Power-performance modeling, analyses and challenges Kirk W. Cameron - - PowerPoint PPT Presentation
11 th Charm++ Workshop: Power-performance modeling, analyses and challenges Kirk W. Cameron Computer Science Virginia Tech This material is based upon work supported by the National Science Foundation under Grant No. 0910784 and 0905187. My
This material is based upon work supported by the National Science Foundation under Grant No. 0910784 and 0905187. 11th Charm++ Workshop:
Source: CAREER: High-performance, Power-aware Computing
From 2007-2012… [6x ↑ Flops/watt] [~2.5x ↑ power consumption] Projections for 2012-2019… [2100 to ~15,000 MFlops/Watt] [66 kW for 1 Petaflop System] [66 MW for 1 Exaflop System} [Need 50,000 Mflops/Watt for 1 Exaflop @ 20 MW by 2019!!!]
Model & Optimize Performance Improve Power-Performance Efficiency Model Effects of Power Profile & Evaluate Power Optimize Effects of Power
[SC04], [SC05], [IPDPS 2005], [IJHPCA 2009], [TPDS 2010]
12
[IEEE Computer 38(11) 2005, TPDS 21(5) 2010, http://scape.cs.vt.edu/software/]
14
[IEEE Computer 38(11) 2005, TPDS 21(5) 2010, http://scape.cs.vt.edu/software/]
16
February 15, 2012 SIAM PP, Savannah, GA 17 / 19
LAPACK MKL PLASM A Power consumption over time Matrix inverse Sources: Piotr Luszczek Hatem Ltaief
February 15, 2012 SIAM PP, Savannah, GA 18 / 19
Bidiagonal Reduction: CPU Power
PLASMA LAPACK
20 40 60 80 100 120 140 160 1 14 27 40 53 66 79 92 105 118 131 144 157 170 183 196 209 222 235 248 261 274 287 300 313 326 339 352 365 378 391 404 417 430 443 456 469 482 495 508 521 534 547 560 573 586 599 612 625 Power (watt) Time (0.02 second) CPU GPU MEM MB
CudaMalloc (Data Movement) convolutionRow convolutionColumn
convolutionTexture_15360_32
21
200 400 600 800 1000 1200 1400 1600 1800
Time Watts
PDU Power Measurements
System +CPU Monitor 200 400 600 800 1000 1200 1400 1600 1800
Time Watts
Granola Enterprise Power Estimates
CPU System Monitor
Granola software gives more detail… …same accuracy as expensive hardware
22
24
Model & Optimize Performance Improve Power-Performance Efficiency Model Effects of Power Profile & Evaluate Power Optimize Effects of Power
[SC 2004], [SC 2005], [IPDPS 2011], [IPDPS 2013]
1 1
N N
– Amdahl’s law for 1 enhancement (parallelism)
Time Degree of Parallelism
So we only get energy savings by reducing time. Right? Then why does PM (e.g. DVFS) save energy? And sometimes without affecting time?
But, overhead is the key to savings energy without loss!
Energy
– Speedup – w: workload – N: number of nodes – f: the clock frequency and f0 is the base value – T1(w, f0): sequential execution time at base frequency f0 – TN(w, f): parallel execution time at N processors at frequency f
1
N N
27
[IPDPS 2007]
– # processors: 256 – CPU frequency: 1200MHz
8 16 32 64 128 256 512 1024 600 1000 1400 0.00 5.00 10.00 15.00 20.00 25.00 30.00 35.00 EDP(104Joulesxseconds) Processors Frequency (MHz)
EDP values for LU
30-35 25-30 20-25 15-20 10-15 5-10 0-5
30
General form of our Iso-energy-efficiency model: : system-wide energy efficiency (baseline): total energy consumption of sequential execution on one processor : the total energy consumption of parallel execution for a given application on p parallel processors : the additional energy overhead required for parallel execution and running extra system components
31
[IPDPS 2011],[IPDPS 2013]
32 Energy efficiency Energy efficiency FT’s system-wide energy efficiency with p and n as variables FT’s system-wide energy efficiency with p and f as variables
while scaling up system size.
Patents: [USPTO: #13/061,565] [UK: #GB2476606B]
33
We need lots of help. Disruptive vs. Incremental. Silver bullet is unlikely. Commodity matters. Markets matter. Tools matter. Wanted: Major catastrophe. Custom system is likely the only answer by 2019. Energy wall? “Victory” is inevitable when you change the game.