What is a Cluster? A cluster is a collection of connected, independent computers that work together to solve a problem.
A Typical Cluster Many standalone computers All of the cluster can work together on a single problem at the same time Portions of the cluster can be working on different problems at the same time Connected together by a network –Larger clusters have separate high speed interconnects Administered as a single “machine”
Some Cluster Acronyms Node – a machine in a cluster Sizes –Kb – kilobyte – thousand bytes – small/medium sized email –Mb – megabyte – million bytes – 2/3 of a 3.5” floppy –Gb – gigabyte – billion bytes – good amount of computer memory or a very old disk drive –Tb – terabyte – trillion bytes – 1/30 th Library of Congress –Pb – petabyte – 1K trillion bytes – 30 Libraries of Congress SMP – symetric multi-processing (many processors) NFS – Network File System HPC – High Performance Computing
Mainframe Vector Supercomputer Mini Computer Workstation PC 1984 Computer Food Chain
How to Build a Supercomputer: 1980’s A supercomputer was a vector SMP (symmetric multi-processor) Custom CPUs Custom memory Custom packaging Custom interconnects Custom operating system Cray* 2 Costs were Extreme: Around ~$5 million/gigaFLOP Technology Evolution Tracking: ~1/3 Moore’s Law Predictions Costs were Extreme: Around ~$5 million/gigaFLOP Technology Evolution Tracking: ~1/3 Moore’s Law Predictions
Mainframe Vector Supercomputer MPP Workstation PC Mini Computer (hitting wall soon) (future is bleak) 1994 Computer Food Chain
How to Build a Supercomputer: 1990’s A supercomputer was an MPP (massively parallel processor) COTS 1 CPUs Costs were High: Around $200K/gigaFLOP Technology Evolution Tracking: ~1/2 Moore’s Law Predictions Costs were High: Around $200K/gigaFLOP Technology Evolution Tracking: ~1/2 Moore’s Law Predictions 1 COTS = Commercial Off The Shelf COTS memory COTS memory Custom packaging Custom packaging Custom interconnects Custom interconnects Custom operating system Custom operating system Intel ® processor based ASCI - Red
NCSA 1990’s Former Cluster ~1,500 processor SGI de-commissioned Too costly to maintain Software too expensive Takes up large amounts of floor space –(Great for tours, looks impressive, nice displays) Gradually being taken out when floor space required Now being used as network file servers
Loki: an Intel ® processor based cluster at Los Alamos National Laborry (LANL How to Build a Supercomputer: 2000’s A Supercomputer is a Cluster COTS 1 CPUs Costs are Modest: Around $4K/gigaFLOP Technology Evolution Tracks Moore’s Law Costs are Modest: Around $4K/gigaFLOP Technology Evolution Tracks Moore’s Law 1 COTS = Commercial Off The Shelf COTS Memory COTS Memory COTS Packaging COTS Packaging COTS Interconnects COTS Interconnects COTS Operating System COTS Operating System
Upcoming Teragrid Clusters Over 4,000 Itanium 2 processors at 4 supercomputer sites –National Center for Supercomputing Applications (NCSA) –San Diego Supercomputer Center (SDSC) –Argonne National Laboratory –California Institute of Technology (Caltech). 13.6 Teraflops computing power (8 teraflops at NCSA) 650 terabytes of disk storage Linked by cross-country 40 Gbit network (16 times faster than the fastest research network currently in existance) –16 minutes to transfer the entire Library of Congress Some uses: –The study of cosmological dark matter –Real-time weather forecasting
Larger Clusters Japan wants to top TOP500 with new cluster 30,000 node cluster planned “Black” clusters (classified) –NSA used to receive large fraction of Cray production Larger ones planned –Scaling problems –Scidac federal mandate to solve scaling problems and enable very large clusters deployment Cooperative widely separated clusters such as SETI
Clustering Today Clustering gained momentum when 3 technologies converged: –1. Very HP Microprocessors –workstation performance = yesterday supercomputers –2. High speed communication –Comm. between cluster nodes >= between processors in an SMP. –3. Standard tools for parallel/ distributed computing & their growing popularity.
What is the Grid ? An infrastructure that couples –Computers (PCs, workstations, clusters, traditional supercomputers, and even laptops, notebooks, mobile computers, PDA, etc) –Databases (e.g., transparent access to human genome database) –Special Instruments (e.g., radio telescope--SETI@Home Searching for Life in galaxy, Austrophysics@Swinburne for pulsars) –People (may be even animals who knows, frogs already planned?) across the local/wide-area networks (enterprise, organisations, or Internet) and presents them as an unified integrated (single) resource.
Network Topologies Cluster has it’s own private network –One or a few outside accessible machines –Most of cluster machines on a private network –Easier to manage –Better security (only have to secure entry machines) –Bandwidth limitations (funneling through a few machines) –Appropriate for smaller clusters –Lower latency between nodes Cluster machines are all on the public network –Academic clusters require this –Some cluster software applications require this –Harder for security (have to secure EVERY machine) –Much higher network bandwidth
Communication Networks 100 Base T (Fast Ethernet) –10 MB/sec (100 Mb/sec) –80-150 microsecond latency –Essentially free Gigabit Ethernet –Typically delivers 30-60 MB/sec –~$1500 / node (going down rapidly)
Message Passing Most parallel computations cluster software requires Message Passing The speed of computations is often dependant on message passing speed as much as raw processor speed Message passing is often done through high speed interconnects because traditional networks are too slow
High Speed Interconnects Myrinet from Myricom (most popular in large clusters) –Proprietary, Myrinet 2000 delivers 200 MB/sec –10-15 microsecond latency –~$1500 / node (going down) –Scales to 1000’s of nodes SCI –Proprietary, good for small clusters –100 MB/sec –~5 microsecond latency Quadrics –Proprietary, very expensive –200 MB/s delivered –5 microsecond latency
InfiniBand – Future of Interconnects? Up to 30 Gbits/second first specifications 15 times faster than fastest high speed interconnects Just now starting to be available commercially Industry standard Will be available from numerous companies
Cluster Software Operating System Choices Linux –Redhat – most popular –Mandrake – similar to Redhat, technically superior FreeBSD, OpenBSD, other BSD’s –Technically superior to Linux’es –Much less popular than Linux’es Windoze
OSCAR Pre-packaged Cluster Software Packaged open source cluster software Designed to support many Unix operating systems –Currently, Redhat Linux –Soon to be released - Mandrake Supported and developed by: –NCSA –IBM –Dell –Intel –Oak Ridge Laboratories Most popular open source cluster software package
Score Pre-Packaged Cluster Software Very popular in Japan Very sophisticated
Scyld/Beuwolf Pre-Packaged Cluster Software Different model – treats cluster of separate machines like one big machine – same process space Oriented towards commercial turn-key clusters Very slick installation Not as flexible – separate machines not accessible
NPACI Rocks Pre-Packaged Cluster Software Based on Redhat Linux Similar to OSCAR Competitor of OSCAR Developed by the San Diego Supercomputer Center and others
OSCAR Overview Open Source Cluster Application Resources Cluster on a CD – automates cluster install process IBM, Intel, NCSA, ORNL, MSC Software, Dell NCSA “Cluster in a BOX” base Wizard driven Nodes are built over network OSCAR <= 64 node clusters for initial target OSCAR will probably be on two 1,000 node clusters Works on PC commodity components RedHat based (for now) Components: Open source and BSD style license