Grid Computing 7700Fall 2005

Lecture 4: Scientific Computing and Hardware

Gabrielle Allenallen@bit.csc.lsu.edu

http://www.cct.lsu.edu/~gallen

Basic Elements

CPU CPU

CPU CPU

DISK

Campus Network (LAN)

Machine Network

CPU CPU

CPU CPU

DISK

Campus Network (LAN)

Machine Network

Wide Area Network

Basic Elements

Distributed systems built from– Computing elements (processors)– Communication elements (networks)– Storage elements (disk, attached or networked)

New elements– Visualization/interactive devices– Experimental and operational devices

Distributed Resources

Local workstations CCT Resources Campus/OCS Resources State/LONI Resources National Centers International Colleagues

Laws Moores Law

– Number of transistors on an integrated circuit will double every 18 months– http://en.wikipedia.org/wiki/Moores_law

“Kryders Law”– Hard disk capacity grows quicker than transistors– http://www.sciam.com/article.cfm?chanID=sa006&colID=30&articleID=000B0C2

2-0805-12D8-BDFD83414B7F0000

Gilders Law– Total bandwidth of communication systems doubles every six months

Metcalfe’s Law– Value of a network is proportional to the square of the number of nodes

Amdahl’s Law– Law of diminishing returns, maximum speedup restricted by slowest

parts– http://en.wikipedia.org/wiki/Amdahls_law

Question: So what about applications?

Compute Elements

Moore’s Law: #transistors on a chip (and clockspeed) increase exponentially (double every 18months)– Transistors = 20*2^[(year-1965)/1.5]– 1975 Intel 8080 has 4500 transistors, 100K

intructions/sec– 2003 Pentium IV has 221,000,000, 8 billion

instructions/sec

Corollary: Price of a given level of supercomputingpower halves every 18 months

Price decrease means that supercomputers nowusually built from “commodity” processors– IA32, PowerPC, “emotion engine”

Compute Elements

Clock speed Cache hierarchy Floating point registers Main memory Internal bandwidths Etc, etc Need powerful operating systems,

compilers, applications to leverage all this

Communication Elements

Links, routers, switches, name servers, protocols Infrastructure evolves slowly (politics, large scale changes,

money) Gilder's Law: total bandwidth of communication systems

doubles every six months Change in LAN to desktops

– 100 mbps shared– 100 mbps switched– 1 gbps– 10 gbps

Clusters: GigE (TCP/IP and MPICH/LAM) standard,Myricom/Quadrics (own MPI drivers) better performance,infiniband/fibrechannel different architecture

Network Speeds

Analog modem: 57 kbps GPRS: 114 kbps Bluetooth: 723 kbps T-1: 1.5 Mbps Eth 10Base-X: 10Mbps 802.11b (WiFi) 11 Mbps T-3: 45 Mbps OC-1: 52 Mbps Fast Eth 100Base-X: 100

Mbps

OC-12: 622 Mbps GigEth 1000Base-X: 1 Gbps OC-24: 1.2 Gbps OC-48: 2.5 Gbps OC-192: 10 Gbps 10 GigEth: 10 Gbps OC-3072: 160 Gbps

My Cox Cable– Upload: 35 KB/s– Download 250 KB/s

CCT “is” to supermike– Up/down: 5000 KB/s

Communication Elements

$30005003005Quadrics

$10005002809Myrinet

$100~130~65100GigabitEthernet

Approximatecost per port

BidirectionalBandwidth(mbps)

PeakBandwidth(mbps)

ShortMessageLatency(microsec)

InterconnectType

Storage Elements

Magnetic tape/Magnetic disk Magnetic disk

– Properties: density/rotation/cost– 1970-1988 density improvements 29% per year– 1988-now density improvements 60% per year– Standard in PCs: 500mb (1995), 2gb(1997), 100gb (2002)– Performance not increasing so fast

• Peak transfer (~100mbs)• Seek times (3-5ms) [bottleneck]

Grids: cost of storage neglibable, high speednetworks make large data libraries attractive

The Future (??)

100 tb/s20 pb160 tb640 t-op/s2008 SC

10 gb/s2 tb16 gb64 g-op/s2008 PC

10 tb/s1280 pb50 tb80 t-op/s2003 SC

100 gb/s32 tb256 gb512 g-op/s2013 PC

1 pb/s320 pb2.6 pb5 p-op/s2013 SC

1 gb/s128 gb512 mb8 g-op/s2003 PC

NetworkDiskMemoryComputeMachine

1 mega = 10^61 giga = 10^91 tera = 10^121 peta = 10^15

TeraGrid:40 TFlop/s6 TB memory1 Petabytes storage10 Gigabits/s

Earth Simulator:40 TFlop/s10 TB memory2.5 Petabytes storage13 Gigabits/s

DOE BlueGene:367 TFlop/s16 TB memory400 Terabyte storage

Supercomputers Definition of supercomputer

– Machine on top500.org ?• http://www.top500.org/lists/plists.php?Y=2005&M=06

– Machine costing over $1M ?– Basically highest end machines

Top 3 (2005)– DOE BlueGene/L (USA) 66K procs/137 TF– IBM BGW (USA) 41K procs/91 TF– NASA Columbia (USA) 10K procs/52TF

Top 3 (2003)– Earth Simulator (JAPAN) 5K procs/36 TF (6)– ASCI Q (USA) 8K procs/14 TF (12)– G5 Cluster (USA) 2k procs/12 TF (14)

Others– 18 IBM (China)– 147 Supermike (LSU !!!)

www.webopedia.com

The fastest type of computer.Supercomputers are very expensive andare employed for specializedapplications

that require immense amounts ofmathematical calculations. For example,

weather forecasting requires asupercomputer. Other uses of

supercomputers include animatedgraphics, fluid dynamic calculations,

nuclear energy research, and petroleumexploration.The chief difference betweena supercomputer and a mainframe is thata supercomputer channels all its power

into executing a few programs as fast aspossible, whereas a mainframe uses its

power to execute many programsconcurrently.

Architectural Classes

Flynn (1972): classification based on the way systemmanipulates instruction and data streams:

SISD Single Instruction Single Data– One instruction stream executed serially.– Conventional workstations

SIMD Single Instruction Multiple Data– Large (many thousands) number of processing units– All execute same instruction on different data in lockstep– Vector processors (NEC SX-6i) acting on arrays of data

MISD Multiple Instruction Single Data– No machines built

MIMD Multiple Instruction Multiple Data– Different to SISD because instructions/data are related

More Classification

Shared Memory Systems– Multiple CPUs sharing same address space– One memory accessed by all processors equally– Location of data not important to user– Can be SIMD (single processor vector processor) or MIMD– OpenMP http://www.openmp.org/index.cgi?faq

Distributed Memory Systems– Each CPU has own memory– CPUs are connected by network– Location of data important– Can be SIMD (lock step example before) or MIMD (large

variety of network topologies)– Distributed processing takes DM-MIMD to extreme

Message Passing

Essential for DM machines, but often alsoused for SM machines for compatibility– MPI Message Passing interface– PVM Parallel Virtual Machine

DM-MIMD

Fast growing section, best performance. Need to balancecomputation and communication performance in machinedesign (and upgrades)

User has to distribute data between processors User has to perform data exchange between processors

explicitly Slow compared to SM machines to access data on other

processors Programming models/algorithms important Programming environments can make this easier (e.g. Cactus

Framework http://www.cactuscode.org handles datadistribution, communications, IO, …)

Same programming models need to be extended to Gridcomputing

ccNUMA

Cache Coherent Non Uniform Memory Access Build systems from SMPs (symmetric

multiprocessing nodes) SMPs consist of up to ~16 processors connected

by a crossbar which share same memory Each node is a SM-MIMD, but with different

memory access times for different processors(memory is physically distributed)

Nodes then connecting in a different way Computational scientists like these machines

DM-MIMD

Processor topology and interconnects veryimportant– Hypercube (with 2^d nodes number of steps between

two nodes at most d, possible to simulate othertopologies)

– Fat tree (simple tree structure with more connections athigher levels to ease conjestion)

– 2D/3D mesh structure (many apps map well to this,avoids expense)

– Crossbars (connecting up to around 64 processors, canbe hierarchical)

Details should be hidden from applicationprogrammers, but for performance need to beaware

Virtual Shared Memory

Kendall Square Research Systems tried toimplement at hardware level

High Performance Fortran– HPF Specification 1993– Simulates a virtual shared memory at a software level– Programming directives distribute data across

processors– Looks like shared memory machine to user

Some vendors have propriety virtual sharedmemory programming models by providing globaladdress space

Network Eras

Past (1969-1988)– ARPANET/NSFNET

Current (1988-2005) Future (2005-)

Historical network maps– http://www.cybergeography.org/atlas/historical.html

Network Infrastructure

Chapter 30 (The Grid 2) Network infrastructure is the foundation

on which Grids are built Composition of local and wide area

services, transport protocols and services,routing protocols and network services,link protocols and physical media

One example of network infrastructure inthe Internet (core protocols TCP/IP)

Protocol Agreed-upon format for transmitting data between two devices

which determines:– The type of error checking to be used– Any data compression method– How sending device indicates it has finished sending a message– How receiving device indicates it has received a message

Various standard protocols: differ in simplicity, reliability,performance.

Computer/device must support the right ones to communicatewith other computers.

Implemented either in hardware or in software http://www.protocols.com/protocols.htm

Slow to Change

Internet has not changed much since 1983 (when TCP/IPdeployed), which does make is stable, but still don’t really haveenvisaged services:– Multicast (one-to-many communication)– Network Reservation– Quality of Service

New protocols peer-to-peer file sharing and instant messaging New technology coupled to applications drive change: e-mail,

web/file-sharing, video streaming

Past: 1969-1988 ARPANET (1969) 56-kbps lines

– Experiment to investigate resourcesharing and remote access

– Added interface message processor(IMP) at each end of network (ourrouters), provided flexibility forlower levels and higher levelapplications

– Success from: freely availabledocumentation and source code;software bundled with newmachines; use for teaching;community development vs.proprietary

NSFNET (1985) 45-mpbs lines– Connect academic HPC centers

ARPANET: 1971

ARPANET: 1980

NSFNET: 1991

Past: 1969-1988

Driving application: e-mail, remote file access,remote job control (drove basic protocols)

Network technology: WAN links lines leased fromtelephone companies. Xerox Palo Alto ResearchCenter (PARC) created Ethernet (3 mbps)(alternatives token ring (IBM), …). Workstationsappear bundled with network protocols. PCs onthe network as interface costs dropped andprocessors became more powerful.

Past: 1969-1988

Protocols and Services– telnet, file transfer protocol, e-mail– Underlying transport protocol TCP (stream of

bytes which can be opened or closed, data canbe sent or received)

– Machine location: Domain Name System (DNS)(replaced list of named files)

• Hierarchical, distributed, redundant

Past: 1969-1988

System Integration– ARPANET: assumed central network operations center– NSFNET: introduced hierarchical system, toplevel backbone

network connecting to regional networks connecting tocampuses

Packet switching strategy was important (using computingpower to optimize communication)

Single communication model was important because itallowed so many people to be connected driving futuredevelopment.

Present: 1988-2005

Internet today: complex structure ofbackbone networks and regional networks

Increased role of private sector (e.g.AT&T, BellSouth), who basically controlour network now.

LSU Campus

LANet

Louisiana statewide network:Office of TelecommunicationsManagement, state agencies,higher education: 6Mbps ->$2450 a month

http://www.state.la.us/otm/lanet/

Quest

Bell South

Baton Rouge: 4 DS3 to New Orleans, 1 DS3 to Houston

Abeline (Internet2)

http://abilene.internet2.edu/maps-lists/Traffic: http://loadrunner.uits.iu.edu/weathermaps/abilene/

National Lambda Rail

http://www.nationallambdarail.org/architecture.html

National Lambda Rail

Global Terabit Research Network

Required Reading

Overview of Recent Supercomputers– http://www.euroben.nl/reports/overview05a.pdf

Concentrate on pages 1 to 32, you do not need tolearn this, just get an appreciation of theconcepts.