SlideShare a Scribd company logo
1 of 35
Download to read offline
Kwanghoon Jhin
kwanghoon.jhin@kr.ibm.com
March, 2017
PowerAI
World Fastest Machine Learning Platform
IBM Power Systems: S822LC for HPC
IBM Storage: Elastic Storage Server
Revision E1
2
출처: nvidia.com
3
4
1950 — Alan Turing creates the “Turing Test”
1952 — Arthur Samuel wrote the first computer learning program. The program was the game of checkers,
and the IBM computer improved at the game the more it played, studying which moves made up winning
strategies and incorporating those moves into its program.
1959 – “Field of study that gives computers the ability to learn without being explicitly programmed” Arthur
Samuel
A (very short) History of AI
6
A (very short) History of AI
IBM 701 vacuum tube plug-in unit
700-series Electronic Data Processing Machine.
1957 — Frank Rosenblatt designed the first neural network for computers (the perceptron), which simulate
the thought processes of the human brain.
A (very short) History of AI
A (very short) History of AI
A (very short) History of AI
“… the navy revealed the embryo of an electronic computer today that it
expects will be able to walk, talk, see, write, reproduce
itself and be conscious of its existence ... Dr. Frank Rosenblatt, a
research psychologist at the Cornell Aeronautical Laboratory, Buffalo, said
Perceptrons might be fired to the planets as mechanical space explorers”
08-JUL-1958, the New York Times.
A (very short) History of AI
과도한 자신감이 거짓 믿음을 만들다.
Perceptrons (1969) by Marvin Minsky, Founder of the MIT AI Lab
– We need to use Multi-Layer Perceptrons
– No one on earth had found a viable way to train MLPs good enough to learn such simple functions.
11
A (very short) History of AI
12
AI winter:
– 1974 to 1980
– 1987 to 1993
Backpropagation
– 1974 by Paul Werbos
– 1982 by Paul Werbos
– 1986 by Geoffrey Hinton
13
A (very short) History of AI
Paul Werbos (1974). Beyond regression: New tools for prediction and analysis in the behavioral sciences.
PhD thesis, Harvard University.
Paul Werbos (1982). Applications of advances in nonlinear sensitivity analysis. In System modeling and
optimization (pp. 762-770). Springer Berlin Heidelberg.
Rumelhart, David E.; Hinton, Geoffrey E.; Williams, Ronald J. (8 October 1986). "Learning representations
by back-propagating errors". Nature. 323 (6088): 533–536.
14
A (very short) History of AI
Backpropagation / 역전파(逆傳播)
1997 — IBM’s Deep Blue beats the world champion at chess.
2006 — Geoffrey Hinton coins the term “deep learning” to explain new algorithms that let computers “see”
and distinguish objects and text in images and videos.
2010 — The Microsoft Kinect can track 20 human features at a rate of 30 times per second, allowing people
to interact with the computer via movements and gestures.
2011 — IBM’s Watson beats its human competitors at Jeopardy.
2011 — Google Brain is developed, and its deep neural network can learn to discover and categorize objects
much the way a cat does.
2012 — In 2012, Hinton and two of his other U of T students, Alex Krizhevsky and Ilya Sutskever, entered an
image recognition contest. Competing teams build computer vision algorithms that learn to identify objects in
millions of pictures; the most accurate wins. The U of T model took the error rate of the best-performing
algorithms to date, and the error rate of the average human, and snapped the difference in half like a dry
twig. The trio created a company, and Google acquired it.
(https://www.thestar.com/news/world/2015/04/17/how-a-toronto-professors-research-revolutionized-artificial-
intelligence.html)
15
A (very short) History of AI
2012 — Google’s X Lab develops a machine learning algorithm that is able to autonomously browse
YouTube videos to identify the videos that contain cats.
2014 — Facebook develops DeepFace, a software algorithm that is able to recognize or verify individuals on
photos to the same level as humans can.
2015 — Amazon launches its own machine learning platform.
2015 — Microsoft creates the Distributed Machine Learning Toolkit, which enables the efficient distribution of
machine learning problems across multiple computers.
2015 — Over 3,000 AI and Robotics researchers, endorsed by Stephen Hawking, Elon Musk and Steve
Wozniak (among many others), sign an open letter warning of the danger of autonomous weapons which
select and engage targets without human intervention.
2016 — Google’s artificial intelligence algorithm beats a professional player at the Chinese board game Go,
which is considered the world’s most complex board game and is many times harder than chess. The
AlphaGo algorithm developed by Google DeepMind managed to win five games out of five in the Go
competition.
16
A (very short) History of AI
http://www.forbes.com/sites/bernardmarr/2016/02/19/a-short-history-of-machine-learning-every-manager-should-read/
17
PowerAI > 왜 지금인가?
긴 ‘인공지능의 겨울(AI Winter)’를 끝내고
‘인공지능의 봄(AI Spring)’을 맞이하다.
Big DataTechnologyAlgorithm
비정형 데이터의 범람하드웨어 능력의 비약적인 발전관습과 한계에 도전 그리고 성취
학계에서는 지금의 시대를 ‘인공지능의 봄’이라고 부릅니다.
그 이유는 이전까지 없었던 아래의 세 가지가 갖추어졌기 때문입니다.
인공지능 설계의 오래된
난제를 해결
수퍼 컴퓨터에서만 가능하던
연산을 책상 위에서 처리
기계학습에 새로운 가능성과
의미를 부여
18
PowerAI > 기계학습에 요구되는 컴퓨팅 구성 요소들
지금 이 시대에 필요한 컴퓨팅 플랫폼
GPU computingFrameworks
한계없는 데이터의 저장과 사용을
위한 소프트웨어 정의 스토리지
가장 혁신적인 I/O 기술을 탑재한
GPU 컴퓨팅 시스템
IBM이 최적화환 프래임워크 및
라이브러리
IBM은 PowerAI라는 이름으로 ‘최적화된 프레임워크’, ‘GPU 컴퓨팅 시스템’
그리고 ‘소프트웨어 정의 스토리지’를 하나의 컴퓨팅 플랫폼으로 제공하고 있습니다.
IBM Optimized Frameworks IBM S822LC for HPC IBM Elastic Storage Server
Software Defined Storage
19
IBM이 최적화 한 ML/DL 프레임워크
IBM은 ibm.biz/powerai 를 통하여 S822LC for HPC에
최적화된 ML/DL 프레임워크를 제공합니다.
OpenBLAS BazelDIGITS NCCLDistributed
Frameworks
Machine Learning / Deep Learning Frameworks
Supporting Libraries
기계학습 플랫폼을 계획하고 구축하는데에 많은 시간을
투자해야 하는 부분이 바로, 프래임워크 최적화입니다.
IBM은 이와 같은 현장의 불편과 불필요한 시간투자를
최소화할 수 있는 ‘최적화’된 프래임워크를 제공합니다.
GPU의 기원
– 1999 – Nvidia, GeForce 256를 출시하면서 “the
world’s first GPU”라고 지칭. 그리고 다음과 같이
GeForce 256 GPU를 설명
• “Single-chip processor with integrated
transform, lighting, triangle setup/clipping, and
rendering engines”
– 2002 – Nvidia의 경쟁사인, ATI Technologies
(AMD)에서 Radeon 9700을 발표하면서 VPU,
‘Visual Processing Unit’이라고 지칭
– 2007 – Nvidia, GPGPU인 Tesla 제품군 발표
Co-Processor로 GPGPU 활용
– HPC
– Machine Learning
20
S822LC for HPC > Graphic Processing Unit
21
GPU 컴퓨팅은 CPU 컴퓨팅과 무엇이 다른가?
Mythbusters at NVISION 08
22
1 2 5 12
9 6 0 3
5 72 9 -15
5 16 -32 -20
18 20
98 -58
78
1 + 2 + 5 + 12 + 9 + 6 + 0 + 3 + 5 + 72 + 9 + (-15) + 5 + 16 + (-32) + (-20) = 78
16개의 수를 합하는 연산을 해야 합니다.
주어진 16개의 수를 차례대로 합하여 총합산을 끝내는 방식(직렬연산)과,
4개씩 짝을 지어 동시에 연산하는 방식(병렬연산)이 있을 수 있겠습니다.
전자의 방식이 일반적인 CPU 연산의 방식이라고 할 수 있고,
후자가 GPU가 취하는 방식입니다.
GPU는 병렬 처리를 효율적으로 처리하기 위한 수천 개의 코어를 가지고 있습니다.
S822LC for HPC > GPU는 연산에 어떤 도움을 주는가?
Although NVLink primarily focuses on connecting multi
also connect Tesla P100 GPUs with IBM Power CPUs
shows how the CPUs are connected with NVLink in th
configuration, each GPU has 180 GBps bidirectional b
and 80 GBps bidirectional bandwidth to the connected
Figure 1-12 CPU to GPU and GPU to GPU interconnect us
All the initialization of the GPU is through the PCIe inte
the side band communication for status, power manag
and running, all data communication is using the NVLi
23
S822LC for HPC > NVLink
NVLink로 PCIe의 한계를 뛰어 넘다.
GPU 제작사인 Nvidia는 GPU 컴퓨팅의 큰 문제 중 하나인
CPU(host)와 GPU(device) 사이 그리고 GPU(device)와
GPU(device) 사이의 병목현상에 주목해 왔습니다. 그리고 그
해결책으로 NVLink라는 bus를 선보였습니다.
Nvidia는 IBM와 손잡고, 세계 최초 NVLink 1.0을 완벽
구현하기에 이르렀으며, S822LC for HPC에 적용되었습니다.
IBM은 NVLink 기술을 높게 평가하여 2.0 버전을 차세대 CPU
플랫폼인 Power9에도 적용하기로 결정했습니다.
PCIe 3.0 x16
32 GB/s
bidirectional
NVLink
160 GB/s
bidirectional
X5
bandwidth
4 GPUs
with NVLink
24
S822LC for HPC > Unified Memory
지금까지의 GPU 메모리 접근의 한계를 없애다.
Unified Memory
• 전세대 GPU, Kepler와 Maxwell의 한계로 멈추어 있던 GPU
사용 메모리 크기 제한(GPU 내 설계에 따른 전용 메모리,
GDDR을 벗어날 수 없음)이 Pascal에서는 제거되어,
시스템에 탑재된 모든 메모리를 사용 가능
• 이제 개발자들이 GPU 메모리 내의 데이터 이동 관리보다
연산 자체에 집중하는 것이 가능
‘두껍고도 수평적으로’ (both Fat and Flat) 설계
• 어느 연결에서도 데이터 전송 병목이 생기지 않도록 설계
• GPU에서도 CPU처럼 시스템 메모리를 직접 접근하고 사용
• 같은 소켓의 GPU간 ‘Fat Pipe’ 구현
보편적 업무와 연산 알고리즘에 최적화
• Startup/Teardown시 폭발적인 성능
• GPU간의 안정적 데이터 이송
• 기존의 부족한 대역폭으로 인한 host-device간의 bus
transfer 문제를 해소
Pascal
GPU
Power8
CPU
Unified Memory
CUDA 8
Allocated beyond GPU memory
Up to 1TB on S822LC for HPC
25
S822LC for HPC > 성능
유사 조건에서 GPU 세대간 성능비교
0
20
40
60
80
100
120
140
x86 with 4x M40 / PCIe Power8 with 4x P100 / NVLink
Training time (minutes): AlexNet and
Caffe to top-1, 50% Accuracy
(Lower is better)
0:00
1:12
2:24
3:36
4:48
6:00
7:12
8:24
x86 with 8x M40 / PCIe Power8 with 4x P100 / NVLink
BVLC Caffe vs IBM Caffe / VGGNet
Time to Top-1 50% accuracy:
(Lower is better)
• IBM	S822LC	20-cores	2.86GHz	512GB	memory		
/	4	NVIDIA	Tesla	P100	GPUs	/	Ubuntu	16.04	/	CUDA	8.0.44	/	cuDNN 5.1		/	IBM	Caffe 1.0.0-rc3	/		Imagenet Data
• Intel	Broadwell	E5-2640v4	20-core	2.6	GHz	512GB	memory	
/	8	NVIDIA	Tesla M40	GPUs /	Ubuntu	16.04	/		CUDA	8.0.44	/	cuDNN 5.1	/	BVLC	Caffe 1.0.0-rc3	/	Imagenet Data
S822LC/HPC	with	4	Tesla	
P100	Tesla	GPUs	is	24%	
Faster	than	8x	Tesla	M40	
GPUs
S822LC/HPC	with	4	Tesla	
P100	GPUs is	2.2x	Faster	
than	4x	Tesla	M40	GPUs
x2.2
24%
26
S822LC for HPC > 성능
유사 조건에서 NVLink와 PCIe와의 비교
X86 + GeForce
Pascal Architecture
on PCIe 3.0 x16
POWER8 + Tesla
Pascal Architecture
on NVLink
ImageNet / Alexnet: Minibatch size = 128
170 ms
78 ms
IBM advantage: data communication
and GPU performance
54%
Pascal Architecture로 제작된 Tesla P100과 GeForce GTX Titan X를
직접 비교한 결과, NVLink의 유무가 전체 처리 속도를 크게
좌우한다는 것을 알 수 있었습니다.
• 데이터 이동시간의 획기적인 단축
• 연산 대기시간 단축으로 전체적인 부하 감소
• 총 연산시간의 실질적인 단축
27
S822LC for HPC > 내부구조 및 구성
2U 19inch Rack Mount, 2-Socket POWER8 CPU,
4 Tesla P100 GPU, 32 DDR4 DIMM slots, 3 PCIe slots.
POWER8 with NVLink (2x)
• 190W
• Integrated NVLink 1.0
Memory DIMM’s Riser (8x)
• 4 IS DDR4 DIMMs per riser
• Single Centaur per riser
• 32 IS DIMM’s total
• 32-1024 GB memory capacity
PCIe slot (3x)
• Gen3 PCIeNvidia GPU
• SXM2 form factor
• NVLink 1.0
• 300 W
• Max of 2 per socket
Power Supplies (2x)
• 1300W
• Common Form Factor
SupplyCooling Fans (4x)
• 80mm Counter- Rotating Fans
• Hot swap
Storage Option (2x)
• 0-2, SATA HDD.SSD
• Tray design for install/removal
• Hot Swap
Service Controller Card
• BMC Content
28
S822LC for HPC > 내부구조 및 구성 > 수냉식 냉각장치
수냉식 냉각장치 옵션 적용
기본 공냉식 냉각장치
수냉식 냉각장치 제공 WATER COOLING OPTION
S822LC for HPC는 대규모 HPC 컴퓨팅 시스템 제작에
필수적인 수냉식 냉각방식을 옵션으로 제공하고
있습니다.
GPU 컴퓨팅에서 필연적으로 따라오는 발열문제를
근본적으로 해결하고 다수의 S822LC for HPC을
구성한 다중 랙을 운영할 때에도 보다 안정적인
컴퓨팅 파워를 확보함과 동시에 전산실 내 공조
시스템의 부하를 줄일 수 있습니다.
수냉식 냉각옵션은 4GPU 구성에만 제공되며, 아래
그림과 같이 방렬판에 직결되어 냉각제가 공급되는
호스가 설치됩니다.
29
S822LC for HPC > Nvidia Tesla 제품군 비교
S822LC for HPC는 ‘P100 (NVLink)’를 사용합니다.
P100	(NVLink) P100	(PCIe) M40 K40 K80
Architecture PASCAL PASCAL Maxwell Kepler Kepler
CUDA	cores 3840 3840 3072 2880
4992
(dual	GPU)
Half-precision	
performance	(FP16)
21.2	TeraFLOPS with	
NVIDIA	GPU	Boost
18.68 TeraFLOPS with	
NVIDIA	GPU	Boost
N/A N/A N/A
Single-Precision
Performance	(FP32)
10.6	TeraFLOPS with	
NVIDIA	GPU	Boost
9.34 TeraFLOPS with	
NVIDIA	GPU	Boost
7	Teraflops	with	
NVIDIA	GPU	Boost
4.29	Teraflops	with	
NVIDIA	GPU	Boost
8.73	Teraflops	with	
NVIDIA	GPU	Boost
Double-Precision
Performance	(FP64)
5.3	TeraFLOPS with	
NVIDIA	GPU	Boost
4.67 TeraFLOPS with	
NVIDIA	GPU	Boost
0.2	Teraflops	with	
NVIDIA	GPU	Boost
1.43	Teraflops	with	
NVIDIA	GPU	Boost
2.91	Teraflops	with	
NVIDIA	GPU	Boost
GPU	Memory 16	GB	HBM2 12GB	or	16GB	HBM2 24	GB	GDDR5 12	GB	GDDR5
24	GB	GDDR5
(12	GB	per	GPU)
Memory	Bandwidth 720	GB/s	HBM2
3072-bit	HBM2	(12GB)
4096-bit	HBM2	(16GB)	
288	GB/s 288	GB/s
480	GB/s
(240	GB/s	per	GPU)
System	Interface
160	GB/s	bidirectional	
interconnect	
bandwidth	with	
NVIDIA	NVLink
32	GB/s
PCI	Express	3.0	x16
32	GB/s	
PCI	Express	3.0	x16
32	GB/s	
PCI	Express	3.0	x16
32	GB/s	
PCI	Express	3.0	x16
Max	Power	
Consumption
300	W 250	W 250	W 235	W
300	W
(150W	per	GPU)
30
IBM Elastic Storage Server
Power Systems
S822L (Server) &
S812L (Mgmt)
EXP24 (GS) or
DCS3700 (GL) or
DeepFlash150 (GF)
Spectrum Scale
Software
Defined Storage
Elastic
Storage
Server
Compute Disks Software
IBM Elastic Storage Server는 IBM Power Systems를
기반으로 Linux 운영체제 그리고 IBM Spectrum Scale이라는
파일 시스템을 중심으로 구성되는 ‘소프트웨어 정의
스토리지’이자 스토리지 어플라이언스입니다.
31
IBM Elastic Storage Server > 개요
IBM Elastic Storage Server는 Spectrum Scale의 모든
장점을 그대로 발휘합니다.
§ Spectrum Scale:
– 다양한 플랫폼상에서 고성능, 고가용 파일시스
템 서비스 및 데이터 관리 서비스를 제공할 수
있도록 설계된 시스템 소프트웨어.
– 1993년 발표된 이후 오랜 기간 안정화 및 버전
업을 통해 성숙화된 파일시스템 솔루션
클러스터 내
최대 노드 수
Architectural Limit 16,384
Maximum Tested 9,300
파일시스템 크기
Architectural Limit
2^99bytes
(524,288 YiB)
Maximum Tested 18PiB
마운드할 수 있는 최대 파일시스템 수 256
파일시스템당
최대 파일 수
Architectural Limit 2^64
Maximum Tested 9,000,000,000
단일 파일시스템에서 테스트된 최대 I/O
대역폭
400GB/s
최대 Dependent fileset 수 10,000
최대 Independent fileset 수 1,000
최대 Global Snapshot 수 256
각각의 Independent fileset에 대한 최대
Snapshot 수
256
병렬/분산 처리 파일시스템
Global Single Name Space
소프트웨어 또는 어플라이언스로 제공
Ethernet/Infiniband를 통한 고속 데이터 처리
32
IBM Elastic Storage Server > 제품군
다양한 구성을 지원하며, 속도중심 또는 용량중심의
제품을 선택할 수 있습니다. All Flash도 지원합니다.
Model GL4
4 Enclosures, 20U
232 NL-SAS, 2 SSD
10+ GB/Sec
Model GL6
6 Enclosures, 28U
348 NL-SAS, 2 SSD
15+ GB/sec
Model GL2
2 Enclosures, 12U
116 NL-SAS, 2 SSD
5+ GB/Sec
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
FC5887
Model GS1
24 SSD
6 GB/Sec
Model GS2
46 SAS + 2 SSD or
48 SSD Drives
2 GB/Sec SAS
12 GB/Sec SSD
Model GS4
94 SAS + 2 SSD or
96 SSD Drives
5 GB/Sec SAS
16 GB/Sec SSD
Model GS6
142 SAS + 2 SSD
7 GB/Sec
Model GF2
2 Enclosures, 10u
360TB, usable Flash
Max Read 26 GB/s
Max Write 16 GB/s
Model GF1
1 Enclosure, 7U
180TB, usable Flash
Max Read 13 GB/s
Max Write 9 GB/s
• GS models	use	2U	24x2.5”	JBODs	or	SSDs > 최대 120TB
• GL models	use	4U	60x3.5”	JBODs	> 최대 2PB
• GF models	use	3U	32xFlash	JBOFs	enclosures	> 최대 360TB
• Support	drives:	1.8TB	SAS,	400GB,	800GB,	1.6TB	SSD	2.5”;	
2TB,4TB,6TB	and	8TB	NL-SAS	3.5”	HDDs
• Supported	NICs:		10GbE,	40GbE	Ethernet	and	EDR	Infiniband
33
IBM Elastic Storage Server > HPC 고객 도입사례
“기존의 일반적인 RAID
스토리지 어레이에서는 1TB
디스크를 리빌드하는데
12시간 전후가 소요 되었는데,
GNR은 2TB 디스크를 1시간
이내에 리빌드가 가능하도록
했으며, 연구 시뮬레이션을
위한 고속 데이터 엑세스를
가능하도록 해 주었다.”
Lothar
Wollschläger
Storage Specialist @ Jülich Supercomputing Centre
IBM Elastic Storage Server
• 총 20대로 구성
• 7 PB (usable capacity)
• 200 GB/s I/O 성능
IBM Elastic Storage Server에서 제공되는 ‘GPFS Native RAID’를 통해 디스크 리빌드(disk rebuild) 시간을
획기적으로 단축시켜 리빌드 중 발생 가능한 2차 장애의 위험성이 현저하게 감소
IBM Systems HardwarePowerAI
World Fastest Machine Learning Platform
THANK YOU!
Architectural Overview
Operation
Training	Layer
(Compute	Cluster)
Data	Layer
(Central	Storage)
Compute
Compute
Compute
Job	Scheduler
Storage	Server
Data
Data
Data
Service	Layer
(Service	Gateway)
Database
Index
Search
Ranking
API
High	Bandwidth	Service	Network
Management	Network
IBM	Elastic	Storage	
Server
IBM	Spectrum	Scale
IBM	S822LC	for	HPC
IBM	Spectrum	LSF
IBM	IB	Switch	EDR

More Related Content

What's hot

[142]편광을 활용한6 dof 전현기
[142]편광을 활용한6 dof 전현기[142]편광을 활용한6 dof 전현기
[142]편광을 활용한6 dof 전현기
NAVER D2
 
제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개
제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개
제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개
Tommy Lee
 
Tdc2013 선배들에게 배우는 server scalability
Tdc2013 선배들에게 배우는 server scalabilityTdc2013 선배들에게 배우는 server scalability
Tdc2013 선배들에게 배우는 server scalability
흥배 최
 
Puppet과 자동화된 시스템 관리
Puppet과 자동화된 시스템 관리Puppet과 자동화된 시스템 관리
Puppet과 자동화된 시스템 관리
Keon Ahn
 
Ceph Day Seoul - Bring Ceph to Enterprise
Ceph Day Seoul - Bring Ceph to EnterpriseCeph Day Seoul - Bring Ceph to Enterprise
Ceph Day Seoul - Bring Ceph to Enterprise
Ceph Community
 

What's hot (20)

[142]편광을 활용한6 dof 전현기
[142]편광을 활용한6 dof 전현기[142]편광을 활용한6 dof 전현기
[142]편광을 활용한6 dof 전현기
 
[154] 데이터 센터의 오픈 소스 open compute project (ocp)
[154] 데이터 센터의 오픈 소스 open compute project (ocp)[154] 데이터 센터의 오픈 소스 open compute project (ocp)
[154] 데이터 센터의 오픈 소스 open compute project (ocp)
 
Community Openstack 구축 사례
Community Openstack 구축 사례Community Openstack 구축 사례
Community Openstack 구축 사례
 
Why OpenStack is Operating System?
Why OpenStack is Operating System?Why OpenStack is Operating System?
Why OpenStack is Operating System?
 
[121]네이버 효과툰 구현 이야기
[121]네이버 효과툰 구현 이야기[121]네이버 효과툰 구현 이야기
[121]네이버 효과툰 구현 이야기
 
제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개
제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개
제2회 난공불락 오픈소스 인프라 세미나 zinst 관리툴 소개
 
[OpenInfra Days Korea 2018] Day 2 - E5: Mesos to Kubernetes, Cloud Native 서비스...
[OpenInfra Days Korea 2018] Day 2 - E5: Mesos to Kubernetes, Cloud Native 서비스...[OpenInfra Days Korea 2018] Day 2 - E5: Mesos to Kubernetes, Cloud Native 서비스...
[OpenInfra Days Korea 2018] Day 2 - E5: Mesos to Kubernetes, Cloud Native 서비스...
 
[오픈소스컨설팅] SELinux : Stop Disabling SELinux
[오픈소스컨설팅] SELinux : Stop Disabling SELinux[오픈소스컨설팅] SELinux : Stop Disabling SELinux
[오픈소스컨설팅] SELinux : Stop Disabling SELinux
 
Tdc2013 선배들에게 배우는 server scalability
Tdc2013 선배들에게 배우는 server scalabilityTdc2013 선배들에게 배우는 server scalability
Tdc2013 선배들에게 배우는 server scalability
 
[2015-05월 세미나] 파이선 초심자의 Openstack
[2015-05월 세미나] 파이선 초심자의 Openstack[2015-05월 세미나] 파이선 초심자의 Openstack
[2015-05월 세미나] 파이선 초심자의 Openstack
 
resource on openstack
 resource on openstack resource on openstack
resource on openstack
 
Puppet과 자동화된 시스템 관리
Puppet과 자동화된 시스템 관리Puppet과 자동화된 시스템 관리
Puppet과 자동화된 시스템 관리
 
[OpenInfra Days Korea 2018] (Track 4) - Grafana를 이용한 OpenStack 클라우드 성능 모니터링
[OpenInfra Days Korea 2018] (Track 4) - Grafana를 이용한 OpenStack 클라우드 성능 모니터링[OpenInfra Days Korea 2018] (Track 4) - Grafana를 이용한 OpenStack 클라우드 성능 모니터링
[OpenInfra Days Korea 2018] (Track 4) - Grafana를 이용한 OpenStack 클라우드 성능 모니터링
 
[오픈소스컨설팅] ARM & OpenStack Community
[오픈소스컨설팅] ARM & OpenStack Community[오픈소스컨설팅] ARM & OpenStack Community
[오픈소스컨설팅] ARM & OpenStack Community
 
[오픈소스컨설팅]Docker on Kubernetes v1
[오픈소스컨설팅]Docker on Kubernetes v1[오픈소스컨설팅]Docker on Kubernetes v1
[오픈소스컨설팅]Docker on Kubernetes v1
 
Real-time Big Data Analytics Practice with Unstructured Data
Real-time Big Data Analytics Practice with Unstructured DataReal-time Big Data Analytics Practice with Unstructured Data
Real-time Big Data Analytics Practice with Unstructured Data
 
Ceph Day Seoul - Bring Ceph to Enterprise
Ceph Day Seoul - Bring Ceph to EnterpriseCeph Day Seoul - Bring Ceph to Enterprise
Ceph Day Seoul - Bring Ceph to Enterprise
 
150625 마이크로커널 운영체제 김지은
150625 마이크로커널 운영체제 김지은150625 마이크로커널 운영체제 김지은
150625 마이크로커널 운영체제 김지은
 
[2B5]nBase-ARC Redis Cluster
[2B5]nBase-ARC Redis Cluster[2B5]nBase-ARC Redis Cluster
[2B5]nBase-ARC Redis Cluster
 
Linux Kernel 101 for Beginner
Linux Kernel 101 for BeginnerLinux Kernel 101 for Beginner
Linux Kernel 101 for Beginner
 

Similar to 제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-AI

OPEN_POWER8_SESSION_20150316
OPEN_POWER8_SESSION_20150316OPEN_POWER8_SESSION_20150316
OPEN_POWER8_SESSION_20150316
기한 김
 
[조진현]Kgc2012 c++amp
[조진현]Kgc2012 c++amp[조진현]Kgc2012 c++amp
[조진현]Kgc2012 c++amp
진현 조
 
Cloud based NGS framework
Cloud based NGS frameworkCloud based NGS framework
Cloud based NGS framework
Hyungyong Kim
 
머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발
머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발
머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발
Jeongkyu Shin
 

Similar to 제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-AI (20)

[IBM 김상훈] AI 최적화 플랫폼 IBM AC922 소개와 활용 사례
[IBM 김상훈] AI 최적화 플랫폼 IBM AC922 소개와 활용 사례[IBM 김상훈] AI 최적화 플랫폼 IBM AC922 소개와 활용 사례
[IBM 김상훈] AI 최적화 플랫폼 IBM AC922 소개와 활용 사례
 
[2017 AWS Startup Day] 스타트업이 인공지능을 만날 때 : 딥러닝 활용사례와 아키텍쳐
[2017 AWS Startup Day] 스타트업이 인공지능을 만날 때 : 딥러닝 활용사례와 아키텍쳐[2017 AWS Startup Day] 스타트업이 인공지능을 만날 때 : 딥러닝 활용사례와 아키텍쳐
[2017 AWS Startup Day] 스타트업이 인공지능을 만날 때 : 딥러닝 활용사례와 아키텍쳐
 
KYSR 1st Seminar 엔비디아 김승규 상무
KYSR 1st Seminar 엔비디아 김승규 상무KYSR 1st Seminar 엔비디아 김승규 상무
KYSR 1st Seminar 엔비디아 김승규 상무
 
Foss open sorucesw_6902
Foss open sorucesw_6902Foss open sorucesw_6902
Foss open sorucesw_6902
 
Ibm과 nvidia가 제안하는 딥러닝 플랫폼
Ibm과 nvidia가 제안하는 딥러닝 플랫폼Ibm과 nvidia가 제안하는 딥러닝 플랫폼
Ibm과 nvidia가 제안하는 딥러닝 플랫폼
 
이기종 멀티코어 기반의 Open cv 응용 사례 및 효율적인 어플리케이션 디자인
이기종 멀티코어 기반의 Open cv 응용 사례 및 효율적인 어플리케이션 디자인이기종 멀티코어 기반의 Open cv 응용 사례 및 효율적인 어플리케이션 디자인
이기종 멀티코어 기반의 Open cv 응용 사례 및 효율적인 어플리케이션 디자인
 
Ic922 ac922 e dm_202008
Ic922 ac922 e dm_202008Ic922 ac922 e dm_202008
Ic922 ac922 e dm_202008
 
김성윤 - 우분투로 슈퍼컴 만들기 (2011Y03M26D)
김성윤 - 우분투로 슈퍼컴 만들기 (2011Y03M26D)김성윤 - 우분투로 슈퍼컴 만들기 (2011Y03M26D)
김성윤 - 우분투로 슈퍼컴 만들기 (2011Y03M26D)
 
[2B2]기계 친화성을 중심으로 접근한 최적화 기법
[2B2]기계 친화성을 중심으로 접근한 최적화 기법[2B2]기계 친화성을 중심으로 접근한 최적화 기법
[2B2]기계 친화성을 중심으로 접근한 최적화 기법
 
SOSCON 2017 - Backend.AI
SOSCON 2017 - Backend.AISOSCON 2017 - Backend.AI
SOSCON 2017 - Backend.AI
 
[E-commerce & Retail Day] 인공지능서비스 활용방안
[E-commerce & Retail Day] 인공지능서비스 활용방안[E-commerce & Retail Day] 인공지능서비스 활용방안
[E-commerce & Retail Day] 인공지능서비스 활용방안
 
델켐매거진5호
델켐매거진5호델켐매거진5호
델켐매거진5호
 
OPEN_POWER8_SESSION_20150316
OPEN_POWER8_SESSION_20150316OPEN_POWER8_SESSION_20150316
OPEN_POWER8_SESSION_20150316
 
Neural stream
Neural streamNeural stream
Neural stream
 
ARTIK 710 IoT class
ARTIK 710 IoT class ARTIK 710 IoT class
ARTIK 710 IoT class
 
[조진현]Kgc2012 c++amp
[조진현]Kgc2012 c++amp[조진현]Kgc2012 c++amp
[조진현]Kgc2012 c++amp
 
Cloud based NGS framework
Cloud based NGS frameworkCloud based NGS framework
Cloud based NGS framework
 
머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발
머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발
머신러닝 및 데이터 과학 연구자를 위한 python 기반 컨테이너 분산처리 플랫폼 설계 및 개발
 
Nvidia architecture
Nvidia architectureNvidia architecture
Nvidia architecture
 
[경북] I'mcloud information
[경북] I'mcloud information[경북] I'mcloud information
[경북] I'mcloud information
 

More from Tommy Lee

More from Tommy Lee (20)

새하늘과 새땅-리차드 미들턴
새하늘과 새땅-리차드 미들턴새하늘과 새땅-리차드 미들턴
새하늘과 새땅-리차드 미들턴
 
하나님의 아픔의신학 20180131
하나님의 아픔의신학 20180131하나님의 아픔의신학 20180131
하나님의 아픔의신학 20180131
 
그리스도인의미덕 통합
그리스도인의미덕 통합그리스도인의미덕 통합
그리스도인의미덕 통합
 
그리스도인의미덕 1장-4장
그리스도인의미덕 1장-4장그리스도인의미덕 1장-4장
그리스도인의미덕 1장-4장
 
예수왕의복음
예수왕의복음예수왕의복음
예수왕의복음
 
Grub2 and troubleshooting_ol7_boot_problems
Grub2 and troubleshooting_ol7_boot_problemsGrub2 and troubleshooting_ol7_boot_problems
Grub2 and troubleshooting_ol7_boot_problems
 
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-CRUI
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-CRUI제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-CRUI
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-CRUI
 
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Ranchers
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Ranchers제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Ranchers
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Ranchers
 
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Asible
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Asible제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Asible
제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-Asible
 
제3회난공불락 오픈소스 인프라세미나 - lustre
제3회난공불락 오픈소스 인프라세미나 - lustre제3회난공불락 오픈소스 인프라세미나 - lustre
제3회난공불락 오픈소스 인프라세미나 - lustre
 
제3회난공불락 오픈소스 인프라세미나 - Nagios
제3회난공불락 오픈소스 인프라세미나 - Nagios제3회난공불락 오픈소스 인프라세미나 - Nagios
제3회난공불락 오픈소스 인프라세미나 - Nagios
 
제3회난공불락 오픈소스 인프라세미나 - MySQL Performance
제3회난공불락 오픈소스 인프라세미나 - MySQL Performance제3회난공불락 오픈소스 인프라세미나 - MySQL Performance
제3회난공불락 오픈소스 인프라세미나 - MySQL Performance
 
제3회난공불락 오픈소스 인프라세미나 - MySQL
제3회난공불락 오픈소스 인프라세미나 - MySQL제3회난공불락 오픈소스 인프라세미나 - MySQL
제3회난공불락 오픈소스 인프라세미나 - MySQL
 
제3회난공불락 오픈소스 인프라세미나 - JuJu
제3회난공불락 오픈소스 인프라세미나 - JuJu제3회난공불락 오픈소스 인프라세미나 - JuJu
제3회난공불락 오픈소스 인프라세미나 - JuJu
 
제3회난공불락 오픈소스 인프라세미나 - Pacemaker
제3회난공불락 오픈소스 인프라세미나 - Pacemaker제3회난공불락 오픈소스 인프라세미나 - Pacemaker
제3회난공불락 오픈소스 인프라세미나 - Pacemaker
 
새하늘과새땅 북톡-3부-우주적회복에대한신약의비전
새하늘과새땅 북톡-3부-우주적회복에대한신약의비전새하늘과새땅 북톡-3부-우주적회복에대한신약의비전
새하늘과새땅 북톡-3부-우주적회복에대한신약의비전
 
새하늘과새땅 북톡-2부-구약에서의총체적구원
새하늘과새땅 북톡-2부-구약에서의총체적구원새하늘과새땅 북톡-2부-구약에서의총체적구원
새하늘과새땅 북톡-2부-구약에서의총체적구원
 
새하늘과새땅 Part1
새하늘과새땅 Part1새하늘과새땅 Part1
새하늘과새땅 Part1
 
제2회 난공불락 오픈소스 인프라 세미나 Kubernetes
제2회 난공불락 오픈소스 인프라 세미나 Kubernetes제2회 난공불락 오픈소스 인프라 세미나 Kubernetes
제2회 난공불락 오픈소스 인프라 세미나 Kubernetes
 
제2회난공불락 오픈소스 세미나 커널튜닝
제2회난공불락 오픈소스 세미나 커널튜닝제2회난공불락 오픈소스 세미나 커널튜닝
제2회난공불락 오픈소스 세미나 커널튜닝
 

Recently uploaded

Grid Layout (Kitworks Team Study 장현정 발표자료)
Grid Layout (Kitworks Team Study 장현정 발표자료)Grid Layout (Kitworks Team Study 장현정 발표자료)
Grid Layout (Kitworks Team Study 장현정 발표자료)
Wonjun Hwang
 
파일 업로드(Kitworks Team Study 유현주 발표자료 240510)
파일 업로드(Kitworks Team Study 유현주 발표자료 240510)파일 업로드(Kitworks Team Study 유현주 발표자료 240510)
파일 업로드(Kitworks Team Study 유현주 발표자료 240510)
Wonjun Hwang
 

Recently uploaded (6)

도심 하늘에서 시속 200km로 비행할 수 있는 미래 항공 모빌리티 'S-A2'
도심 하늘에서 시속 200km로 비행할 수 있는 미래 항공 모빌리티 'S-A2'도심 하늘에서 시속 200km로 비행할 수 있는 미래 항공 모빌리티 'S-A2'
도심 하늘에서 시속 200km로 비행할 수 있는 미래 항공 모빌리티 'S-A2'
 
Grid Layout (Kitworks Team Study 장현정 발표자료)
Grid Layout (Kitworks Team Study 장현정 발표자료)Grid Layout (Kitworks Team Study 장현정 발표자료)
Grid Layout (Kitworks Team Study 장현정 발표자료)
 
오픈소스 위험 관리 및 공급망 보안 솔루션 'Checkmarx SCA' 소개자료
오픈소스 위험 관리 및 공급망 보안 솔루션 'Checkmarx SCA' 소개자료오픈소스 위험 관리 및 공급망 보안 솔루션 'Checkmarx SCA' 소개자료
오픈소스 위험 관리 및 공급망 보안 솔루션 'Checkmarx SCA' 소개자료
 
클라우드 애플리케이션 보안 플랫폼 'Checkmarx One' 소개자료
클라우드 애플리케이션 보안 플랫폼 'Checkmarx One' 소개자료클라우드 애플리케이션 보안 플랫폼 'Checkmarx One' 소개자료
클라우드 애플리케이션 보안 플랫폼 'Checkmarx One' 소개자료
 
[OpenLAB] AWS reInvent를 통해 바라본 글로벌 Cloud 기술동향.pdf
[OpenLAB] AWS reInvent를 통해 바라본 글로벌 Cloud 기술동향.pdf[OpenLAB] AWS reInvent를 통해 바라본 글로벌 Cloud 기술동향.pdf
[OpenLAB] AWS reInvent를 통해 바라본 글로벌 Cloud 기술동향.pdf
 
파일 업로드(Kitworks Team Study 유현주 발표자료 240510)
파일 업로드(Kitworks Team Study 유현주 발표자료 240510)파일 업로드(Kitworks Team Study 유현주 발표자료 240510)
파일 업로드(Kitworks Team Study 유현주 발표자료 240510)
 

제4회 한국IBM과 함께하는 난공불락 오픈소스 인프라 세미나-AI

  • 1. Kwanghoon Jhin kwanghoon.jhin@kr.ibm.com March, 2017 PowerAI World Fastest Machine Learning Platform IBM Power Systems: S822LC for HPC IBM Storage: Elastic Storage Server Revision E1
  • 3. 3
  • 4. 4
  • 5. 1950 — Alan Turing creates the “Turing Test” 1952 — Arthur Samuel wrote the first computer learning program. The program was the game of checkers, and the IBM computer improved at the game the more it played, studying which moves made up winning strategies and incorporating those moves into its program. 1959 – “Field of study that gives computers the ability to learn without being explicitly programmed” Arthur Samuel A (very short) History of AI
  • 6. 6 A (very short) History of AI IBM 701 vacuum tube plug-in unit 700-series Electronic Data Processing Machine.
  • 7. 1957 — Frank Rosenblatt designed the first neural network for computers (the perceptron), which simulate the thought processes of the human brain. A (very short) History of AI
  • 8. A (very short) History of AI
  • 9. A (very short) History of AI
  • 10. “… the navy revealed the embryo of an electronic computer today that it expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence ... Dr. Frank Rosenblatt, a research psychologist at the Cornell Aeronautical Laboratory, Buffalo, said Perceptrons might be fired to the planets as mechanical space explorers” 08-JUL-1958, the New York Times. A (very short) History of AI 과도한 자신감이 거짓 믿음을 만들다.
  • 11. Perceptrons (1969) by Marvin Minsky, Founder of the MIT AI Lab – We need to use Multi-Layer Perceptrons – No one on earth had found a viable way to train MLPs good enough to learn such simple functions. 11 A (very short) History of AI
  • 12. 12 AI winter: – 1974 to 1980 – 1987 to 1993
  • 13. Backpropagation – 1974 by Paul Werbos – 1982 by Paul Werbos – 1986 by Geoffrey Hinton 13 A (very short) History of AI
  • 14. Paul Werbos (1974). Beyond regression: New tools for prediction and analysis in the behavioral sciences. PhD thesis, Harvard University. Paul Werbos (1982). Applications of advances in nonlinear sensitivity analysis. In System modeling and optimization (pp. 762-770). Springer Berlin Heidelberg. Rumelhart, David E.; Hinton, Geoffrey E.; Williams, Ronald J. (8 October 1986). "Learning representations by back-propagating errors". Nature. 323 (6088): 533–536. 14 A (very short) History of AI Backpropagation / 역전파(逆傳播)
  • 15. 1997 — IBM’s Deep Blue beats the world champion at chess. 2006 — Geoffrey Hinton coins the term “deep learning” to explain new algorithms that let computers “see” and distinguish objects and text in images and videos. 2010 — The Microsoft Kinect can track 20 human features at a rate of 30 times per second, allowing people to interact with the computer via movements and gestures. 2011 — IBM’s Watson beats its human competitors at Jeopardy. 2011 — Google Brain is developed, and its deep neural network can learn to discover and categorize objects much the way a cat does. 2012 — In 2012, Hinton and two of his other U of T students, Alex Krizhevsky and Ilya Sutskever, entered an image recognition contest. Competing teams build computer vision algorithms that learn to identify objects in millions of pictures; the most accurate wins. The U of T model took the error rate of the best-performing algorithms to date, and the error rate of the average human, and snapped the difference in half like a dry twig. The trio created a company, and Google acquired it. (https://www.thestar.com/news/world/2015/04/17/how-a-toronto-professors-research-revolutionized-artificial- intelligence.html) 15 A (very short) History of AI
  • 16. 2012 — Google’s X Lab develops a machine learning algorithm that is able to autonomously browse YouTube videos to identify the videos that contain cats. 2014 — Facebook develops DeepFace, a software algorithm that is able to recognize or verify individuals on photos to the same level as humans can. 2015 — Amazon launches its own machine learning platform. 2015 — Microsoft creates the Distributed Machine Learning Toolkit, which enables the efficient distribution of machine learning problems across multiple computers. 2015 — Over 3,000 AI and Robotics researchers, endorsed by Stephen Hawking, Elon Musk and Steve Wozniak (among many others), sign an open letter warning of the danger of autonomous weapons which select and engage targets without human intervention. 2016 — Google’s artificial intelligence algorithm beats a professional player at the Chinese board game Go, which is considered the world’s most complex board game and is many times harder than chess. The AlphaGo algorithm developed by Google DeepMind managed to win five games out of five in the Go competition. 16 A (very short) History of AI http://www.forbes.com/sites/bernardmarr/2016/02/19/a-short-history-of-machine-learning-every-manager-should-read/
  • 17. 17 PowerAI > 왜 지금인가? 긴 ‘인공지능의 겨울(AI Winter)’를 끝내고 ‘인공지능의 봄(AI Spring)’을 맞이하다. Big DataTechnologyAlgorithm 비정형 데이터의 범람하드웨어 능력의 비약적인 발전관습과 한계에 도전 그리고 성취 학계에서는 지금의 시대를 ‘인공지능의 봄’이라고 부릅니다. 그 이유는 이전까지 없었던 아래의 세 가지가 갖추어졌기 때문입니다. 인공지능 설계의 오래된 난제를 해결 수퍼 컴퓨터에서만 가능하던 연산을 책상 위에서 처리 기계학습에 새로운 가능성과 의미를 부여
  • 18. 18 PowerAI > 기계학습에 요구되는 컴퓨팅 구성 요소들 지금 이 시대에 필요한 컴퓨팅 플랫폼 GPU computingFrameworks 한계없는 데이터의 저장과 사용을 위한 소프트웨어 정의 스토리지 가장 혁신적인 I/O 기술을 탑재한 GPU 컴퓨팅 시스템 IBM이 최적화환 프래임워크 및 라이브러리 IBM은 PowerAI라는 이름으로 ‘최적화된 프레임워크’, ‘GPU 컴퓨팅 시스템’ 그리고 ‘소프트웨어 정의 스토리지’를 하나의 컴퓨팅 플랫폼으로 제공하고 있습니다. IBM Optimized Frameworks IBM S822LC for HPC IBM Elastic Storage Server Software Defined Storage
  • 19. 19 IBM이 최적화 한 ML/DL 프레임워크 IBM은 ibm.biz/powerai 를 통하여 S822LC for HPC에 최적화된 ML/DL 프레임워크를 제공합니다. OpenBLAS BazelDIGITS NCCLDistributed Frameworks Machine Learning / Deep Learning Frameworks Supporting Libraries 기계학습 플랫폼을 계획하고 구축하는데에 많은 시간을 투자해야 하는 부분이 바로, 프래임워크 최적화입니다. IBM은 이와 같은 현장의 불편과 불필요한 시간투자를 최소화할 수 있는 ‘최적화’된 프래임워크를 제공합니다.
  • 20. GPU의 기원 – 1999 – Nvidia, GeForce 256를 출시하면서 “the world’s first GPU”라고 지칭. 그리고 다음과 같이 GeForce 256 GPU를 설명 • “Single-chip processor with integrated transform, lighting, triangle setup/clipping, and rendering engines” – 2002 – Nvidia의 경쟁사인, ATI Technologies (AMD)에서 Radeon 9700을 발표하면서 VPU, ‘Visual Processing Unit’이라고 지칭 – 2007 – Nvidia, GPGPU인 Tesla 제품군 발표 Co-Processor로 GPGPU 활용 – HPC – Machine Learning 20 S822LC for HPC > Graphic Processing Unit
  • 21. 21 GPU 컴퓨팅은 CPU 컴퓨팅과 무엇이 다른가? Mythbusters at NVISION 08
  • 22. 22 1 2 5 12 9 6 0 3 5 72 9 -15 5 16 -32 -20 18 20 98 -58 78 1 + 2 + 5 + 12 + 9 + 6 + 0 + 3 + 5 + 72 + 9 + (-15) + 5 + 16 + (-32) + (-20) = 78 16개의 수를 합하는 연산을 해야 합니다. 주어진 16개의 수를 차례대로 합하여 총합산을 끝내는 방식(직렬연산)과, 4개씩 짝을 지어 동시에 연산하는 방식(병렬연산)이 있을 수 있겠습니다. 전자의 방식이 일반적인 CPU 연산의 방식이라고 할 수 있고, 후자가 GPU가 취하는 방식입니다. GPU는 병렬 처리를 효율적으로 처리하기 위한 수천 개의 코어를 가지고 있습니다. S822LC for HPC > GPU는 연산에 어떤 도움을 주는가?
  • 23. Although NVLink primarily focuses on connecting multi also connect Tesla P100 GPUs with IBM Power CPUs shows how the CPUs are connected with NVLink in th configuration, each GPU has 180 GBps bidirectional b and 80 GBps bidirectional bandwidth to the connected Figure 1-12 CPU to GPU and GPU to GPU interconnect us All the initialization of the GPU is through the PCIe inte the side band communication for status, power manag and running, all data communication is using the NVLi 23 S822LC for HPC > NVLink NVLink로 PCIe의 한계를 뛰어 넘다. GPU 제작사인 Nvidia는 GPU 컴퓨팅의 큰 문제 중 하나인 CPU(host)와 GPU(device) 사이 그리고 GPU(device)와 GPU(device) 사이의 병목현상에 주목해 왔습니다. 그리고 그 해결책으로 NVLink라는 bus를 선보였습니다. Nvidia는 IBM와 손잡고, 세계 최초 NVLink 1.0을 완벽 구현하기에 이르렀으며, S822LC for HPC에 적용되었습니다. IBM은 NVLink 기술을 높게 평가하여 2.0 버전을 차세대 CPU 플랫폼인 Power9에도 적용하기로 결정했습니다. PCIe 3.0 x16 32 GB/s bidirectional NVLink 160 GB/s bidirectional X5 bandwidth 4 GPUs with NVLink
  • 24. 24 S822LC for HPC > Unified Memory 지금까지의 GPU 메모리 접근의 한계를 없애다. Unified Memory • 전세대 GPU, Kepler와 Maxwell의 한계로 멈추어 있던 GPU 사용 메모리 크기 제한(GPU 내 설계에 따른 전용 메모리, GDDR을 벗어날 수 없음)이 Pascal에서는 제거되어, 시스템에 탑재된 모든 메모리를 사용 가능 • 이제 개발자들이 GPU 메모리 내의 데이터 이동 관리보다 연산 자체에 집중하는 것이 가능 ‘두껍고도 수평적으로’ (both Fat and Flat) 설계 • 어느 연결에서도 데이터 전송 병목이 생기지 않도록 설계 • GPU에서도 CPU처럼 시스템 메모리를 직접 접근하고 사용 • 같은 소켓의 GPU간 ‘Fat Pipe’ 구현 보편적 업무와 연산 알고리즘에 최적화 • Startup/Teardown시 폭발적인 성능 • GPU간의 안정적 데이터 이송 • 기존의 부족한 대역폭으로 인한 host-device간의 bus transfer 문제를 해소 Pascal GPU Power8 CPU Unified Memory CUDA 8 Allocated beyond GPU memory Up to 1TB on S822LC for HPC
  • 25. 25 S822LC for HPC > 성능 유사 조건에서 GPU 세대간 성능비교 0 20 40 60 80 100 120 140 x86 with 4x M40 / PCIe Power8 with 4x P100 / NVLink Training time (minutes): AlexNet and Caffe to top-1, 50% Accuracy (Lower is better) 0:00 1:12 2:24 3:36 4:48 6:00 7:12 8:24 x86 with 8x M40 / PCIe Power8 with 4x P100 / NVLink BVLC Caffe vs IBM Caffe / VGGNet Time to Top-1 50% accuracy: (Lower is better) • IBM S822LC 20-cores 2.86GHz 512GB memory / 4 NVIDIA Tesla P100 GPUs / Ubuntu 16.04 / CUDA 8.0.44 / cuDNN 5.1 / IBM Caffe 1.0.0-rc3 / Imagenet Data • Intel Broadwell E5-2640v4 20-core 2.6 GHz 512GB memory / 8 NVIDIA Tesla M40 GPUs / Ubuntu 16.04 / CUDA 8.0.44 / cuDNN 5.1 / BVLC Caffe 1.0.0-rc3 / Imagenet Data S822LC/HPC with 4 Tesla P100 Tesla GPUs is 24% Faster than 8x Tesla M40 GPUs S822LC/HPC with 4 Tesla P100 GPUs is 2.2x Faster than 4x Tesla M40 GPUs x2.2 24%
  • 26. 26 S822LC for HPC > 성능 유사 조건에서 NVLink와 PCIe와의 비교 X86 + GeForce Pascal Architecture on PCIe 3.0 x16 POWER8 + Tesla Pascal Architecture on NVLink ImageNet / Alexnet: Minibatch size = 128 170 ms 78 ms IBM advantage: data communication and GPU performance 54% Pascal Architecture로 제작된 Tesla P100과 GeForce GTX Titan X를 직접 비교한 결과, NVLink의 유무가 전체 처리 속도를 크게 좌우한다는 것을 알 수 있었습니다. • 데이터 이동시간의 획기적인 단축 • 연산 대기시간 단축으로 전체적인 부하 감소 • 총 연산시간의 실질적인 단축
  • 27. 27 S822LC for HPC > 내부구조 및 구성 2U 19inch Rack Mount, 2-Socket POWER8 CPU, 4 Tesla P100 GPU, 32 DDR4 DIMM slots, 3 PCIe slots. POWER8 with NVLink (2x) • 190W • Integrated NVLink 1.0 Memory DIMM’s Riser (8x) • 4 IS DDR4 DIMMs per riser • Single Centaur per riser • 32 IS DIMM’s total • 32-1024 GB memory capacity PCIe slot (3x) • Gen3 PCIeNvidia GPU • SXM2 form factor • NVLink 1.0 • 300 W • Max of 2 per socket Power Supplies (2x) • 1300W • Common Form Factor SupplyCooling Fans (4x) • 80mm Counter- Rotating Fans • Hot swap Storage Option (2x) • 0-2, SATA HDD.SSD • Tray design for install/removal • Hot Swap Service Controller Card • BMC Content
  • 28. 28 S822LC for HPC > 내부구조 및 구성 > 수냉식 냉각장치 수냉식 냉각장치 옵션 적용 기본 공냉식 냉각장치 수냉식 냉각장치 제공 WATER COOLING OPTION S822LC for HPC는 대규모 HPC 컴퓨팅 시스템 제작에 필수적인 수냉식 냉각방식을 옵션으로 제공하고 있습니다. GPU 컴퓨팅에서 필연적으로 따라오는 발열문제를 근본적으로 해결하고 다수의 S822LC for HPC을 구성한 다중 랙을 운영할 때에도 보다 안정적인 컴퓨팅 파워를 확보함과 동시에 전산실 내 공조 시스템의 부하를 줄일 수 있습니다. 수냉식 냉각옵션은 4GPU 구성에만 제공되며, 아래 그림과 같이 방렬판에 직결되어 냉각제가 공급되는 호스가 설치됩니다.
  • 29. 29 S822LC for HPC > Nvidia Tesla 제품군 비교 S822LC for HPC는 ‘P100 (NVLink)’를 사용합니다. P100 (NVLink) P100 (PCIe) M40 K40 K80 Architecture PASCAL PASCAL Maxwell Kepler Kepler CUDA cores 3840 3840 3072 2880 4992 (dual GPU) Half-precision performance (FP16) 21.2 TeraFLOPS with NVIDIA GPU Boost 18.68 TeraFLOPS with NVIDIA GPU Boost N/A N/A N/A Single-Precision Performance (FP32) 10.6 TeraFLOPS with NVIDIA GPU Boost 9.34 TeraFLOPS with NVIDIA GPU Boost 7 Teraflops with NVIDIA GPU Boost 4.29 Teraflops with NVIDIA GPU Boost 8.73 Teraflops with NVIDIA GPU Boost Double-Precision Performance (FP64) 5.3 TeraFLOPS with NVIDIA GPU Boost 4.67 TeraFLOPS with NVIDIA GPU Boost 0.2 Teraflops with NVIDIA GPU Boost 1.43 Teraflops with NVIDIA GPU Boost 2.91 Teraflops with NVIDIA GPU Boost GPU Memory 16 GB HBM2 12GB or 16GB HBM2 24 GB GDDR5 12 GB GDDR5 24 GB GDDR5 (12 GB per GPU) Memory Bandwidth 720 GB/s HBM2 3072-bit HBM2 (12GB) 4096-bit HBM2 (16GB) 288 GB/s 288 GB/s 480 GB/s (240 GB/s per GPU) System Interface 160 GB/s bidirectional interconnect bandwidth with NVIDIA NVLink 32 GB/s PCI Express 3.0 x16 32 GB/s PCI Express 3.0 x16 32 GB/s PCI Express 3.0 x16 32 GB/s PCI Express 3.0 x16 Max Power Consumption 300 W 250 W 250 W 235 W 300 W (150W per GPU)
  • 30. 30 IBM Elastic Storage Server Power Systems S822L (Server) & S812L (Mgmt) EXP24 (GS) or DCS3700 (GL) or DeepFlash150 (GF) Spectrum Scale Software Defined Storage Elastic Storage Server Compute Disks Software IBM Elastic Storage Server는 IBM Power Systems를 기반으로 Linux 운영체제 그리고 IBM Spectrum Scale이라는 파일 시스템을 중심으로 구성되는 ‘소프트웨어 정의 스토리지’이자 스토리지 어플라이언스입니다.
  • 31. 31 IBM Elastic Storage Server > 개요 IBM Elastic Storage Server는 Spectrum Scale의 모든 장점을 그대로 발휘합니다. § Spectrum Scale: – 다양한 플랫폼상에서 고성능, 고가용 파일시스 템 서비스 및 데이터 관리 서비스를 제공할 수 있도록 설계된 시스템 소프트웨어. – 1993년 발표된 이후 오랜 기간 안정화 및 버전 업을 통해 성숙화된 파일시스템 솔루션 클러스터 내 최대 노드 수 Architectural Limit 16,384 Maximum Tested 9,300 파일시스템 크기 Architectural Limit 2^99bytes (524,288 YiB) Maximum Tested 18PiB 마운드할 수 있는 최대 파일시스템 수 256 파일시스템당 최대 파일 수 Architectural Limit 2^64 Maximum Tested 9,000,000,000 단일 파일시스템에서 테스트된 최대 I/O 대역폭 400GB/s 최대 Dependent fileset 수 10,000 최대 Independent fileset 수 1,000 최대 Global Snapshot 수 256 각각의 Independent fileset에 대한 최대 Snapshot 수 256 병렬/분산 처리 파일시스템 Global Single Name Space 소프트웨어 또는 어플라이언스로 제공 Ethernet/Infiniband를 통한 고속 데이터 처리
  • 32. 32 IBM Elastic Storage Server > 제품군 다양한 구성을 지원하며, 속도중심 또는 용량중심의 제품을 선택할 수 있습니다. All Flash도 지원합니다. Model GL4 4 Enclosures, 20U 232 NL-SAS, 2 SSD 10+ GB/Sec Model GL6 6 Enclosures, 28U 348 NL-SAS, 2 SSD 15+ GB/sec Model GL2 2 Enclosures, 12U 116 NL-SAS, 2 SSD 5+ GB/Sec 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 FC5887 Model GS1 24 SSD 6 GB/Sec Model GS2 46 SAS + 2 SSD or 48 SSD Drives 2 GB/Sec SAS 12 GB/Sec SSD Model GS4 94 SAS + 2 SSD or 96 SSD Drives 5 GB/Sec SAS 16 GB/Sec SSD Model GS6 142 SAS + 2 SSD 7 GB/Sec Model GF2 2 Enclosures, 10u 360TB, usable Flash Max Read 26 GB/s Max Write 16 GB/s Model GF1 1 Enclosure, 7U 180TB, usable Flash Max Read 13 GB/s Max Write 9 GB/s • GS models use 2U 24x2.5” JBODs or SSDs > 최대 120TB • GL models use 4U 60x3.5” JBODs > 최대 2PB • GF models use 3U 32xFlash JBOFs enclosures > 최대 360TB • Support drives: 1.8TB SAS, 400GB, 800GB, 1.6TB SSD 2.5”; 2TB,4TB,6TB and 8TB NL-SAS 3.5” HDDs • Supported NICs: 10GbE, 40GbE Ethernet and EDR Infiniband
  • 33. 33 IBM Elastic Storage Server > HPC 고객 도입사례 “기존의 일반적인 RAID 스토리지 어레이에서는 1TB 디스크를 리빌드하는데 12시간 전후가 소요 되었는데, GNR은 2TB 디스크를 1시간 이내에 리빌드가 가능하도록 했으며, 연구 시뮬레이션을 위한 고속 데이터 엑세스를 가능하도록 해 주었다.” Lothar Wollschläger Storage Specialist @ Jülich Supercomputing Centre IBM Elastic Storage Server • 총 20대로 구성 • 7 PB (usable capacity) • 200 GB/s I/O 성능 IBM Elastic Storage Server에서 제공되는 ‘GPFS Native RAID’를 통해 디스크 리빌드(disk rebuild) 시간을 획기적으로 단축시켜 리빌드 중 발생 가능한 2차 장애의 위험성이 현저하게 감소
  • 34. IBM Systems HardwarePowerAI World Fastest Machine Learning Platform THANK YOU!