Jaunzems, Stig Viaene, Hooman Peiro,
Evangelos Savvidis, Steffen Grohsschmiedt, Qi Qi, Gayana Chandrasekara, Nikolaos
Stanogias, Daniel Bali, Ioannis Kerkinos, Peter Buechler, Pushparaj Motamari, Hamid
Afzali, Wasif Malik, Lalith Suresh, Mariano Valles, Ying Lieu, Fanti Machmount Al
Samisti, Braulio Grana, Adam Alpire, Zahin Azher Rashid, ArunaKumari Yedurupaka,
Tobias Johansson , Roberto Bampi.
www.hops.io
@hopshadoop