Networking - DESY · Industry Exhibits Exhibitor Forum HPC Games Posters SCinet99 Webcasts...
Transcript of Networking - DESY · Industry Exhibits Exhibitor Forum HPC Games Posters SCinet99 Webcasts...
COMPAQ ALPHA
1
1/29/2007 Supercomputing 99 1
Networking
1/29/2007 Supercomputing 99, Processors 2
Conference facilities
• State of the art show floor network:– multiple OC-192 (10GBit/s) rings– Dense Wave Division Multiplex (DWDM)
• External network at OC48 (2.5 Gbit/s)• In house Gigabit Ethernet• Demo of 10-G Ethernet (formerly Xnet)
COMPAQ ALPHA
2
1/29/2007 Supercomputing 99, Processors 3
ASCI Networking plans
1/29/2007 Supercomputing 99, Processors 4
Internet development forecast
• Talk of Vinton Cerf (excerpts)
COMPAQ ALPHA
3
1/29/2007 Supercomputing 99, Processors 5
Internet Hosts (000s) 1989-2006
0
100000
200000
300000
400000
500000
600000
700000
800000
900000
1000000
1989
1991
1993
1995
1997
1999
2001
2003
2005
hosts
COMPAQ ALPHA
4
1/29/2007 Supercomputing 99, Processors 7
Observations
• 75% of traffic on Internet is WWW• Data Domination (20% voice, 80% data)• Traffic growth 100-1000%/year reported• 300 M - 1000 M users by Dec 2000
• Internet MUCH faster growing than CPU power
1/29/2007 Supercomputing 99, Processors 8
Internet-enabled Devices• Information appliances
– 1997 - 3 M, 1998 - 6 M, 2002 - 56 M (IDC)• WebTV, Palm-Pilot, Nokia 9000,Sony,
Nintendo, Sega games• Refrigerator (and the bathroom scales)• Automobiles, household appliances (turning
a box of soap into a service) • Web-server on a chip (see next slide)
COMPAQ ALPHA
5
1/29/2007 Supercomputing 99, Processors 9
UMASS Web server on a chipborn 10 AM, 14 July 1999
• TCP/IP code itself fits in about 256 bytes (12-bit)
• PIC 12C509A, running at 4MHz
• 24LC256 i2c EEPROM • HTTP 1.0 and RFC 1122
compliant
• eternity.cs.umass.edu:9080/index0.html
Our 25 year mission: to go whereno network has gone before!
Space: the final frontier
COMPAQ ALPHA
6
1/29/2007 Supercomputing 99, Processors 11
1/29/2007 Supercomputing 99, Processors 12
COMPAQ ALPHA
7
1/29/2007 Supercomputing 99, Processors 13
•End-to-end information flow across the solar system
•Layered architecture for evolvability and interoperability
•IP-like protocol suite tailored to operate over long round trip light times
•Integrated communications and navigation services
1/29/2007 Supercomputing 99, Processors 14
Interplanetary Internet Status
• Part of the Mars Mission Plan• Possible Earth/Moon mission 2001• Low Mars Orbit and Areosynchronous
satellites by 2008• Mars Outposts by 2010• Possible Orbiting manned mission 2018• Possible Manned Mars station 2030??• Stable Interplanetary backbone 2040?
COMPAQ ALPHA
8
1/29/2007 Supercomputing 99 15
Grid Computing / Batch
1/29/2007 Supercomputing 99, Processors 16
COMPAQ ALPHA
9
1/29/2007 Supercomputing 99, Processors 17
Grid computing
• geographically distributed computing• similar to Metacomputing in Europe• several toolkits to enable GRID computing
– (compare with UNICORE in Europe)• GLOBUS• LEGION• others
1/29/2007 Supercomputing 99, Processors 18
GLOBUS projectshttp://www.globus.org
• GUSTO – Testbed (as shown on SC 98)
• CACTUS – parallel finite difference simulation codes
• CMT– Microtomography
• Flash– Seamless access to remote computing
COMPAQ ALPHA
10
1/29/2007 Supercomputing 99, Processors 19
LEGION
• Worldwide virtual computer• Middleware that connects computer resources• http://legion.virginia.edu• Used e.g. at DoD, NASA, NPACI
– (NPACI: National Partnership for Advanced Computational Infrastructure)
1/29/2007 Supercomputing 99, Processors 20
Legion status monitor (Java)
COMPAQ ALPHA
11
1/29/2007 Supercomputing 99, Processors 21
Batch systems
• PBS (Portable Batch system)– developed at NASA/Ames– mature public domain batch system focused on parallel
computing (sophisticated job scheduling)– no AFS support
• LSF (Platform Computing)– the market leader in US
• Codine (Gridware, formerly Genias, Chord)
1/29/2007 Supercomputing 99, Processors 22
Batch system genealogy
COMPAQ ALPHA
12
1/29/2007 Supercomputing 99 23
Linux Clusters
1/29/2007 Supercomputing 99, Processors 24
Linux and SGI
• Demo of an Itanium cluster running Linux• Statement to release software in the public domain• Committed to Linux• SGI Linux high performance Clusters
– Beowulf style– Advanced Cluster Environment– new product line SGI 1400 (PIII, Redhat 6)
COMPAQ ALPHA
13
1/29/2007 Supercomputing 99, Processors 25
Linux and IBM
• Committed to Linux as well• Hardware cluster activities (see e.g. next slide)
– web servers (Netfinity servers)
• Involvement in Software– mainly focused on desktop and web(Lotus Domino, Websphere, DB2, ViaVoice)
1/29/2007 Supercomputing 99, Processors 26
Large Clusters
• Chiba City– built by ANL, IBM and VALinux– 256 Dual Pentium beowulf system
• Product “Cluster City” from VA Linux– comes with VACM management software(GPL)– allows complete remote access to all resources
COMPAQ ALPHA
14
1/29/2007 Supercomputing 99 27
Other topics
1/29/2007 Supercomputing 99, Processors 28
Top 500 Supercomputers
• 1. Sandia National Labs (ASCI Red) 9632 Intel Proc. 2.3TFlops• 2. Lawrence Livermore (ASCI Blue) 5808 IBM 604e 2.1TFlops• 3. Los Alamos (ASCI Blue Mntn) 48 Origin2000/128 1.6TFlops• 5. Uni Tokio Hitachi SR8000 128 Proc 873GFlops• 9. Deutscher Wetterdienst Offenbach 812Proc T3E 671GFlops• 20. FZ Juelich 540 Processor T3E1200 448GFlops• 500. USA(Banking) Sun HPC 10000 48 Proc. 33GFlops
COMPAQ ALPHA
15
1/29/2007 Supercomputing 99, Processors 29
Supercomputer Trends
• Strong influence of ASCI• US:Europe:Japan ~ 4:2:1• Doubled speed every 1.2 years (other: 1.6 years)• Increasing number of commercial installations• Increasing role of cluster solutions• Constant number of vector computers
1/29/2007 Supercomputing 99, Processors 30
Zero administration terminal
• Login facility at conference provided by Sun• Equipped with Sunray 1
– Terminal is basically a remote framebuffer – Virtual framebuffer on server is sent to terminal– operated on separate Ethernet– smartcard with “hot desking”, “plug and work”
• Cost ~1000DM + Monitor + 1/25 SunServer• Problem: Security and maintenance on server
COMPAQ ALPHA
16
1/29/2007 Supercomputing 99 31
Keynote Address
State of the Field Talks
Invited Talks
Technical Papers
Tutorials
Awards
Panels
Birds-of-a-Feather (BOFS)
Education Program
Research Exhibits
Industry Exhibits Exhibitor Forum
HPC Games
Posters
SCinet99
Webcasts
Compilers Grid ComputingHigh-Performance Networking Industrial and Commercial Applications I/O Low-Level Architecture MPI Non-Numerical Algorithms Numerical Algorithms Ocean and Climate Performance Profiling Scheduling Scientific Applications Special Purpose SystemsVisualization Wide Area ApplicationsFernbach Award and Gordon Bell Finalists
1/29/2007 Supercomputing 99, Processors 32
COMPAQ ALPHA
17
1/29/2007 Supercomputing 99 33
Processor Architectures
1/29/2007 Supercomputing 99, Processors 34
Alpha Roadmap
1997 1998 19991995 1996 2000 2001
EV5/333 EV5/333 2116421164
EV6/575 EV6/575 2126421264
EV68/1000 EV68/1000 2126421264
PCA56/533 PCA56/533 21164PC21164PC
EV56/600EV56/6002116421164
0.5μm
0.35μm
0.35μm
0.35μm
EV67/750 EV67/750 2126421264
0.28μm
PCA57/600 PCA57/600 21164PC21164PC
0.28μm
0.18μm
0.18μm
Higher Performance
Low
er C
o st
...
EV8EV80.13μm
EV7/1000 EV7/1000 2136421364
COMPAQ ALPHA
COMPAQ ALPHA
18
1/29/2007 Supercomputing 99, Processors 35
Int RegMap
Branch Predictors
21364 Core
FETCH MAP QUEUE REG EXEC DCACHEStage: 0 1 2 3 4 5 6
L2 cache1.5MB6-Set
IntIssue
Queue(20)
Exec
4 Instructions / cycle
RegFile(80)
Victim Buffer
L1 DataCache64KB2-Set
FPRegMap
FP ADDDiv/Sqrt
FP MUL
Addr
80 in-flight instructionsplus 32 loads and 32 stores Addr
Miss Address
Next-LineAddress
L1 Ins.Cache64KB2-Set
Exec
Exec
ExecRegFile(80)
FP Issue
Queue(15)
RegFile(72)
COMPAQ ALPHA
1/29/2007 Supercomputing 99, Processors 36
Integrated Memory Controller
• Direct RAMbus– High data capacity per pin– 800 MHz operation– 30ns CAS latency pin to pin
• 6 GB/sec read or write bandwidth• 100s of open pages• Directory based cache coherence• ECC SECDED
COMPAQ ALPHA
COMPAQ ALPHA
19
1/29/2007 Supercomputing 99, Processors 37
Integrated Network Interface• Direct processor-to-processor
interconnect• 10 GB/second per processor• 15ns processor-to-processor latency• Out-of-order network with adaptive
routing• Asynchronous clocking between
processors• 3 GB/second I/O interface per processor
COMPAQ ALPHA
1/29/2007 Supercomputing 99, Processors 38
COMPAQ ALPHA
21364 System Block Diagram
364M
IO364
M
IO364
M
IO364
M
IO
364M
IO364
M
IO364
M
IO364
M
IO
364M
IO364
M
IO364
M
IO364
M
IO
COMPAQ ALPHA
20
1/29/2007 Supercomputing 99, Processors 39
IBM Power4
1/29/2007 Supercomputing 99, Processors 40
IBM Power4
COMPAQ ALPHA
21
1/29/2007 Supercomputing 99, Processors 41
IBM Power4
1/29/2007 Supercomputing 99, Processors 42
Intel Itanium (IA64)
COMPAQ ALPHA
22
1/29/2007 Supercomputing 99, Processors 43
Intel Itanium
1/29/2007 Supercomputing 99 44
Special Purpose Systems
COMPAQ ALPHA
23
1/29/2007 Supercomputing 99, SPECIAL PURPOSE SYSTEMS
45
Processing-in-memory (PIM) chips that integrate processor logic into memory devices offer a new opportunityfor bridging the growing gap between processor and memory speeds, especially for applications with high
memory-bandwidth requirements.
Mary Hall, Peter Kogge*, Jeff Koller, Pedro Diniz, Jacqueline Chame, Jeff Draper, JeffLaCoss,John Granacki, Jay Brockman*, Apoorv Srivastava, William Athas, Vincent Freeh*,Jaewook Shin, Joonseok ParkUSC Information Sciences Institute * University of Notre DameMarina del Rey, CA 90292 Notre Dame, IN 46556
Data-IntensiVe Architecture DIVA
1/29/2007 Supercomputing 99, SPECIAL PURPOSE SYSTEMS
46
Koji Hashimoto et al.Kyushu University, Fuji Xerox Co.,Ltd, Pharmaceutical Co.,Ltd,Hokkaido University of Education, Shimane University,National Institute for Advanced Interdisciplinary Research
MOE MolecularOrbital calculationEngine
COMPAQ ALPHA
24
1/29/2007 Supercomputing 99, SPECIAL PURPOSE SYSTEMS
47
$7.0/Mflops Astrophysical N -Body Simulationwith Treecode on GRAPE-5Atsushi Kawai, Toshiyuki Fukushige and Junichiro MakinoUniversity of Tokyo
GRAvity PipE GRAPE GORDON BELL PRIZE winner
1/29/2007 Supercomputing 99, SPECIAL PURPOSE SYSTEMS
48
GRAvity PipE GRAPE GORDON BELL PRIZE winner
COMPAQ ALPHA
25
1/29/2007 Supercomputing 99 49
High-speed Interconnect
1/29/2007 Supercomputing 99, High speed interconnect
50
GSN Physical Layer(HIPPI-6400-PH)
– 6400 Mbits (800 Mbytes)/sec bandwidth– Independent full speed, half duplex channels– 4 Virtual Circuits (multiplexing facility)– Small (32 byte) fixed size micropacket– Credit-based flow control– End-to-end & Link-to-link checksums– Automatic retransmit to correct flawed data– Support for legacy HIPPI-800 traffic
Gigabyte System Network GSN
COMPAQ ALPHA
26
1/29/2007 Supercomputing 99, High speed interconnect
51
Silicon Graphics SUMAC TM
ASIC– 32.5 x 32.5 mm 624
pin ceramic CGA– AC GSN port– IC 2 x 64 bit 100 MHz
host interface– 1.25 million cells– 17 watts– Available now
GSNAC
SRC DSTIC
GSN SUMAC Chip
1/29/2007 Supercomputing 99, High-speed interconnect
52
Infiniband: Future I/O + Next Generation I/O (NGIO)
COMPAQ ALPHA
27
1/29/2007 Supercomputing 99, High-speed interconnect
53
Infiniband
1/29/2007 Supercomputing 99, High-speed interconnect
54
Myrinet
COMPAQ ALPHA
28
1/29/2007 Supercomputing 99, High-speed interconnect
55
Myrinet
1/29/2007 Supercomputing 99, High-speed interconnect
56
Myrinet
COMPAQ ALPHA
29
1/29/2007 Supercomputing 99, High-speed interconnect
57
Myrinet
1/29/2007 Supercomputing 99, High-speed interconnect
58
Myrinet
COMPAQ ALPHA
30
1/29/2007 Supercomputing 99, High-speed interconnect
59
ATOLL
1/29/2007 Supercomputing 99, High-speed interconnect
60
ATOLL
COMPAQ ALPHA
31
1/29/2007 Supercomputing 99, High-speed interconnect
61
ATOLL
1/29/2007 Supercomputing 99, Processors 62