Multiple users fighting for resources = bad G Don't allow them to do so directly |omputer users are greedy G Det the system allocate resources [sers like to know job status without having to keep an open session
| |
| | î
Do-it-yourself (DIY) ^S|AR Rocks Pelican P| (formerly Parallel Knoppix) Microsoft Windows ||E ^penMosix (closed March 2008) |lustermatic (no activity since 2005)
| |
| { | Advantages G |ontrol G Dearning Experience Disadvantages G |ontrol G Administration
| |
| { |
ardware Requirements ead Node Deployment G |ore Software Requirements G |luster-specific Software G |onfiguration Adding compute nodes
| |
|
Several commodity computers: G cpu/motherboard G memory G ethernet card G hard drive (recommended, in most cases) Network switch |ables, etc.
| |
| î
|ore system G system logger, core utilities, mail, etc. G Dinux Kernel Network Filesystem (NFS) server support Additional Packages G Secure Shell (SS) server G iptables (firewall) G nfs-utils G portmap G Network Time Protocol (NTP) | |
| î
Additional Packages (cont.) G inetd/xinetd ± For FTP, globus, etc. G Message Passing Interface (MPI) package G Scheduler ± PBS, SGE, |ondor, etc. G Ganglia ± Simplified |luster ³ealth´ Dogging dependency: Apache Web Server
| Install compute node image on the compute node G Manually G PXE Boot (pxelinux, etherboot, etc.) G Redat Kickstart G etc. |onfigure host name, NFS, etc. ... for each node!
| |
| a
Software updates in head node require update in compute node Failed nodes must be temporarily removed from head node configuration files
| |
|
| But what if my boss wants a 200-node cluster? G Monster.com G ^R come up with your own automation scheme G ^R [se ^S|AR or Rocks
| |
| | î
Do-it-yourself (DIY) ^S|AR Rocks Pelican P| (formerly Parallel Knoppix) Microsoft Windows ||E ^penMosix (closed March 2008) |lustermatic (no activity since 2005)
| |
| î| ^pen Source |luster Application Resources Fully-integrated software bundle to ease deployment and management of a cluster Provides G Management Wizard G |ommand-line tools G System Installation Suite
| |
|
Install ^S|AR-approved Dinux distribution Install ^S|AR distribution |reate node image(s) Add nodes Start computing
| |
| î|a
Download/install/remove ^S|AR packages Build a cluster image Add/remove cluster nodes |onfigure networking Reimage or test a node with the Network Boot Manager
| |
| î||
Everything the Wizard offers yume G [pdate node packages | - The |luster |ommand |ontrol Tools G provide cluster-wide versions of common commands G |oncurrent execution G example 1: copy a file from the head node to all visualization nodes G example 2: execute a script on all compute nodes | |
| | | cexec: execution of any standard command on all cluster nodes ckill: terminates a user specified process cget: retrieves files or directories from all cluster nodes cpush: distribute files or directories to all cluster nodes cpushimage: update the system image on all cluster nodes using an image captured by the SystemImager tool | |
|
/etc/c .conf describes cluster configuration cluster gcb { gcb.fiu.edu head node dead placeholder change command line to 1 indexing compute-0-[0-8] first set of nodes exclude 5 offline node in the range (killed by J. Figueroa) } -------
| |
| a The ^S|AR Password Installer and [ser
Management Synchronize user accounts Set up passwordless SS Periodically check for changes in passwords
| |
| î î System Installation Suite Installs Dinux systems over a network Image-based Allows different images for different nodes Nodes can be booted from network, floppy, or |D.
| |
| | î
Do-it-yourself (DIY) ^S|AR Rocks Pelican P| (formerly Parallel Knoppix) Microsoft Windows ||E ^penMosix (closed March 2008) |lustermatic (no activity since 2005)
| |
|
# Disadvantages G Tight-coupling of software G ighly-automated Advantages G ighly-automated... G But also flexible
| |
|
# The following 25 slides are property of [| Regants
|luster-fork ± execute command on nodes (serially) |luster-kill ± kill a process on all nodes |luster-probe ± get information about cluster status |luster-ps ± query nodes for a running process by name
| |
| &
Basis: Redhat Kickstart file G plus XMD flexibility G and dynamic stuff (i.e. support for ³macros´) Image Docation: /export/home/install |ustomization: rolls and extend-compute.xml |ommand: rocks-dist
| |
| ù
# X-Windows Related G X, fonts, display manager Display wall G XDMX, |hromium, SAGE
| |
| a '| ù
X configuration G xorg.conf G Xinitrc Display Manager |onfiguration
| |
| a Rocks Directory: /var/411 G |ommon configuration files: Autofs-related /etc/group, /etc/passwd, /etc/shadow /etc/services, /etc/rpc G All encrypted elper |ommand G rocks-user-sync
| |
| î|
Rocks is now installed |hoose an MPI runtime G MPI| G ^penMPI G DAM-MPI Start compiling and executing
| |
|
| Dive|D for instant cluster creation Advantages G Easy to use G A lot of built-in software Disadvantages G Not persistent G Difficult to add software
| |
| a
î
Windows Server 200 |ompute |luster Edition (||E) Microsoft |ompute |luster pack (||P) Microsoft MPI (based on MPI|2) Microsoft Scheduler
