HPC Midlands - Loughborough University and Rolls Royce HPC Collaboration
Energy and Power Consumption Trends in HPC - …luszczek/conf/ppam2013_energy_power...Piotr Luszczek...
Transcript of Energy and Power Consumption Trends in HPC - …luszczek/conf/ppam2013_energy_power...Piotr Luszczek...
Piotr Luszczek
Energy and Power Consumption Trends in HPC
Fun Things to Do with (not so) Big Data
Innovative Computing Laboratory University of Tennessee
PPAM 2013, Warsaw, Poland 1950 1952 1960 1964 1970 1971 1976 1982 1986 1988 1992 1993 1997 2000 2008 2009 2010 2011
1.00E+000
1.00E+002
1.00E+004
1.00E+006
1.00E+008
1.00E+010
1.00E+012
1.00E+014
1.00E+016
EDSAC 1UNIVAC 1
IBM 7090CDC6600
CDC7600
IBM 360/195
Cray 1
Cray X-MPCray 2
TMC CM-2
TMC CM-5Cray T3D
ASCI Red
K Computer
Tianhe-1ACray XT5 Jaguar
IBM Cell Roadrunner
ASCII White
5 Decades of the LINPACK: the LINPACK Law
PPAM 2013, Warsaw, Poland 1950 1952 1960 1964 1970 1971 1976 1982 1986 1988 1992 1993 1997 2000 2008 2009 2010 2011
1.00E+000
1.00E+002
1.00E+004
1.00E+006
1.00E+008
1.00E+010
1.00E+012
1.00E+014
1.00E+016
EDSAC 1UNIVAC 1
IBM 7090CDC6600
CDC7600
IBM 360/195
Cray 1
Cray X-MPCray 2
TMC CM-2
TMC CM-5Cray T3D
ASCI Red
K Computer
Tianhe-1ACray XT5 Jaguar
IBM Cell Roadrunner
ASCII White
5 Decades of the LINPACK: the LINPACK Law
19901991
19931995
19981999
20012003
20052009
20101
10
100
1000
10000
100000
1000000
10000000
PPAM 2013, Warsaw, Poland 4
Exaflop by 2020?
● Horst Simon's wager– A €2000 bet that it won't happen.
PPAM 2013, Warsaw, Poland 5
Number One Obstacle to Exaflop?
● Time to run
PPAM 2013, Warsaw, Poland 6
Projections from 2009
PPAM 2013, Warsaw, Poland 7
Let's Plugin the Numbers
● Memory: 64 PB– Problem size: ~90 million unknowns
● Performance:– 1 Eflop/s
● Time to run:– 5.5 days!
N = √64 PB/8 bytes
time ≈
23N 3
rmax
PPAM 2013, Warsaw, Poland 8
How Long a TOP500 Run Takes
PPAM 2013, Warsaw, Poland 9
Projecting Execution Time to Exascale
PPAM 2013, Warsaw, Poland 10
Number Two Obstacle?
● Power– Total power consumption– Gains in power efficiency
PPAM 2013, Warsaw, Poland 11
Power Consumption
● Power consumption is an issue throughout the chip industry● Supercomputing is hit hard from all sides
– CPU/GPU/Accelerator/Coprocessor– Memory– Interconnect
● Power consumption projections for the exascale system continue evolving
PPAM 2013, Warsaw, Poland 12
Power Consumption in Perspective
K computer Exa-scaleGoogle Apple
Facebook NSA
20 MW
12 MW
+65 MW
50 MW…
100 MW
65 MW…
200 MW
Microsoft
17 MW
Tianhe-2
PPAM 2013, Warsaw, Poland 13
Power Trends Towards Exascale
Hoover Dam
Siekierki
Żerań
PPAM 2013, Warsaw, Poland 14
Another Way to Look at Power and Performance
PPAM 2013, Warsaw, Poland 15
How to Measure the Cost of the System
● Initial system cost has always been of questionable value– Often, part of a deal with the vendor and the price cannot be
disclosed● Large supercomputer installations create good “bragging rights”
opportunity– Vendor might ship early and/or development parts which are
discounted● Getting the system up and running for a press release
– The early systems are “ASAP quality”● Life time costs have become prominent over the years
– Energy is the biggest component
PPAM 2013, Warsaw, Poland 16
Energy Trends