Diese Präsentation wurde erfolgreich gemeldet.
Wir verwenden Ihre LinkedIn Profilangaben und Informationen zu Ihren Aktivitäten, um Anzeigen zu personalisieren und Ihnen relevantere Inhalte anzuzeigen. Sie können Ihre Anzeigeneinstellungen jederzeit ändern.

AWSでの機械学習におけるデータレイク・GPU実行環境

329 Aufrufe

Veröffentlicht am

ABEJA Cloud AI Night in GTC2018で使用した、AWSでの機械学習におけるデータレイク・GPU実行環境についての発表資料です。

Veröffentlicht in: Software
  • Als Erste(r) kommentieren

AWSでの機械学習におけるデータレイク・GPU実行環境

  1. 1. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. CA . E A 1A E A C8 E 8E A 1 CG 8 , B , A E 2 1 M a ST NK P WJUY I 03
  2. 2. yr cVsf 3J EGA I GSNE M )JCIENJ J EIKJEIN W w aVd b l k )GJO ,LJIN n V ) JONA & ELA N )JIIA N 3 )GJO L EG )GJO . 3 )JIBEC ypi )GJO N D LOMNA PEMJL mw G MNE (A IMN GF )GJO ,JLH NEJI 5KM JLFM iV u f 3 I CAHAIN )JIMJGA )2 v M P . SNDJI O S Vr u Ve x s l hV e aVd ( 01 2 b tzVn ) ONJ GEIC 2 H G MNE 2J ( G I EIC ) )JIN EIAL ALPE A iVw g w JLF K AM JLF J M JLF3 EG fn Ve ( G EAL , NJL CA NAQ S mVixVf SI HJ ( A MDEBN G MNE) DA G MNE 3 K A O A EIAMEM NDAI N EKAGEIA w aVd cVsf KK NLA H G MNE MA L D , m n ELA NJLS ALPE A bVo )J A AKGJS )J A)JHHEN )J A EKAGEIA ) 8 H TJI 2AR 3 DEIA 2A LIEIC JGGS AFJCIENEJI ,https://aws.amazon.com/jp/products/
  3. 3. * ,8 6S a ( A S A 8 C AS ) , +02 01 M 8 W 6a a h S6 hf ( cf d a a a a a A 8 aA 6 6 aS W h 6 6 a 54 A W G ) , +02 01 New *4 20 3 1 ) https://aws.amazon.com/jp/about-aws/global-infrastructure/
  4. 4. 41 - ( ) https://aws.amazon.com/jp/compliance/pci-data-privacy-protection-hipaa-soc-fedramp-faqs/
  5. 5. W* S A ( ) A
  6. 6. 1 S * 1 S A ) W 0( A ) 1 10( 7
  7. 7. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
  8. 8. 1 5
  9. 9. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
  10. 10. W TS • X • X • • • ce • 13 • A • • • h • X • ce • X • •
  11. 11. 14
  12. 12. Batch / Real time 15
  13. 13. 20 Batch / Real time 16
  14. 14. Batch / Real time AWS AWS ETL 17
  15. 15. Batch / Real time AWS AWS ETL Amazon Mechanical TurkAWS ML 18
  16. 16. Batch / Real time 19 AI/ML/DL
  17. 17. Batch / Real time 20
  18. 18. Batch / Real time 21
  19. 19. Batch / Real time AWS AWS ETL Amazon Mechanical TurkAWS ML AWS 22
  20. 20. 23 Batch / Real time AWS AWS ETL Amazon Mechanical TurkAWS ML AWS
  21. 21. Amazon Kinesis AWS Glue Amazon S3 Amazon Glacier Amazon ECS Amazon EC2 Lambda function Amazon DynamoDB Amazon RDS Amazon Redshift Amazon Athena Amazon ES Amazon EMR Amazon QuickSight Amazon Kinesis Analytics Amazon Kinesis Streams Amazon SageMaker Amazon Machine Learning AWS Batch AWS IoT AWS Greengrass Lambda function Ingest Store Compute Database Analyze Intelligence The Edge AWS
  22. 22. 3F RG W A B RG S LI B 3 )3B D3E M Q C ( B ) C RG IGK T ) 3 B 3 RG 2 C B RG A A 25
  23. 23. Redshift EMR (Yarn/Spark) Aggregator Model Development &Training / Amazon Athena Realtime Event Processor Agg. Cached Data Aggregator Realtime Search Kinesis Analytics Decision on the Cloud Feedback Aggregator Quicksight Kinesis Streams Kinesis Firehose Amazon S3 AWS Greengrass Lambda function AWS Glue AWS IOT ML Service
  24. 24. 27 https://robotstart.info/2018/03/27/toyota-aws.html
  25. 25. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
  26. 26. Batch / Real time AWS ML 29
  27. 27. Services Amazon Rekognition Image Platform Engines TensorFlow Caffe Torch Theano CNTK Keras CPU IoT MobileInfrastructure GPU MXNet AWS Deep Learning AMI (Amazon Linux / Ubuntu / Windows) Amazon Polly Amazon Lex Amazon Comprehend Amazon Translate Amazon Transcribe Amazon Rekognition Video Amazon Machine Learning Apache Spark Amazon EMR Amazon Kinesis AWS Batch Amazon ECS Amazon SageMaker 30
  28. 28. Services Amazon Rekognition Image Platform Engines TensorFlow Caffe Torch Theano CNTK Keras CPU IoT MobileInfrastructure GPU MXNet AWS Deep Learning AMI (Amazon Linux / Ubuntu / Windows) Amazon Polly Amazon Lex Amazon Comprehend Amazon Translate Amazon Transcribe Amazon Rekognition Video Amazon Machine Learning Apache Spark Amazon EMR Amazon Kinesis AWS Batch Amazon ECS Amazon SageMaker 31
  29. 29. 8 2 563 1 0 72 1 1 1 1 7 2 11 0 72 2017/10/26
  30. 30. N U UG UG • 0 1- A 3 a - 0 N I D • V Ad 3 N U P T AV l e • U b e NVIDIA Roadmap GTC 2017 33 https://aws.amazon.com/jp/ec2/instance-types/p3/
  31. 31. 34
  32. 32. (8 A: ) 1 > 03 ( 12 DCE GI • 1 3: > 1 > DCE GI 1 > 3 3: > DCE GI 0 20 40 60 80 100 120 140 K80 P100 V100 Mixed/FP16 Perf (TFLOPS) 0 2 4 6 8 10 12 14 16 K80 P100 V100 FP32 Perf (TFLOPS) 0 1 2 3 4 5 6 7 8 K80 P100 V100 FP64 Perf (TFLOPS) 0 1000 2000 3000 4000 5000 6000 K80 P100 V100 Resnet-50 8 GPU (Images/sec) 1 0 2 .
  33. 33. https://speakerdeck.com/ayemos/powering-by-aws-gpu-instances-in-cookpad-inc
  34. 34. https://speakerdeck.com/rtechkouhou/hua-xiang-chu-li-ji-jie-xue-xi-hefalsep3huo-yong
  35. 35. 3 38
  36. 36. 39 https://aws.amazon.com/machine-learning/amis/
  37. 37. 40 CUDA9, nvidia-docker AMI nvidia-docker NVIDIA GPU Cloud Docker Docker https://ngc.nvidia.com https://aws.amazon.com/marketplace/pp/B076K31M1S http://docs.nvidia.com/ngc/ngc-aws-setup-guide/index.html
  38. 38. C 5 Instance Name vCPUs RAM EBS Bandwidth Network Bandwidth c5.large 2 4 GiB Up to 2.25 Gbps Up to 10 Gbps c5.xlarge 4 8 GiB Up to 2.25 Gbps Up to 10 Gbps c5.2xlarge 8 16 GiB Up to 2.25 Gbps Up to 10 Gbps c5.4xlarge 16 32 GiB 2.25 Gbps Up to 10 Gbps c5.9xlarge 36 72 GiB 4.5 Gbps 10 Gbps c5.18xlarge 72 144 GiB 9 Gbps 25 Gbps 41 https://aws.amazon.com/jp/ec2/instance-types/c5/
  39. 39. / 42 • FM I S L N X • G U A • N cS P a W • X Trained models and Lambdas Extracted IntelligenceInferences and take actions locally on device AWS Cloud for training PREVIEW AVAILABLE https://aws.amazon.com/jp/greengrass/ml/
  40. 40. 43 TuSimple’s deep learning requirements for training versus implementation are very different. The models are created and trained in a multiple-GPU- based Amazon Web Services cloud environment. https://www.oreilly.com/ideas/self-driving-trucks-enter-the-fast-lane-using-deep-learning
  41. 41. 228 2 • e WL L e S • 22 1 nw P GM p f eT AN S s • ,L FG e o t r S l D C • e W g F 2 1 2 0 2 Xa 44 https://aws.amazon.com/jp/deeplens/
  42. 42. • O O D S e • A O • W 45

×