Transcript Diapositiva 1 - Taiwan Evolutionary Intelligence Laboratory

TEIL-Clusters
Peng Chun-Jen
Taiwan Evolutionary Intelligence Laboratory
2016/11/23 Group Meeting Presentation
Outline
• Brief review on Computer Networking
• Common Services in Clusters
• Parallel Computing
Outline
• Brief review on Computer Networking
– Server host
– Server Application
• Common Services in Clusters
• Parallel Computing
Server
• A Program or Device that provides services for other
programs or devices
• Software : Client-Server model
– Split tasks and workload between
service provider (server) and requester (client)
• Hardware : Server-host
– runs one or more server at the same time
– Could be a client at the same time
Server (Device)
• Host: any computer connected to a network
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
http://https://max-media.imgix.net/transfers/2016/10/12/97772b5e3b8ec2fad77d4ad66e3b10edf3d6c806.gif
http://i.lv3.hbo.com/assets/images/series/westworld/homepage/2016/september/160922-westworld-key-art-1920.jpg
Server (Device)
• Host: any computer connected to a network
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
• Appearance:
– Workstation 工作站
Tower Server 塔式
• Slightly larger than PC
https://www.gamesclan.com/immagini/dedicati-h.jpg
• Scalability
Server (Device)
• Host: any computer connected to a network
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
• Appearance:
– Tower Server 塔式
– Rack Server 機架式
• Easier to maintain servers in bays or slots of rack
• 22U ~ 42U (1U = 44.45mm)
https://s.aolcdn.com/hss/storage/midas/c4ffd6e734b74ca2a7a
fac615a9cc603/202504538/google-servers-2015-08-20-01.jpg
Server (Device)
• Host: any computer connected to a network
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
• Appearance:
– Tower Server 塔式
– Rack Server 機架式
– Blade Server 刀鋒伺服器
http://www.colomachine.com/colomachinev3/Blade-System/CB1100/Web-Image/pcblade-diag1.gif
Server (Device)
• Host: any computer connected to a network
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
• Appearance:
– Tower Server 塔式
– Rack Server 機架式
– Blade Server 刀鋒伺服器
http://siliconangle.com/files/2013/08/smtp-server.jpg
– Multi-node 多節點 (窮人版刀鋒伺服器)
Server (Device)
• Appearance: Tower, Rack, Blade
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
– Hubs(集線器), Switch(交換器), Router(路由器)
Server (Device)
• Appearance: Tower, Rack, Blade
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
– Hubs(集線器) : boardcast network, half-duplex(半雙工)
– Switch(交換器)
– Router(路由器)
http://aitkotw.blogspot.tw/2015/04/computer-hubs-and-types-of-hub.html
Server (Device)
• Appearance: Tower, Rack, Blade
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
– Hubs(集線器)
: boardcast network, half-duplex(半雙工)
– Switch(交換器) : memorize MAC, full-duplex(全雙工)
– Router(路由器) :
http://http://www.grd.ru/images/price/1372128268.jpeg
Server (Device)
• Appearance: Tower, Rack, Blade
• Physical network node:
– Host: Computers
– Hubs(集線器) :
– Switch(交換器) :
– Router(路由器) :
ip-sharer (ip 分享器)
http://www.wiringdiagrams21.com/wp-content/uploads/2012/01/SamsungUA40D6000SRLEDTVtoLANNetworkCableConnectionDiagram.jpg
http://www.itworldcanada.com/wp-content/uploads/2013/09/Haweui-AR3200-enterprise-router-cropped.jpg
Server (Device)
• Appearance: Tower, Rack, Blade
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
– Hubs(集線器), Switch(交換器), Router(路由器)
• Network:
– Ethernet : Carrier Sense Multiple Access/Collision Detection
– InfiniBand : More efficient mechanism than TCP
Server (Device)
• Appearance: Workstation, Rack, Blade
https://www.asus.com/tw/support/faq/1000761/
http://image.wuji8.com/upload/5/7c/57c634e3b9cf9d27706e3ed864917c22.jpg
• Physical network node:
– Host: Computers (i.e. server-hosts, user-host)
– Hubs(集線器), Switch(交換器), Router(路由器)
• Network:
– Ethernet : 100Mbps ~ 10Gbps
– InfiniBand : 20Gbps
Server (Device)
• Appearance : Workstation, Rack, Blade
• Network node: Host, Hubs, Switch, Router
• Network
: Ethernet, InfiniBand
• Goal: (RASUM)
=>
Cloud Service
– Reliability
- Usability
– Avalibility
- Manageability
– Scalability
- Modularity?
Server (Device)
• TEIL-Cluster1 at EE2-R202
Server (Device)
• TEIL-Cluster2 at BL-615
Server (Device)
• TEIL-Cluster2 at BL-615
Server (Program)
• Architecture:
– A computer(host) could be a client, a server, or both
– Server-hosts have more resources then client-hosts?
• Master-Slave :
• Client-Server:
• Peer-to-Peer:
Server (Program)
• Master-Slave :
– service provider (slave) and requester (master)
– unidirectional control
• Client-Server:
– service provider (server) and requester (client)
– Server can automatically send information to the Clients
Server (Program)
• Architecture:
Server (Program)
• Client-Server:
– Problem: Scalability and Reliability
– Heavy Load on Server
• P2P:
– Increase speed in system
– Requires better hardware for each hosts
– Security issues: confidentiality, malware, traffic
https://storage.googleapis.com/peer5/Billy/client%20server%20vs%20p2p.png
Server (Program)
• Daemon ( Demon vs. Daemon XD )
– Program that provides a certain service
– init daemon
http://daemon.pl/daemon.jpg
• /etc/init.d/daemon start or stop or status or restart
• chkconfing daemon on or off
– systemd
• systemctl start or stop or status or restart daemon
• systemctl enable or disable
Outline
• Brief review on Computer Networking
• Common Services in Clusters
– High Performance Computing
– NFS
– NIS
• Parallel Computing
Server (Program)
• DNS Server : Domain Name System
• DHCP Server : Dynamic Host Configuration
• NFS Server : Network File System
• NIS Server : Network Information Service
• NAT Server : Network Address Translation
• Web Server: Apache, Nginx
• Monitoring: collectd, Ganglia
Network File System (NFS)
• NFS
– Mount Remote File System
– Deault port 2049
• RPC(Remote Procedure Call)
– Randomly pick free port < 1024
Network File System (NFS)
• /etc/fstab
– mount at boot
– Problem: traffic & lock
• Autofs
– mount when access
– Problem: need root privilege to mount
Network Information Service(NIS)
• Sun Yellow Pages (yp)
MPI vs. Spark
• Explosion need in mid-2000
– Analyzing internet-scale data and flood of genomics data
– Scalability, accuracy, large-scale data storage
http://www.dursi.ca/hpc-is-dying-and-mpi-is-killing-it/
MPI vs. Spark
• Explosion need of HPC in mid-2000
– Analyzing internet-scale data and flood of genomics data
– Scalability, accuracy, large-scale data storage
http://www.dursi.ca/hpc-is-dying-and-mpi-is-killing-it/
MPI vs. Spark
• Explosion need in mid-2000
• People create new parallel computing stacks
• HPC Community don’t consider other platforms as
"real" high performance computing
• HPC is wedded to 25-year old technology stack which
doesn't meet the needs of those communities
MPI vs. Spark
• MPI (Message Passing Interface)
– "killer apps" at initial of cluster computing
– useful collective operations for sending and receiving
messages
• Why MPI is the wrong tool for today
– wrong level of abstraction for app writers & tool builders
– MPI is more than you need for modest levels of parallelism
MPI vs. Spark
• API accessibility : Spark > Hadoop > MPI
• Speed : MPI > Hadoop = Spark
• Spark is suitable for iterative computation (in memory)
e.g. Machine Learning / Data analytics
• High demand of experts in ML + Distributed Compute
Reference
1.
2.
3.
http://linux.vbird.org/
https://www.server-world.info/en/
http://www.dursi.ca/hpc-is-dying-and-mpi-is-killing-it/