Announcing the 3rd Graph500 List!

Report
Announcing the 9th Graph500 List!
Graph500 Co-Founders:
David A. Bader, Georgia Tech
Andrew Lumsdaine, Indiana University
Richard Murphy, Micron Technology, Inc.
Marc Snir, Argonne National Laboratory and U. of Illinois
Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company,
for the United States Department of Energy’s National Nuclear Security Administration
under contract DE-AC04-94AL85000.
What is the Graph500?
•New benchmark to complement the Top 500 for large-scale
data analysis problems
•Executive Committee:
– David A. Bader, Peter Kogge, Andrew Lumsdaine, Rich Murphy
•International Multidisciplinary Steering Committee
– Jeremy Kepner (Chair)
– Jim Ang, Brian Barrett, Jon Berry, Bill Brantley, Almadena Chtchelkanova, John
Daly, John Feo, Michael Garland, John Gilbert, Bill Gropp, Bill Harrod, Bruce
Hendrickson, Anton Korzh, Jure Leskovec, Bob Lucas, Mike Merrill, Hans Meuer,
David Mizell, Shoaib Mufti, Richard Murphy, Nick Nystrom, Fabrizio Petrini, Wilf
Pinfold, Steve Poole, Arun Rodrigues, Rob Schreiber, John Simmons, Thomas
Sterling, Blair Sullivan, T.C. Tuan, Jeff Vetter, Mike Vildibill
•Three Kernels
–Search (Concurrent Search, the Ranking Kernel)
–Optimization (Single Source Shortest Path, released)
–Edge Oriented (Maximal Independent Set, in specification)
History of the Graph500
• Graph500 announced at ISC10 (June 2010)
• 1st Graph500 List: 9 machines at SC10 (Nov. 2010)
• 2nd Graph500 List: 29 machines at ISC11 (June 2011)
• 3rd Graph500 List: 51 machines at SC11 (Nov. 2011)
• 4th Graph500 List: 88 entries at ISC 12 (June 2012)
• 5th Graph500 List: 124 entries at SC12 (Nov. 2012)
• 6th Graph500 List: 142 entries at ISC13 (June 2013)
• 7th Graph500 List: 160 entries at SC13 (Nov. 2013)
• 8th Graph500 List: 174 entries at ISC14 (June 2014)
• 9th Graph500 List: 183 entries at SC14 (Nov. 2014) [TODAY!]
Five Business Areas
•Cybersecurity
– 15 Billion Log Entires/Day (for
large enterprises)
– Full Data Scan with End-to-End
Join Required
•Medical Informatics
– 50M patient records, 20-200
records/patient, billions of
individuals
– Entity Resolution Important
•Social Networks
– Example, Facebook, Twitter
– Nearly Unbounded Dataset Size
•Data Enrichment
– Easily PB of data
– Example: Maritime Domain
Awareness
• Hundreds of Millions of
Transponders
• Tens of Thousands of Cargo
Ships
• Tens of Millions of Pieces of Bulk
Cargo
• May involve additional data
(images, etc.)
•Symbolic Networks
– Example, the Human Brain
– 25B Neurons
– 7,000+ Connections/Neuron
www.GRAPH500.org
The Rankings
th
9
Graph 500 List
(followed by special highlights)
# of Entries
124
142
183
174
160
88
9
28
51
1st 2nd 3rd 4th 5th 6th 7th 8th 9th
•Will present the Top3 now, and highlight the Top10
•Ruud van der Pas on Oracle SPARC Graph500
•Torsten Hoefler: “Green” Graph500 measure
9th Graph 500 List
Country
Amsterdam
Australia
Canada
China
France
Germany
Italy
Japan
Luxembourg
Poland
Russia
South Korea
Switzerland
Taiwan
UK
USA
Grand Total
# Entries
2
1
5
8
2
3
2
52
1
1
8
1
6
6
4
81
183
% Entries
1.1%
0.5%
2.7%
4.4%
1.1%
1.6%
1.1%
28.4%
0.5%
0.5%
4.4%
0.5%
3.3%
3.3%
2.2%
44.3%
9th Graph 500: Trends -- TEPS
Slide credit: Scott Beamer
9th Graph 500: Trends -- Cores
Slide credit: Scott Beamer
9th Graph 500: Trends -- Efficiency
Slide credit: Scott Beamer
9th Graph500 List, Top Vendors
Vendor
IBM
Cray
Intel
Convey
Dell
Count
35
22
14
13
13
9th Graph500: Top 10
Rank
1
2
Machine
Vendor
Type
Installation Site
DOE/NNSA/LLNL
Lawrence Livermore Nat.
Sequoia
IBM
BlueGene/Q Lab.
RIKEN AICS
Country Nodes Cores
Mem.
(TB) Scale GTEPS
USA
98304 1572864 1.57 PB
41
23751
Japan
82944 663552 1.33 PB
40
19585
USA
Custom
3
K computer
Fujitsu
DOE/SC/Argonne
Mira
IBM
49152 786432
768
40
14982
4
JUQUEEN
IBM
BlueGene/Q Forschungszentrum Juelich Germany 16384 262144
256
38
5848
5
IBM
BlueGene/Q CINECA
Italy
8192 131072
128
37
2567
6
Fermi
Tianhe-2
(MilkyWay-2)
NUDT
MPP
China
8192 196608
512
36
2061
7
Blue Joule
IBM
BlueGene/Q Daresbury Laboratory
UK
4096
65536
64
36
1427
7
DIRAC
IBM
BlueGene/Q University of Edinburgh
UK
4096
65536
64
36
1427
7
Zumbrota
IBM
BlueGene/Q EDF R&D
France
4096
65536
64
36
1427
7
Avoca
IBM
BlueGene/Q Victorian Life Sciences CI
Australia
4096
65536
64
36
1427
7
Turing
IBM
BlueGene/Q CNRS/IDRIS-GENCI
France
4096
65536
64
36
1427
BlueGene/Q Argonne Nat. Lab.
Changsha
Observations on Graph500
for Nov. 2014
Peter M. Kogge
McCourtney Prof. CSE
Univ. of Notre Dame
Graph 500, SC14
18
Gathering Observations
• Based on same approach as for Top500
• Annotated each report with architecture class
–
–
–
–
Heavyweight (such as Xeon)
Lightweight (such as Blue Gene)
Hybrid (with GPU)
Other (such as XMT or Convey)
• Used machine characteristics to look for
significant trends
– Use all multi-year data here
Graph 500, SC14
19
Still Relatively Flat Performance
1.E+05
1.E+03
1.E+02
1.E+01
1.E+00
1.E-01
Heavyweight
Lightweight
Hybrid
Graph 500, SC14
Other
1/1/15
1/1/14
1/1/13
1/2/12
1/1/11
1.E-02
1/1/10
TEPS (Billions/sec)
1.E+04
Trend
20
Blow-up of Top of Range
1.E+04
Heavyweight
Lightweight
Graph 500, SC14
Hybrid
Other
1/1/15
1/1/14
1/1/13
1/2/12
1/1/11
1.E+03
1/1/10
TEPS (Billions/sec)
1.E+05
Trend
21
Not Much Change from “per Node”
1.E+02
1.E+00
1.E-01
1.E-02
1.E-03
1.E-04
Heavyweight
Lightweight
Hybrid
Graph 500, SC14
Other
1/1/15
1/1/14
1/1/13
1/2/12
1/1/11
1.E-05
1/1/10
GTEPS/Node
1.E+01
Trend
22
Not Much Change in Product Metric
1.E+17
1.E+16
1.E+14
1.E+13
1.E+12
1.E+11
1.E+10
1.E+09
1.E+08
1.E+07
1.E+06
Lightweight
Graph 500, SC14
Hybrid
1/1/15
1/1/14
1/1/13
1/1/11
Heavyweight
1/2/12
1.E+05
1/1/10
GTEPS*# Vertics
1.E+15
Other
23
We Can See Comm Cost
1.E+03
GTEPS/Node
1.E+02
1.E+01
1.E+00
1.E-01
1.E-02
1.E-03
1.E-04
1.E-05
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
Nodes
Heavyweight
Lightweight
Graph 500, SC14
Hybrid
Other
24
Continued Near Perfect Weak Scaling
1.E+05
TEPS (Billions/sec)
1.E+04
1.E+03
Shared Memory
Machines
1.E+02
1.E+01
1.E+00
1.E-01
1.E-02
20
25
30
35
40
45
Vertex Scale (log2(#vertices))
Heavyweight
Lightweight
Graph 500, SC14
Hybrid
Other
Trend
25
As a Function of Nodes
1.E+05
TEPS (Billions/sec)
1.E+04
1.E+03
Single Node
Performance
1.E+02
1.E+01
1.E+00
1.E-01
1.E-02
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
Hybrid
Other
1.E+05
Nodes
Heavyweight
Lightweight
Graph 500, SC14
Trend
26
As a Function of Cores
1.E+05
TEPS (Billions/sec)
1.E+04
1.E+03
Single Node
Shared Memory
Machines
1.E+02
1.E+01
1.E+00
1.E-01
1.E-02
1.E+00
1.E+01
Heavyweight
1.E+02
1.E+04
1.E+03
Lightweight
Cores
Graph 500, SC14
Hybrid
1.E+05
1.E+06
Other
1.E+07
Trend
27
Blue Gene Comparison
GTEPS/Node
1.E+00
1.E-01
1.E-02
1.E-03
1.E-04
1/1/10
1/1/11
1/2/12
1/1/13
1/1/14
1/1/15
Nodes
BGQ
Graph 500, SC14
BGP
28
Highlights of the
th
9
Graph500 List
• K Computer and IBM BlueGene/Q dominate the top of the list
• Several new observations:
•
•
•
•
Sequoia reclaiming #1 is a big drop in efficiency (MTEPS/core) for #1
Biggest decrease in efficiency for #1 (and first time to go down in 3
years)
Is this the first time Graph500 has had a full allocation (1.5M cores) on
Sequoia?
First increase in SCALE for #1 in 2 years
Benchmark is maturing (things staying constant):
•
•
•
•
•
2nd list in a row of modest improvement for top (in contrast to rapid
growth or flatline)
Sum of top 8 machines is still large fraction (>70% of submitted TEPS)
#1 system has about 25% of total TEPS
SCALEs of top 8 submissions relatively the same
Graph 500, SC14
29
ANNOUNCEMENT
•The 10th Graph500 List will be released at
ISC15,
•June 2015
Graph 500, SC14
30

similar documents