Experience and Future Directions
Andrew A. Chien
CTO, Entropia, Inc.
SAIC Chair Professor
Computer Science and Engineering, UCSD
National Computational Science Alliance
Invited Talk, USENIX Windows, August 4, 2000
Critical Enabling Technologies
The Alliance’s Windows Supercluster
– Design and Performance
Other Windows Cluster Efforts
Future
– Terascale Clusters
– Entropia
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
100
Microprocessors
MIPS R2000 (125)
MIPS R3000 (40)
10
Cray 1S (12.5)
Cray X-MP (8.5)
Cray Y-MP (6)
HP 7000 (15)
R4000 (10)
R4400 (6.7)
1
Vector supercomputers
X86/Alpha
(1)
1975 1980 1985 1990 1995
Micros: 10MF -> 100 MF -> 1GF -> 3GF -> 6GF (2001?)
Year
Introduced
=> Memory system performance catching up (2.6 GB/s 21264 memory
BW)
GigSAN/GigE: 110 MB/s
UW Scsi: 40 MB/s
LAN: 10Mb/s -> 100Mb/s -> ?
SAN: 12MB/s -> 110MB/s
(Gbps) -> 1100MB/s -> ?
– Myricom, Compaq, Giganet, Intel,...
Network bandwidths limited by system internal memory bandwidths
Cheap and very fast communication hardware
FastE: 12 MB/s
Ethernet 1MB/s
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
HD Storage
Networks
Graphical Interfaces
Audio/Graphics
Clustering, Performance,
Multiprocess Protection
Mass store, HP networking,
Management, Availability, etc.
SMP support
Basic device access
1981 1985 1990 1995 1999
Desktop (PC) operating systems now provide
– richest OS functionality
– best program development tools
– broadest peripheral/driver support
– broadest application software/ISV support
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Cluster management and resource integration (“use like” one system)
Delivered communication performance
– IP protocols inappropriate
Balanced systems
– Memory bandwidth
– I/O capability
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Goals
– Enable tightly coupled and distributed clusters with high efficiency and low effort (integrated solution)
– Provide usable access thru convenient standard parallel interfaces
– Deliver highest possible performance and simple programming model
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Early 1990’s, Gigabit testbeds
– 500Mbits (~60MB/s) @ 1 MegaByte packets
– IP protocols not for Gigabit SAN’s
Cluster Objective: High performance communication to small and large messages
Performance Balance Shift: Networks faster than I/O, memory, processor
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
User-level network access
Lightweight protocols
– flow control, reliable delivery
– tightly-coupled link, buffer, and I/O bus management
Poll-based notification
Streaming API for efficient composition
Many generations 1994-1999
– [IEEE Concurrency, 6/97]
– [Supercomputing ’95, 12/95]
Related efforts: UCB AM, Cornell U-Net,RWCP PM,
Princeton VMMC/Shrimp, Lyon BIP => VIA standard
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
250
200
150
100
50
0
1995 1996 1997 1998 1999
MB/s
20MB/s -> 200+ MB/s (10x)
– Much of advance is software structure: API’s and implementation
– Deliver *all* of the underlying hardware performance
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
25
20
15
10
5 microseconds
0
1995 1996 1997 1998 1999
100 m s to 2 m s overhead (50x)
– Careful design to minimize overhead while maintaining throughput
– Efficient event handling, fine-grained resource management and interlayer coordination
– Deliver *all* of the underlying hardware performance
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
MPI Put/Get
Global
Arrays
Fast Messages
Myrinet
Server-
Net
Giganet
VIA
SMP
BSP
WAN
Scheduling
& Mgmt (LSF)
Performance
Tools
HPVM 1.0 (8/1997)
HPVM 1.2 (2/1999)
- multi, dynamic, install
HPVM 1.9 (8/1999)
- giganet, smp
Turnkey Cluster Computing; Standard API’s
Network hardware and API’s increase leverage for users, achieve critical mass for system
Each involved new research challenges and provided deeper insights into the research issues
– Drove continually better solutions (e.g. multi-transport integration, robust flow control and queue management)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
120
100
80
60
40
20
0
FM on Myrinet
MPI on FM-Myrinet
• N
1/2
~ 400 Bytes message size (bytes)
Delivers underlying performance for small messages, endpoints are the limits
100MB/s at 1K vs 60MB/s at 1000K
– >1500x improvement
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
80
90
70
60
50
40
30
20
10
FM on Giganet VIA
MPI-FM on Giganet VIA
• N
1/2
~ 400 Bytes
0
4
10
24
29
44
39
68
49
92
60
16
70
40
80
64
90
88
10
11
2
11
13
6
12
16
0
13
18
4
14
20
8
15
23
2
16
25
6 message size (bytes)
FM Protocol/techniques portable to Giganet VIA
Slightly lower performance, comparable N
1/2
Commercial version: WSDI (stay tuned)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Procs
<space>
Fixed Size
Frames
Variable Size
Data
Increasing
Addresses
Networks
Fixed Size Trailer
+ Length/Flag
Solution: Uniform notify and poll (single Q representation)
Scalability: n into k (hash); arbitrary SMP size or number of NIC cards
Key: integrate variable-sized messages; achieve single DMA transfer
– no pointer-based memory management, no special synchronization primitives, no complex computation
Memory format provides atomic notification in single contiguous memory transfer (bcopy or DMA)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Myrinet (latency)
Single Transport Integrated
8.3
m s
Myrinet (BW) 101MB/s
Shared Memory (latency) 3.4
m s
Shared Memory (BW) 200+MB/s
8.4
m s
101MB/s
3.5
m s
200+MB/s
No polling or discontiguous access performance penalties
Uniform high performance which is stable over changes of configuration or the addition of new transports
– no custom tuning for configuration required
Framework is scalable to large numbers of SMP processors and network interfaces
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Cray T3E
MF/Proc Flops/Byte Flops/NetworkRT
1200 ~2 ~2,500
SGI Origin2000 500
HPVM NT Supercluster 600
~0.5
~8
~1,000
~12,000
IBM SP2 (4 or 8-way) 2.6-5.2GF ~12-25 ~150-300K
Beowulf (100Mbit) 600 ~50 ~200,000
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Windows
Early prototypes in CSAG
– 1/1997, 30P, 6GF
– 12/1997, 64P, 20GF
Alliance’s Supercluster
– 4/1998, 256P, 77GF
– 6/1999, 256P*, 109GF
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
128 HP Kayak XU
Dual PIII 550 MHz/1GB RAM
AS-PCG MPI Performance - 2D Navier Stokes Kernel
20
Engineering Fluid Flow Problem
18
128 300 MHz Intel Pentium II +
128 550 MHz Pentium III Xeon
16
14
12
10
Origin
550 MHz
Using NT, Myrinet Interconnect, and HPVM
8
6
4
2
0
300 MHz
SGI O2000, 250 MHz R10000
NT Cluster: Intel 550 MHz PIII Xeon HP Kayak
NT Cluster: Intel 300MHz PII HP Kayak
Cluster: 128 550MHz + 128 300 MHz
#207 in Top 500
Supercomputing Sites
0 32
D. Tafti, NCSA
64 96 128
Processors
160 192 224 256
Rob Pennington (NCSA), Andrew Chien (UCSD)
Front-End Systems
Fast Ethernet
File servers
LSF master
Internet
• Apps development
• Job submission
LSF Batch
Job Scheduler
FTP to Mass Storage
Daily backups
128 GB Home
200 GB Scratch
128 Compute Nodes, 256 CPUs
Windows NT, Myrinet and HPVM
128 Dual 550 MHz Systems
Infrastructure and Development Testbeds
Windows 2K and NT
8 4p 550 + 32 2p 300 + 8 2p 333
(courtesy Rob Pennington, NCSA)
MILC – QCD
Navier-Stokes Kernel
Zeus-MP – Astrophysics CFD
Large-scale Science and Engineering codes
Comparisons to SGI O2K and Linux clusters
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
12
10
IA-32/Win NT, 300 MHz PII
250 MHz SGI O2K
T3E 900
IA-32/Win NT 550MHz Xeon
8
6
4
2
0
0 50
Processors
Src: D. Toussaint and K. Orginos, Arizona
100
10000
9000
8000
7000
6000
5000
4000
3000
2000
1000
0
1 4 16 32 64 96 128 192 256
# procs
SGI O2K
Janus (ASCI Red)
NT Supercluster
550 Mhz
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
AS-PCG MPI Performance - 2D Navier Stokes Kernel
20
18
16
14
12
10
8
6
4
2
0
0 32 64
128 300 MHz Intel Pentium II +
128 550 MHz Pentium III Xeon
96
SGI O2000, 250 MHz R10000
NT Cluster: Intel 550 MHz PIII Xeon HP Kayak
NT Cluster: Intel 300MHz PII HP Kayak
Cluster: 128 550MHz + 128 300 MHz
128
Processors
160 192 224 256
Source: Danesh Tafti, NCSA
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Zeus-MP (256P, Mike Norman)
ISIS++ (192P, Robert Clay)
ASPCG (256P, Danesh Tafti)
Cactus (256P, Paul Walker/John Shalf/Ed Seidel)
MILC QCD (256P, Lubos Mitas)
QMC Nanomaterials (128P, Lubos Mitas)
Boeing CFD Test Codes, CFD Overflow (128P, David Levine) freeHEP (256P, Doug Toussaint)
ARPI3D (256P, weather code, Dan Weber)
GMIN (L. Munro in K. Jordan)
DSMC-MEMS (Ravaioli)
FUN3D with PETSc (Kaushik)
SPRNG (Srinivasan)
MOPAC (McKelvey)
Astrophysical N body codes (Bode)
=> Little code retuning and quickly running ...
Parallel Sorting (Rivera
– CSAG),
18.3 GB Minutesort World Record
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Sort max data disk-to-disk in 1 minute
“Indy sort”
– fixed size keys, special sorter, and file format
HPVM/Windows Cluster winner for 1999 (10.3GB) and 2000 (18.3GB)
– Adaptation of Berkeley NOWSort code (Arpaci and
Dusseau)
Commodity configuration ($$ not a metric)
– PC’s, IDE disks, Windows
– HPVM and 1Gbps Myrinet
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Kayak
Netserver
HPVM & 1Gbps Myrinet
Kayak
32 HP Kayaks
3Ware Controllers
4 x 20GB IDE disks
32 HP Netservers
2 x 16GB SCSI disks
Concurrent read/bucket-sort/communicate is bottleneck – faster I/O infrastructure required (busses and memory, not disks)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Gossip: “Windows platforms are not reliable”
– Larger systems => intolerably low MTBF
Our Experience: “Nodes don’t crash”
– Application runs of 1000s of hours
– Node failure means an application failure; effectively not a problem
Hardware
– Short term: Infant mortality (1 month burn-in)
– Long term
• ~1 hardware problem/100 machines/month
• Disks, network interfaces, memory
• No processor or motherboard problems.
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
NT Cluster Usage by Number of Processors
May1999 to Jul2000
500000
400000
300000
200000
100000
0
1 - 31 32 - 63 64 - 256
Number of Processors
Lots of large jobs
Runs up to ~14,000 hours (64p * 9 days)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Sandia’s Kudzu Cluster (144 procs, 550 disks, 10/98)
Cornell’s AC 3 Velocity Cluster (256 procs, 8/99)
Others (sampled from vendors)
– GE Research Labs (16, Scientific)
– Boeing (32, Scientific)
– PNNL (96, Scientific)
– Sandia (32, Scientific)
– NCSA (32, Scientific)
– Rice University (16, Scientific)
– U. of Houston (16, Scientific)
– U. of Minnesota (16, Scientific)
– Oil & Gas (8,Scientific)
– Merrill Lynch (16, Ecommerce)
– UIT (16, ASP/Ecommerce)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
The AC 3 Velocity
64 Dell PowerEdge 6350 Servers
• Quad Pentium III 500 MHz/2 MB Cache Processors (SMP)
• 4 GB RAM/Node
• 50 GB Disk (RAID 0)/Node
GigaNet Full Interconnect
• 100 MB/Sec Bandwidth between any 2 Nodes
• Very Low Latency
2 Terabytes Dell PowerVault 200S Storage
• 2 Dell PowerEdge 6350 Dual Processor File Servers
• 4 PowerVault 200S Units/File Server
• 8 36 GB/Disk Drives/PowerVault 200S
• Quad Channel SCSI Raid Adapter
• 180 MB/sec Sustained Throughput/ Server #381 in Top 500
2 Terabyte PowerVault 130T Tape Library
• 4 DLT 7000 Tape Drives
• 28 Tape Capacity
(courtesy David A. Lifka, Cornell TC)
Supercomputing Sites
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Recent AC 3 Additions
8 Dell PowerEdge 2450 Servers (Serial Nodes)
• Pentium III 600 MHz/512 KB Cache
• 1 GB RAM/Node
• 50 GB Disk (RAID 0)/Node
7 Dell PowerEdge 2450 Servers (First All NT Based AFS Cell)
• Dual Processor Pentium III 600 MHz/512 KB Cache
• 1 GB RAM/Node Fileservers, 512 MB RAM/Node Database servers
• 1 TB SCSI based RAID 5 Storage
• Cross platform filesystem support
64 Dell PowerEdge 2450 Servers (Protein Folding, Fracture
Analysis)
• Dual Processor Pentium III 733 Mhz/256 KB Cache
• 2 GB RAM/Node
• 27 GB Disk (RAID 0)/Node
• Full Giganet Interconnect
3 Intel ES6000 & 1 ES1000 Gigabit switches
• Upgrading our Server Backbone network to Gigabit Ethernet
(courtesy David A. Lifka, Cornell TC)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
3
Only commercially supported technology
– Rapid spinup and spinout
– Package technologies for vendors to sell as integrated systems
=> All of commercial packages were moved from SP2 to Windows, all users are back and more!
Users: “I don’t do windows” =>
“I’m agnostic about operating systems, and just focus on getting my work done.”
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Protein Folding
The cooperative motion of ion and water through the gramicidin ion channel. The effective quasi-article that permeates through the channel includes eight water molecules and the ion. Work of Ron Elber with
Bob Eisenberg, Danuta Rojewska and Duan Pin.
Reaction path study of lig and diffusion in leghemoglobin. The ligand is CO (white) and it is moving from the binding site, the heme pocket, to the protein exterior. A study by
Weislaw Nowak and Ron Elber.
http://www.tc.cornell.edu/reports/NIH/resource/CompBiologyTools/
(courtesy David A. Lifka, Cornell TC)
Protein Folding Per/Processor Performance
Results on different computers for a protein structures:
Machine
Blue Horizon (SP
San Diego)
Linux cluster
Velocity (CTC)
Velocity+ (CTC)
System
AIX 4
CPU
Power3
Linux 2.2
PentiumIII
Win 2000 PentiumIII
Xeon
Win 2000 PentiumIII
CPU speed
[MHz]
222
650
500
733 compiler xlf
PGF 3.1
df v6.1
df v6.1
Results on different computers for ( a / b or b proteins):
Machine
Blue Horizon (SP
San Diego)
Linux cluster
System
AIX 4
Linux 2.2
CPU
Power3
PentiumIII
CPU speed
[MHz]
222
650 compiler xlf
PGF 3.1
Velocity (CTC)
Velocity+ (CTC)
Win 2000 PentiumIII
Xeon
Win 2000 PentiumIII
500
733 df v6.1
df v6.1
(courtesy David A. Lifka, Cornell TC)
Energy evaluations per second
44.3
59.1
46.0
59.2
Energy evaluations per second
15.0
21.0
16.9
22.4
AC 3 Corporate Members
-Air Products and Chemicals
-Candle Corporation
-Compaq Computer Corporation
-Conceptual Reality Presentations
-Dell Computer Corporation
-Etnus, Inc.
-Fluent, Inc.
-Giganet, Inc.
-IBM Corporation
-ILOG, Inc.
-Intel Corporation
-KLA-Tencor Corporation
-Kuck & Associates, Inc.
-Lexis-Nexis
-MathWorks, Inc.
-Microsoft Corporation
-MPI Software Technologies, Inc.
-Numerical Algorithms Group
-Portland Group, Inc.
-Reed Elsevier, Inc.
-Reliable Network Solutions, Inc.
-SAS Institute, Inc.
-Seattle Lab, Inc.
-Visual Numerics, Inc.
-Wolfram Research, Inc.
(courtesy David A. Lifka, Cornell TC)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Good performance
Lots of Applications
Good reliability
Reasonable Management complexity (TCO)
Future is bright; uses are proliferating!
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
NT Supercluster, NCSA
– http://www.ncsa.uiuc.edu/General/CC/ntcluster/
– http://www-csag.ucsd.edu/projects/hpvm.html
AC3 Cluster, TC
– http://www.tc.cornell.edu/UserDoc/Cluster/
University of Southampton
– http://www.windowsclusters.org/
=> application and hardware/software evaluation
=> many of these folks will work with you on deployment
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Communication Hardware
– Myrinet, http://www.myri.com/
– Giganet, http://www.giganet.com/
– Servernet II, http://www.compaq.com/
Cluster Management and Communication Software
– LSF, http://www.platform.com/
– Codeine, http://www.gridware.net/
– Cluster CoNTroller, MPI, http://www.mpi-softtech.com/
– Maui Scheduler http://www.cs.byu.edu/
– MPICH, http://www-unix.mcs.anl.gov/mpi/mpich/
– PVM, http://www.epm.ornl.gov/pvm/
Microsoft Cluster Info
– Win2000 http://www.microsoft.com/windows2000/
– MSCS,http://www.microsoft.com/ntserver/ntserverenterprise/ exec/overview/clustering.asp
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Terascale Clusters
Entropia
10+ Teraflops in 2000?
NSF currently running a $36M Terascale competition
Budget could buy
– an Itanium cluster (3000+ processors)
– ~3TB of main memory
? #1 in Top 500 ?
Supercomputing Sites
– > 1.5Gbps high speed network interconnect
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
• COTS, SHV enable larger, cheaper, faster systems
•
Supercomputers (MPP’s) to…
•
Commodity Clusters (NT Supercluster) to…
• Entropia
Idea: Assemble large numbers of idle PC’s in people’s homes, offices, into a massive computational resource
– Enabled by broadband connections, fast microprocessors, huge PC volumes
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Entropia network: ~30,000 machines (and growing fast!)
– 100,000, 1Ghz => 100 TeraOp system
– 1,000,000, 1Ghz => 1,000 TeraOp system (1 PetaOp)
IBM ASCI White (12 TeraOp, 8K processors, $110 Million system)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Millions have demonstrated willingness to donate their idle cycles
“Great Cause” Computing
– Current: Find ET, Large Primes, Crack DES…
– Next: find cure for cancer, muscular dystrophy, air and water pollution, …
• understand human genome, ecology, fundamental properties of matter, economy
Participate in science, medical research, promoting causes that you care about!
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Heterogeneity (machine, configuration, network)
Scalability (thousands to millions)
Reliability (turn off, disconnect, fail)
Security (integrity, confidentiality)
Performance
Programming
. . .
Entropia: harnessing the computational power of the Internet
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Power : a network with unprecedented power and scale
Empower : ordinary people to participate in solving the great social challenges and mysteries of our time
Solve : team solving fascinating technical problems
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Windows clusters are powerful, successful high performance platforms
– Cost effective and excellent performance
– Poised for rapid proliferation
Beyond clusters are Internet computing systems
– Radical technical challenges, vast and profound opportunities
For more information see
– HPVM http://www-csag.ucsd.edu/
– Entropia http://www.entropia.com/
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
NT Cluster Team Members
– CSAG (UIUC and UCSD Computer Science) – my research group
– NCSA Leading Edge Site -- Robert Pennington’s team
Talk materials
• NCSA (Rob Pennington, numerous application groups)
• Cornell TC (David Lifka)
• Boeing (David Levine)
• MPISoft (Tony Skjellum)
• Giganet (David Wells)
• Microsoft (Jim Gray)
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA
Entropia, Inc -- University of California, San Diego (UCSD/CSE) -- NCSA