[225]yarn 기반의 deep learning application cluster 구축 김제민
32 likes6,241 views
[225]yarn 기반의 deep learning application cluster 구축 김제민
![[225]yarn 기반의 deep learning application cluster 구축 김제민](https://image.slidesharecdn.com/225yarndeeplearningapplicationcluster-161025031031/85/225-yarn-deep-learning-application-cluster-5-320.jpg)
![• dl_app_caffe.properties
[application]
username=tony
appname='c3 deeplearning app caffe mnist example'
docker_image=naver/c3_dl-caffe
user_dev_workspace_path=./user_dev_workspace
user_shell_script=dl_app_start.sh
user_shell_args='-gpu all'
[from_hdfs]
hdfs=/user/tony/caffe_test/*
input_path=data
[to_hdfs]
output_path=output
hdfs=/user/tony/caffe_mnist_example_output
overwrite=true
• dl_app_start.sh
#!/usr/bin/env bash
./create_mnist.sh
./train_lenet.sh $@
dlapp-launcher 실행 예
.
├── dl_app_caffe.properties
└── user_dev_workspace
├── create_mnist.sh
├── data/
├── dl_app_start.sh
├── examples/
├── lenet_solver.prototxt
├── lenet_train_test.prototxt
├── output/
└── train_lenet.sh](https://image.slidesharecdn.com/225yarndeeplearningapplicationcluster-161025031031/85/225-yarn-deep-learning-application-cluster-83-320.jpg)
![dlapp-launcher 실행 예
$ dlapp-launcher dl_app_caffe.properties
C3 DL Cluster Version check : ok
CheckDLAppProperties : ok
user_dev_workspace_path = /path/to/user_dev_workspace
user_dev_workspace_exclude_list = []
user_dev_workspace dir size = 4 kb
DL App Submission : OK
DL App ID = application_aaaaaaaaaaa_bbbb
DL App Dashboard
* While app is running , check : http://XXXXXXXXX/cluster/app/application_aaaaaaaaaaa_bbbb
* After app is completed, check : http://XXXXXXXXX/applicationhistory/app/application_aaaaaaaaaaa_bbbb
DL App Status and stdout/stderr Log URLs
* dlapp-status application_aaaaaaaaaaa_bbbb](https://image.slidesharecdn.com/225yarndeeplearningapplicationcluster-161025031031/85/225-yarn-deep-learning-application-cluster-84-320.jpg)
![DL App Tools
• dlapp-status [ DL App ID ]
• dlapp-list [Username]
• dlapp-softkill [Username] [ DL App ID ]
• dlapp-kill [Username] [ DL App ID ]](https://image.slidesharecdn.com/225yarndeeplearningapplicationcluster-161025031031/85/225-yarn-deep-learning-application-cluster-85-320.jpg)
![DL App Toolset
• dlapp-launcher dl_app.properties
• dlapp-shell dl_app.properties
• dlapp-run
• dlapp-status [DL App ID]
• dlapp-list [Username]
• dlapp-softkill [Username] [DL App ID]
• dlapp-kill [Username] [DL App ID]](https://image.slidesharecdn.com/225yarndeeplearningapplicationcluster-161025031031/85/225-yarn-deep-learning-application-cluster-98-320.jpg)
[225]yarn 기반의 deep learning application cluster 구축 김제민
- 1. YARN, Docker 기반의 Deep Learning Application Cluster 구축 김제민 NAVER Search
- 2. 하고 있고, 해 왔던 것들
- 3. Naver Search • 검색 모델링 • 검색 데이터 정제 • 검색 서비스 • 대용량 데이터 처리 • 분산 데이터 처리 • 실시간 데이터 처리
- 4. C3 Hadoop Cluster 소개 • Apache Ambari • Cluster 배포/관리/모니터 • Hadoop 2.7.1 • Hive, Spark, HBase, Oozie etc ..
- 6. Deep Learning • Machine Learning 기법 중의 하나 • Artificial Neural Network 인데 Deep함 < AlexNet > • 60 million parameters and 650,000 neurons From “ImageNetClassification with Deep Convolutional Neural Networks”
- 7. Deep Learning Breakthrough • 학습의 어려움 극복 • 2006, “A fast learning algorithm for deep belief nets” by Geoffrey E. Hinton • Hardware • GPU 성능의 발전 • NVIDIA CUDA
- 8. Deep Learning 은 거의 모든 분야에서 절대적으로 성능이 우월함
- 9. 이제 Deep Learning 으로 Go, Go !!!
- 10. • 다수의 프로젝트 • 다수의 연구원/개발자 But • GPU 자원은 제한적
- 12. GPU 자원의 효과적인 공유 필요성 ? GPU Cluster
- 13. Deep Learning Application 개발 환경 • GPU • NVIDIA CUDA • Deep Learning Frameworks • Caffe • TensorFlow • Torch • Theano • Keras
- 14. 그래서 결론적으로 지금 우리에게 필요한 것은 …
- 15. Multi-tenantDeep Learning Application 실행 환경 • 다양한 Deep Learning Framework 지원 • GPU Cluster 자원 관리 ? GPU Cluster DL Application DL Application DL Application
- 16. Hadoop YARN 기반의 Multi-tenantDeep Learning Application 실행 환경 ? GPU Cluster Hadoop YARN DL Application DL Application DL Application
- 17. Why YARN ? • Mesos ? • YARN의 장점 • Capacity Scheduler : Queue기반의 ResourceScheduling • Enterprise환경에서 ResourcePlanning 용이 • 기타 • 기존 클러스터(C3)와의 호환성 및 통합 가능성 • Cluster 배포 및 관리 시스템 기존재 (Ambari) • YARN Cluster 운영 Knowhow
- 18. 목차 1. Technical Issues 2. Deep Learning Application Toolset 개발 3. C3 Deep Learning Cluster 구축
- 20. Technical Issues 1) YARN Cluster에서 Shell Script Application 실행 2) 다양한 Deep Learning Framework 환경 제공 3) YARN Cluster에서 GPU Resource Scheduling
- 21. 1. 1) YARN Cluster 에서 Shell Script Application 실행
- 22. Shell Script Application on YARN ? YARN Cluster CPU Core 1 Memory 1G Main main.sh dl_app.tar.gz data/ examples/ output/ main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh data/ examples/ output/ ü main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh CPU Core 1 Memory 1G tar DL Application
- 23. Hadoop YARN • Hadoop YARN ( Hadoop v2) • Yet Another ResourceNegotiator • MapReduce이외에 일반적인 Application 실행 가능 • CPU,Memory 기반의 ResourceScheduling • YARN 구조 • ResourceManager ( Master) • Node Manager ( Slaves ) • 최소 실행 단위 : Container
- 24. YARN 구조 Resource Manager (Master) Node Manager (Slave) Server Container Container Container Container NM NM
- 25. YARN Application Resource Manager (Master) NM Application Master Container NM NM Container Request
- 26. YARN Application Resource Manager (Master) NM Application Master Container Container NM NM Container Container
- 27. Shell Script Application on YARN ? YARN Cluster CPU Core 1 Memory 1G Main main.sh dl_app.tar.gz data/ examples/ output/ main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh data/ examples/ output/ ü main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh CPU Core 1 Memory 1G tar DL Application
- 28. YARN Application 개발하기 • ApplicationMaster ResourceManager • org.apache.hadoop.yarn.client.api.async.AMRMClientAsync • InterfaceAMRMClientAsync.CallbackHandler • ApplicationMaster NodeManager • org.apache.hadoop.yarn.client.api.async.NMClientAsync • InterfaceNMClientAsync.CallbackHandler
- 29. YARN Distributed Shell (Example) • https://github.com/apache/hadoop-common/tree/trunk/hadoop-yarn-project/hadoop-yarn/hadoop-yarn- applications/hadoop-yarn-applications- distributedshell/src/main/java/org/apache/hadoop/yarn/applications/distributedshell • ApplicationMaster.java • Client.java • DSConstants.java • Log4jPropertyHelper.java
- 30. C3 Distributed Shell • Yarn Distributed Shell 기반 • 추가 기능 • CONTAINTER_INDEX, NUM_CONTAINERS 환경 변수 • YARN-5696 • Server Node 에 따른 Resource요청 • 특정 Server Node 에 Resource요청 • Server Node Blacklist설정 • YARN-4703
- 31. Technical Issues ü YARN Cluster에서 Shell Script Application 실행 2) 다양한 Deep Learning Framework 환경 제공 3) YARN Cluster에서 GPU Resource Scheduling è C3 Distributed Shell
- 32. 1. 2) 다양한 Deep Learning Framework 환경 제공
- 33. Classical Approach • 모든 DL Framework를 모든 서버에 설치 • Installation, Upgrade, Update 등 유지 보수의 어려움 • 같은 DL Framework 에서 Multi Version 지원 이슈 • 서로 상이한 DL Framework간의 Library 의존성 충돌 • DevOps Tools : • Dimensions • ( DL Framworks ) * ( Versions ) * ( 의존 libs ) * ( Server 환경)
- 34. 각 Deep Learning Framework 환경들이 깔끔히 Isolation되어 관리되면 좋을텐데..
- 35. Docker ?
- 36. 그렇지만 GPU는 ? CUDA는 ? Docker 에서 가능한가?
- 37. Docker 에서 CUDA 사용하기 • Docker Image에 NVIDIA User Level Libraries 와 CUDA Library 를 설치 • 제약 : Host에 설치된 NVIDIA Driver Major,Minor Version과 일치 해 야함 • Docker 실행시 NVIDIA Device File 들을 Docker에 노출 $ docker run … --device=/dev/nvidiactl --device=/dev/nvidia-uvm --device=/dev/nvidia
- 38. Docker 에서 CUDA 사용하기 Host Server NVIDIA Kernel Module 361.48 User Level NVIDIA Libs 361.48 NVIDIA Driver Version Major 361 Minor Docker Container(Image) User Level NVIDIA Libs 361.48 GPU Application CUDA Libs /dev/nvidiactl /dev/nvidia-uvm /dev/nvidia
- 40. Nvidia-Docker !!
- 41. Nvidia-Docker • Docker 상에서 GPU Device를 사용시 발생하는 Host 의존성을 해결 • GPU Resource Isolation 지원 !!
- 42. Nvidia-Docker Host Server NVIDIA Kernel Module 361.48 User Level NVIDIA Libs 361.48 Docker Container(Image) GPU Application CUDA Libs volume User Level NVIDIA Libs 361.48
- 43. Nvidia-Docker 실행 • 기본 실행 • nvidia-docker <docker-options> <docker-command> <docker- args> • GPU isolation • NV_GPU 환경 변수에 GPU Device ID 목록을 설정 • NV_GPU='0,1' nvidia-docker <docker-options> <docker-command> … • Docker Image 제약 • Nvidia-docker 용으로 미리 build된 image 사용
- 44. Nvidia-docker 기반으로 Deep Learning Framework 환경을 Docker Image로 제공
- 45. Technical Issues ü YARN Cluster에서 Shell Script Application 실행 ü 다양한 Deep Learning Framework 환경 제공 3) YARN Cluster에서 GPU Resource Scheduling è C3 Distributed Shell è Nvidia-Docker
- 46. 1. 3) YARN Cluster에서 GPU Resource Scheduling
- 47. YARN GPU Scheduling 현상황 • 현재 YARN 은 CPU와 Memory 만으로 스케쥴링 • GPU Scheduling은 지원하지 않음 • 새로운 자원 형태를 Scheduling 하기 위해서는 부가적인 개발이 필요함
- 48. YARN GPU Scheduling 개발 주요 Source Code 변경 대상 • YARN Protocol • YARN Configuration • Resource 추상화 객체 • Dominant Resource Calculator *YARN JIRA : YARN-5517
- 49. YARN Protocol • yarn_protos.proto • yarn_service_protos.proto Implementation Details • yarn_protos.proto message ResourceProto { optional int32 memory = 1; optional int32 virtual_cores = 2; optional int32 gpu_cores = 3; } enum SchedulerResourceTypes { MEMORY = 0; CPU = 1; GPU = 2; } • yarn_service_protos.proto Implementation Details • yarn_protos.proto message ResourceProto { optional int32 memory = 1; optional int32 virtual_cores = 2; optional int32 gpu_cores = 3; } enum SchedulerResourceTypes { MEMORY = 0; CPU = 1; GPU = 2; } • yarn_service_protos.proto
- 50. YARN Configuration • org.apache.hadoop.yarn.conf.YarnConfiguration Implementation Details • org.apache.hadoop.yarn.conf.YarnConfiguration public static final String RM_SCHEDULER_MINIMUM_ALLOCATION_GCORES = YARN_PREFIX + "scheduler.minimum-allocation-gcores"; public static final int DEFAULT_RM_SCHEDULER_MINIMUM_ALLOCATION_GCORES = 0; public static final String RM_SCHEDULER_MAXIMUM_ALLOCATION_GCORES = YARN_PREFIX + "scheduler.maximum-allocation-gcores"; public static final int DEFAULT_RM_SCHEDULER_MAXIMUM_ALLOCATION_GCORES = 8; public static final String NM_GCORES = NM_PREFIX + "resource.gcores"; public static final int DEFAULT_NM_GCORES = 0;
- 51. Resource 객체 • org.apache.hadoop.yarn.util.resource.Resource Implementation Details • org.apache.hadoop.yarn.util.resource.Resources public static Resource createResource(int memory) { return createResource(memory, (memory > 0) ? 1 : 0, 0); } public static Resource createResource(int memory, int cores) { return createResource(memory, cores, 0); } public static Resource createResource(int memory, int cores, int gcores) { Resource resource = Records.newRecord(Resource.class); resource.setMemory(memory); resource.setVirtualCores(cores); resource.setGpuCores(gcores); return resource; } public int getGpuCores() { return 0; } public void setGpuCores(int gcores) { throw new RuntimeException("NONE cannot be modified!"); } public int compareTo(Resource o) { int diff = 0 - o.getMemory(); if (diff == 0) { diff = 0 - o.getVirtualCores(); if (diff == 0) { diff = 0 - o.getGpuCores(); } } return diff; }
- 52. Dominant Resource Calculator • org.apache.hadoop.yarn.util.resource.DominantResourceCalculator
- 53. YARN UI 변화 UI Changes * 테스트용 클러스터 화면 캡쳐
- 54. YARN Cluster CPU Core 1 Memory 1G GPU 3 Main main.sh tar dl_app.tar.gz data/ examples/ output/ main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh data/ examples/ output/ ü main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh CPU Core 1 Memory 1G GPU 3 C3 Distributed Shell : GPU 추가
- 55. Technical Issues ü YARN Cluster에서 Shell Script Application 실행 ü 다양한 Deep Learning Framework 환경 제공 ü YARN Cluster에서 GPU Resource Scheduling è C3 Distributed Shell è Nvidia-Docker è YARN with GPU Scheduling 개발
- 56. 2. Deep Learning Application Launcher 및 Toolset 개발
- 57. Deep Learning Application Toolset 개발 • GPU Device ID Manager • DL(Deep Learning) App Concept • DL App Launcher • DL App Tools • DL App Log • DL App 개발 환경
- 58. GPU Device ID 할당 문제점 • 문제점 : YARN은 GPU 갯수만 관리함 • Deep Learning Application에 명시적으로 GPU Device ID를 할당하는 로직 필요
- 59. GPU Device ID Manager • Cluster 전체의 GPU Device ID Allocation 정보를 Zookeeper 를 통해 관리 • GPU Device ID Manager • python kazoo lib • 분산 Lock • GPU Device ID 할당/해제 • Garbage Collection { ”host01": { "gpu0": "application_id_01", "gpu1": “application_id_01”, "gpu2": “application_id_02”, "gpu3": “application_id_03” }, “host02": { "gpu0": “application_id_04", … }, … }
- 60. DL(Deep Learning) App 구현 Concept • DL App 구성 요소 • Deep Learning Framework 환경 (Caffe, TensorFlow …) • User Program ( Source Code) • Input Data • Output Data Input Output DL Framework 환경 Source Code
- 61. YARN Cluster CPU Core 1 Memory 1G GPU 3 Main main.sh tar dl_app.tar.gz data/ examples/ output/ main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh data/ examples/ output/ ü main.sh create_mnist.sh lenet_solver.prototxt lenet_train_test.prototxt train_lenet.sh CPU Core 1 Memory 1G GPU 3 C3 Distributed Shell C3 Distributed Shell User Program ( Source Code) Input Data ?
- 62. DL(Deep Learning) App 구현 Concept • DL App 구성 요소 • Deep Learning Framework 환경 (Caffe, TensorFlow …) • User Program ( Source Code) • Input Data • Output Data Input Output DL Framework 환경 Source Code
- 63. DL(Deep Learning) App 구현 Concept • Input/Output Data • Data Repository Tech. • NFS • Ceph, GlusterFS, Lustre ü HDFS • C3 HDFS Storage를 저장소로 사용
- 64. DL(Deep Learning) App 구현 Concept • Docker Container의 데이터는 volatile함 • Docker Container안에 permanent한 작업 영역(dir)이 필요함 • DL App Workspace (Directory) • DL App 실행을 위한 File/Dir들이 존재하는 영역 • User Program Src Code, Input/OutputData 들이 위치함 • Host Server상의 Directory이며 , Docker Container에 volumemount 되는 영역
- 65. DL(Deep Learning) App 구현 Concept /User_Dev_Workspace /DL_App_Workspace Deep Learning App Input Output User Container
- 66. DL App 실행 과정
- 67. DL App 실행 과정 YARN Container Yarn Container 시작
- 68. DL App 실행 과정 YARN Container GPU Device ID 할당 GPU Device ID Manager GPU 0,3 Zookeeper
- 69. YARN Container DL App 실행 과정 DL App Workspace Source Code User Workspace 복사 GPU 0,3
- 70. YARN Container DL App 실행 과정 DL App Workspace Input Source Code Input Data 복사 GPU 0,3
- 71. YARN Container DL App 실행 과정 DL App Workspace( NVIDIA Docker ) Input Source Code NVIDIA Docker Cotainer 실행 GPU 0,3
- 72. YARN Container DL App 실행 과정 DL App Workspace( NVIDIA Docker ) Input Source Code Docker Volume DL App Workspace Mount GPU 0,3
- 73. YARN Container DL App 실행 과정 ( NVIDIA Docker ) Input Source Code DL App Docker 실행 GPU 0,3
- 74. YARN Container DL App 실행 과정 ( NVIDIA Docker ) Input Source Code Output DL App Docker 실행 GPU 0,3
- 75. YARN Container DL App 실행 과정 Input Source Code Output DL App Workspace DL App Docker 종료 GPU 0,3
- 76. YARN Container DL App 실행 과정 Input Source Code Output DL App Workspace OutputData 복사 GPU 0,3
- 77. DL App 실행 과정 YARN Container GPU Device ID 해제 GPU Device ID Manager GPU 0,3 Zookeeper
- 78. DL App 실행 과정 YARN Container Yarn Container 종료
- 79. DL App 실행 과정 Yarn Container 종료
- 80. 원하는 DL App Properties • Docker Image • 실행 Src Code Dir Path • main 실행 script • Resource : CPU,Mem,GPU • Input Data 경로 • Output Data 경로 DL App 생성 및 실행하기 C3 Distributed Shell DL App 실행 과정 2) GPU ID 할당 3) User Workspace 복사 4) Input Data 복사 5) DL App Workspace 구성 완료 6) DL App NVIDIA Docker 실행 7) Output Data 복사 8) GPU ID 해제 YARN Cluster Shell Script Application Shell Script Application 생성
- 81. dl_app.properties • Docker Image • 실행 Src Code Dir Path • main 실행 script • Resource : CPU,Mem,GPU • Input Data 경로 • Output Data 경로 C3 Distributed Shell DL App 실행 과정 2) GPU ID 할당 3) User Workspace 복사 4) Input Data 복사 5) DL App Workspace 구성 완료 6) DL App NVIDIA Docker 실행 7) Output Data 복사 8) GPU ID 해제 YARN Cluster dlapp-launcher dl_app.properties DL App Launcher Shell Script Application Shell Script Application 생성
- 82. • Caffe LeNet MNIST Example • caffe/examples/mnist 기반 • . ├── dl_app_caffe.properties └── user_dev_workspace ├── create_mnist.sh ├── data/ ├── dl_app_start.sh ├── examples/ ├── lenet_solver.prototxt ├── lenet_train_test.prototxt ├── output/ └── train_lenet.sh dlapp-launcher 실행 예
- 83. • dl_app_caffe.properties [application] username=tony appname='c3 deeplearning app caffe mnist example' docker_image=naver/c3_dl-caffe user_dev_workspace_path=./user_dev_workspace user_shell_script=dl_app_start.sh user_shell_args='-gpu all' [from_hdfs] hdfs=/user/tony/caffe_test/* input_path=data [to_hdfs] output_path=output hdfs=/user/tony/caffe_mnist_example_output overwrite=true • dl_app_start.sh #!/usr/bin/env bash ./create_mnist.sh ./train_lenet.sh $@ dlapp-launcher 실행 예 . ├── dl_app_caffe.properties └── user_dev_workspace ├── create_mnist.sh ├── data/ ├── dl_app_start.sh ├── examples/ ├── lenet_solver.prototxt ├── lenet_train_test.prototxt ├── output/ └── train_lenet.sh
- 84. dlapp-launcher 실행 예 $ dlapp-launcher dl_app_caffe.properties C3 DL Cluster Version check : ok CheckDLAppProperties : ok user_dev_workspace_path = /path/to/user_dev_workspace user_dev_workspace_exclude_list = [] user_dev_workspace dir size = 4 kb DL App Submission : OK DL App ID = application_aaaaaaaaaaa_bbbb DL App Dashboard * While app is running , check : http://XXXXXXXXX/cluster/app/application_aaaaaaaaaaa_bbbb * After app is completed, check : http://XXXXXXXXX/applicationhistory/app/application_aaaaaaaaaaa_bbbb DL App Status and stdout/stderr Log URLs * dlapp-status application_aaaaaaaaaaa_bbbb
- 85. DL App Tools • dlapp-status [ DL App ID ] • dlapp-list [Username] • dlapp-softkill [Username] [ DL App ID ] • dlapp-kill [Username] [ DL App ID ]
- 87. DL App Log 확인 • YARN History,Timeline Server 기능 활용 • YARN Application stdout/stderr Log 는 Web 또는 yarn 명령어을 통 해 확인 가능 • YARN Container 의 Attempt ID 및 Container ID 를 통해 확인 • dlapp-status tool을 통해 stdout/stderr Log URL 제공
- 88. dlapp-status 실행 예 $ dlapp-status application_aaaaaaaaaaa_bbbb DL App application_aaaaaaaaaaa_bbbb is FINISHED Username : tony DL App ID : application_aaaaaaaaaaa_bbbb DL App Name : 'c3 deeplearning app caffe mnist example' startTime : 2016-09-26 22:24:11.575000 finishTime : 2016-09-26 22:24:57.703000 elapsedTime : 0:00:46.13 stdout http://XXXXX/applicationhistory/logs/XXXXX/container_aaaaaaaaaaa_bbbb_01_000004/container_aaaaaaaaaaa_bbbb _01_000004/tony/stdout/?start=-4096 stderr http://XXXXX/applicationhistory/logs/XXXXX/container_aaaaaaaaaaa_bbbb_01_000004/container_aaaaaaaaaaa_bbbb _01_000004/tony/stderr/?start=-4096
- 89. Example
- 91. dlapp-list 실행 예
- 92. dlapp-kill 실행 예
- 93. DL App Development 환경의 필요성 • DL App을 개발하려면 GPU 장비가 필요함 • 초기 DL App개발 시점에서 기본적인 검증 용도로 DL Cluster를 이용하는 것은 매우 번거로움 • 코드의 사소한 실수를 수정하는 과정의 반복 • Input/outputData등의 Directory 구조 확인
- 94. DL App Development 환경 • DL App 개발 전용 GPU 서버 구축 • dlapp-shell • dlapp-launcher 와 동일한 Docker Image 및 Workspace로 실행되는 docker bash 환경 제공 • dlapp-run : dlapp-shell 환경에서 C3 DL Cluster 에 dl app 을 실행
- 95. YARN Container DL App 실행 과정 ( NVIDIA Docker ) Input Source Code DL App Docker 실행 GPU 0,3
- 96. DL App 개발 서버 DL App Shell ( NVIDIA Docker ) : Bash Shell Source Code GPU 0,3
- 98. DL App Toolset • dlapp-launcher dl_app.properties • dlapp-shell dl_app.properties • dlapp-run • dlapp-status [DL App ID] • dlapp-list [Username] • dlapp-softkill [Username] [DL App ID] • dlapp-kill [Username] [DL App ID]
- 100. Hadoop YARN 기반의 Multi-tenantDeep Learning Application 실행 환경 ? GPU Cluster Hadoop YARN DL Application DL Application DL Application
- 101. DL Application C3 Deep Learning Cluster Zookeeper Dev. Environment Dev. Server Resource Manager YARN Node Manager YARN Docker Registry C3 HDFS Storage YARN with GPU Scheduling Node Manager YARN Node Manager YARN . . . C3 Distributed Shell NVIDIA Docker GPU Device ID Manager Input Output DL App Toolset dlapp-launcher dlapp-shell dlapp-run
- 102. DL Framework Docker Images DL Solution Docker Image Caffe naver/c3_dl-caffe Torch naver/c3_dl-torch Theano naver/c3_dl-theano TensorFlow naver/c3_dl-tensorflow Theano+Keras naver/c3_dl-theano-keras TensorFlow+Keras naver/c3_dl-tensorflow-keras Base (CUDA/cuDNN) naver/c3_dl-base
- 103. • Resource 관리 • GPU 자원 Utilization 증가 • GPU ResourcePlanning ( YARN Queue) • Administration • 작업 이력 관리 • 프로젝트 관리 • 사용자 관리 • DevOps • 규격화된 개발 환경 제공 • 개발 환경 공유 개선점
- 104. • 검색 품질 향상을 위한 seq2seq 기반의 검색 키워드 변환 • word2vec,CNN 기반의 문서 주제 분류 • CNN 기반의 Image Tagger • 이외 다수의 프로젝트가 실행중 현재 C3 DL Cluster에서는 … * CNN : Convolutional Neural Network
- 105. To Do
- 106. To Do • GPU Resource Scheduling 고도화 • 대용량 Input Data • Input Data Caching • Input Data Feeder • Distributed Deep Learning
- 107. Q&A
- 108. Thank You