ncar computing update
DESCRIPTION
NCAR Computing Update. Tom Engel Scientific Computing Division National Center for Atmospheric Research Computing in the Atmospheric Sciences Workshop 11 September 2003. NCAR. Managed by UCAR Established 1959 66 Member & 20 Academic Affiliate Institutions. UCAR. - PowerPoint PPT PresentationTRANSCRIPT
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
NCAR Computing Update
Tom EngelScientific Computing Division
National Center for Atmospheric ResearchComputing in the Atmospheric Sciences Workshop
11 September 2003
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
NCAR
• Managed by UCAR• Established 1959• 66 Member & 20 Academic
Affiliate Institutions
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Member Institutions
University Corporation for Atmospheric Research
Board of Trustees
President Richard Anthes
Finance & Administration
Katy Schmoll, VP
Corporate Affairs
Jack Fellows, VP
UCAR Office ofPrograms
Jack Fellows, Director
NCARTim Killeen, Director
Larry Winter, Deputy Director
Education& Outreach
Roberta Johnson
ConstellationObservingSystem for
MeteorologyIonosphere
Climate(COSMIC)
Cooperative ProgramFor Operational
MeteorologyEducation and
Training(COMET)
GPS Science & Technology Program (GST)
UnidataVisiting Scientists
Programs(VSP)
Environmental& Societal
Impacts Group(ESIG)
High Altitude
Observatory(HAO)
Mesoscale &Microscale
Meteorological Division(MMM)
ScientificComputing
Division(SCD)
ResearchApplicationsPrograms
(RAP)
Joint Officefor Science
Support(JOSS)
TimothySpangler
BillKuo
MaryMarlino
Robert Harriss
MichaelKnölker
Bob Gall
BrantFoote
Al Kellie
RandolphWare
MohanRamamurthy
MegAustin
KarynSawyer
AtmosphericChemistry Division
(ACD)
AtmosphericTechnology Division
(ATD)
Advanced StudyProgram(ASP)
Climate & GlobalDynamics Division
(CGD)
MauriceBlackmon
AlCooper
DavidCarlson
DanielMcKenna
Nat’l ScienceDigital Library
(NSDL)
DavidFulker
6/03
Digital Library
for Earth System
Education
(DLESE)
Global Learning and
Observation to Benefit the
Environment(GLOBE)
Jack Fellows
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Scientific Computing Division (SCD)
• enable the best atmospheric research in the world by providing and advancing high-performance computing technologies
• offer computing, research datasets, data storage, networking, and data analysis tools to advance NCAR's scientific research agenda
Purpose
Facility
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
SCD’s Computational Facilities
Two Computational Facilities One Infrastructure
FY03 ComputationalResource Usage
Community42%
Climate Simulation Laboratory
58%
• Climate Simulation Laboratory• Community Facility
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Climate36%
Oceanography11%
Astrophysics12%
Basic Fluid Dynamics2%
Weather Prediction22%
Upper Atmosphere8%
Other6%
Cloud Physics3%
FY2003† Community Facility Computational Resource Usage by
Discipline
† To date: Oct’02-Aug’03
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Oceanography5% Other
1%
Climate94%
FY2003† CSL Facility Computational Resource Usage by Discipline
† To date: Oct’02-Aug’03
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Climate70%
Cloud Physics1%
Other3%
Upper Atmosphere3%
Weather Prediction9%
Basic Fluid Dynamics1%
Astrophysics5%
Oceanography8%
FY2003† Total Computational Resource Usage by Discipline
† To date: Oct’02-Aug’03
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Advanced Research Computing System
• At last CAS ... Al Kellie announced the award of the ARCS Contract to IBM:– Fall ’01: 1.1 TFLOP upgrade to POWER3 system
(blackforest) ... was underway during last CAS– Fall ’02: 4.8 TFLOP† POWER4 system (bluesky)– Fall ’03: Federation switch upgrade– Fall ’04: 8.75 TFLOP† system (bluesky upgrade)
• A few things have changed ...– Fall ’02
• POWER4 capability short of contract commitment• additional initiative funding from NCAR
– bluesky: 38-frame, p690 (Regatta) system, 6.3 TF† capability performance commitments
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
NCAR Computational Facility
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Peak Computing Capacity at NCAR
0.0
1.0
2.0
3.0
4.0
5.0
6.0
7.0
8.0
9.0
Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03
Pea
k T
FL
OP
s at
NC
AR
blackforestWH-I
blackforestupgrade
bluesky
Cray C90/16
HP SPP2000
SGI Origin2000
blackforestWH-II
ARCSPhase 1
ARCSPhase 2
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Bluesky diurnal workload distribution
bluesky diurnal workload distribution (Jan-Aug '03)(1024 total batch CPUs)
0
100
200
300
400
500
600
700
800
900
1000
0 2 4 6 8 10 12 14 16 18 20 22
time of day
Ave
rag
e C
PU
s al
loca
ted
CSL
COM
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Production Workload Distribution - bluesky
Cumulative Workload Distribution - bluesky (Feb'03-Aug'03)
0%
5%
10%
15%
20%
25%
30%
1 p
e
2-3
pe
4 p
e
5-7
pe
8 p
e
9-15
pe
16 p
e
17-2
3 p
e
24 p
e
25-3
1 p
e
32 p
e
33-6
3 p
e
64 p
e
65-1
27 p
e
128
pe
129-
255
pe
256
pe
257-
511
pe
512
pe
513-
1023
pe
1024
pe
>10
24 p
e
% o
f to
tal j
ob
s/C
PU
tim
e
% of alljobs
% of totalreservedCPUhours
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Production Workload Distribution - blackforest
Cumulative Workload Distribution - blackforest (Feb'03-Aug'03)
0%
5%
10%
15%
20%
25%
30%
35%
40%
1 p
e
2-3
pe
4 p
e
5-7
pe
8 p
e
9-15
pe
16 p
e
17-2
3 p
e
24 p
e
25-3
1 p
e
32 p
e
33-6
3 p
e
64 p
e
65-1
27 p
e
128
pe
129-
255
pe
256
pe
257-
511
pe
512
pe
513-
1023
pe
1024
pe
>10
24 p
e
% o
f to
tal j
ob
s/C
PU
tim
e
% of alljobs
% of totalreservedCPUhours
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
August ’03 queue-wait times - bluesky
8-way LPARs
CSL Community
Average Job
Wait Time (min)Standard Deviation # Jobs
Average Job Wait Time (min)
Standard Deviation # Jobs
Premium 0:24 0:59 422 0:06 0:22 519
Regular 0:35 1:23 1366 0:13 1:31 2412
Economy 1:49 4:24 296 0:34 1:30 1059
Stand-by 2:14 8:19 116 0:09 1:06 1932
32-way LPARs
CSL Community
Average Job
Wait Time (min)Standard Deviation # Jobs
Average Job Wait Time (min)
Standard Deviation # Jobs
Premium 0:26 1:29 25 0:27 0:45 48
Regular 0:29 1:09 152 0:32 0:48 116
Economy 0:54 2:02 131 0:41 1:23 164
Stand-by 1:00 2:19 287 0:07 0:23 46
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
August ’03 queue-wait times - blackforest
Winterhawk-2 Nodes
CSL Community
Average Job
Wait Time (min)Standard Deviation # Jobs
Average Job Wait Time (min)
Standard Deviation # Jobs
Premium 0:13 0:30 509 0:00 0:01 5892
Regular 0:11 0:29 2258 0:02 0:18 7290
Economy 0:37 1:31 1644 0:13 0:41 1397
Stand-by 0:59 2:50 130 0:22 1:18 912
Nighthawk-2 Nodes
CSL Community
Average Job
Wait Time (min)Standard Deviation # Jobs
Average Job Wait Time (min)
Standard Deviation # Jobs
_nh 0:00 0:02 453 0:00 0:00 114
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
NCAR MSS Net Growth
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
2.5
Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03
Net
Gro
wth
Rat
e (T
B/d
ay)
UniqueTotal
Dual copyinstituted
CoSinstituted
ARCSPhase 1
ARCSPhase 2
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
NCAR MSS Growth vs. Sustained Computing
0
100
200
300
400
Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03
un
iqu
e b
ytes
/ m
illi
on
flo
atin
g p
oin
t o
per
atio
ns
Unique
Total
Dual copyinstituted
CoSinstituted
ARCSPhase 1
ARCSPhase 2
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
HPM Statistics
HPM Statistics gathered between 07/28/03 and 09/03/03
bluesky8-way
bluesky32-way
blackforestWH-2
blackforestNH-2
Average % Efficiency of User Application Code 3.9 4.3 5.4 9.2
Average MFLOPs per CPU 143.7 141.9 54.9 63.2
Average Daily Peak MFLOPS per CPU 600.2 337.7 238.5 160.3
Average Loads per TLB miss 1491.0 1063.4 1476.1 63.2
Average Data TLB miss rate (/sec) 204475.5 117412.6 - -
Average Instruction TLB miss rate (/sec) 1931.9 2110.8 - -
Average L1 cache hit rate (%) 88.7 89.8 - -
Average L2 cache miss rate (%) 6.1 5.0 - -
Average L3 cache miss rate (%) 32.9 31.1 - -
Average L2 load bandwidth (MB/sec) 1020.9 869.2 - -
Average L3 load bandwidth (MB/sec) 38.8 30.8 - -
Average memory load bandwidth (MB/sec) 116.4 92.8 - -
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
bluesky workload computation rate
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
ARCS Status – Fall ‘03
We thought we had our future planned ...
Federation
IPCC
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Commitment IPCC ScenariosExtra
extra
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
“bluesky upgrade”
• IBM p690 augmentation to bluesky (Sep ’03)– +448 1.3 GHz POWER4 processors (in 14 node x 32
processor configuration) – 2.3 TFLOPs peak
– +.896 TB memory, +10.5 TB disk
• Maintenance, Federation Switch, 2-year ARCS extension renegotiated– Federation now 2H04 option
reduces risk during critical IPCC runs
– Federation ECIP participation 3+4Q03
– 10 TFLOP POWER5 option
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Peak TFLOPs at NCAR
Peak TFLOPs at NCAR
0
2
4
6
8
10
12
Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04
IBM p690/32 Regatta-H (bluesky)
SGI Origin3800/128(chinook)
IBM p690 (16)Regatta (bluedawn)
Compaq ES40/32(prospect)
IBM POWER3(blackforest)
IBM POWER3(babyblue)
SGI Origin2000/128(ute)
HP SPP-2000/64(sioux)
CRI Cray C90/16(antero)
Cray J90 series
Cray T3D
CRI Cray YMP/8(shavano)
blackforestWH-I
ARCS Phase 1blackforest upgrade
ARCS Phase 2bluesky
Cray C90/16
HP SPP2000
SGI Origin2000
blackforestWH-II
ARCS Phase 3 (IPCC)bluesky
SGI Origin3800
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Sustained GFLOPs at NCAR
Sustained GFLOPs at NCAR
0
100
200
300
400
500
Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04
IBM p690/32 Regatta-H (bluesky)
SGI Origin3800/128(chinook)
IBM p690 (16)Regatta (bluedawn)
Compaq ES40/32(prospect)
IBM POWER3(blackforest)
IBM POWER3(babyblue)
SGI Origin2000/128(ute)
HP SPP-2000/64(sioux)
CRI Cray C90/16(antero)
Cray J90 series
Cray T3D
CRI Cray YMP/8(shavano)
blackforestWH-I
ARCS Phase 1blackforest upgrade
ARCS Phase 2bluesky
Cray C90/16
HP SPP2000
SGI Origin2000
blackforestWH-II
ARCS Phase 3 (IPCC)bluesky
SGI Origin3800
Copyright © 2003 University Corporation for Atmospheric Research Sponsored by the National Science Foundation
Thank You
Questions?