predicting performance impact of dvfs for realistic memory systems
DESCRIPTION
Predicting Performance Impact of DVFS for Realistic Memory Systems. Rustam Miftakhutdinov Eiman Ebrahimi Yale N. Patt. Dynamic Voltage/Frequency Scaling. V. f. Image source: intel.com. Impact of Frequency Scaling. time. energy. power. f opt. frequency. Impact of Frequency Scaling. - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/1.jpg)
Predicting Performance Impact of DVFSfor Realistic Memory Systems
Rustam MiftakhutdinovEiman Ebrahimi
Yale N. Patt
![Page 2: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/2.jpg)
2
V
f
Dynamic Voltage/Frequency Scaling
Image source: intel.com
![Page 3: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/3.jpg)
3
fopt
Impact of Frequency Scaling
frequency
time
power
energy
![Page 4: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/4.jpg)
4
fo
Impact of Frequency Scaling
power
time
frequency
![Page 5: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/5.jpg)
5
fopt
Prediction Overview
instructions
frequency
energy perinstruction
100K 200K 300K0
fo freq.
time
fo freq.
powerfo
fo freq.fopt
energy
our work
×
![Page 6: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/6.jpg)
6
Outline
Intro to performance prediction
Why realistic memory systems?
Variable memory latency
Prefetching
✓
![Page 7: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/7.jpg)
7
V
f
Why Realistic Memory System?
![Page 8: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/8.jpg)
8
Prior Work
• Stall time
• Leading loads (2010) S. Eyerman et al. G. Keramidas et al. B. Rountree
Evaluated withconstant access latency memory system
![Page 9: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/9.jpg)
9
Energy Savings
Constant Access Latency
Realistic DRAM Realistic DRAM + Streaming Prefetcher
0123456789 Oracle
Stall timeLeading loadsOur predictor
Norm
. Ene
rgy
Savi
ngs
(%)
< 0.1
Gmean of relative savings for 13 memory-intensive SPEC 2006 benchmarks.Baseline: most energy-efficient static frequency for SPEC 2006
*
![Page 10: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/10.jpg)
10
Energy Savings
Constant Access Latency
Realistic DRAM Realistic DRAM + Streaming Prefetcher
0123456789 Oracle
Stall timeLeading loadsOur predictor
Norm
. Ene
rgy
Savi
ngs
(%)
< 0.1
Gmean of relative savings for 13 memory-intensive SPEC 2006 benchmarks.Baseline: most energy-efficient static frequency for SPEC 2006
*
![Page 11: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/11.jpg)
11
Outline
Intro to performance prediction
Why realistic memory systems?
Variable memory latency
Prefetching
✓✓
![Page 12: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/12.jpg)
12
Execution Example
chipactivity
memoryrequests A
BC
DE
1 2 3 4time
![Page 13: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/13.jpg)
13
T = Tmemory + Tcomputeindependent offrequency
proportional tocycle time
![Page 14: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/14.jpg)
14
to
Linear Modelexecution time T
cycle time t
Tmemory
Tcompute
0
![Page 15: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/15.jpg)
15
Measuring Tmemory
chipactivity
memoryrequests
time
![Page 16: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/16.jpg)
16
Measuring Tmemory
chipactivity
memoryrequests
time
![Page 17: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/17.jpg)
17
Causes of Request Dependences
next
next
next
Pointer Chasing
instruction window
miss miss
Finite Chip Resources
![Page 18: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/18.jpg)
18
Measuring Tmemory
chipactivity
memoryrequests
time
![Page 19: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/19.jpg)
Critical Path Algorithm
at Tstart 1. record Tstart and Tmemory
TendTstart time
Tmemory
19
at Tend 2. compute path = Tmemory(Tstart) + (Tend - Tstart)old critical path request latency
3. set Tmemory = max(Tmemory, path)
new Tmemory
(length of critical path)
![Page 20: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/20.jpg)
20
to
Linear Modelexecution time T
cycle time t
Tmemory
Tcompute
0
![Page 21: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/21.jpg)
21
Linear Model
to
execution time T
cycle time t
Tmemory
Tcompute
0
to cycletime
Tm
time
fo freq.
time
fo freq.
power
fo freq.fopt
energy
×
![Page 22: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/22.jpg)
22
Critical Path: Variable Access Latency
chipactivity
memoryrequests
time
Leading Loads: Constant Access Latency
timechipactivity
memoryrequests
![Page 23: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/23.jpg)
23
to
Leading Loadsexecution time T
cycle time t
Tmemory
Tcompute
0
leading loads
![Page 24: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/24.jpg)
24
Leading Loads
to
execution time T
cycle time t
Tmemory
Tcompute
0
leading loads
to cycletime
Tm
time
fo freq.
time
fo freq.
power
fo freq.fopt
energy
×
![Page 25: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/25.jpg)
25
Energy Savings
Constant Access Latency
Realistic DRAM0
1
2
3
4
5
6
7
8 OracleStall timeLeading loadsOur predictor
Norm
. Ene
rgy
Savi
ngs (
%)
Gmean of relative savings for 13 memory-intensive SPEC 2006 benchmarks.Baseline: most energy-efficient static frequency for SPEC 2006
*
![Page 26: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/26.jpg)
26
Outline
Intro to performance prediction
Why realistic memory systems?
Variable memory latency
Prefetching
✓✓✓
![Page 27: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/27.jpg)
27
chipactivity
memoryrequests
time
Prefetcher OFF
Prefetcher ON
chipactivity
memoryrequests
Streaming Workload
![Page 28: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/28.jpg)
28
Limited Bandwidth Modelexecution time T
cycle time t
Tdemand
TcomputeTmemorymin
tcrossover0
![Page 29: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/29.jpg)
Energy Savings
29Gmean of relative savings for 13 memory-intensive SPEC 2006 benchmarks.Baseline: most energy-efficient static frequency for SPEC 2006
*
Constant Access Latency
Realistic DRAM Realistic DRAM + Streaming Prefetcher
0123456789 Oracle
Stall timeLeading loadsOur predictor
Norm
. Ene
rgy
Savi
ngs
(%)
< 0.1
![Page 30: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/30.jpg)
30
Recap
Intro to performance prediction
Why realistic memory systems?
Variable memory latency
Prefetching
✓✓✓✓
![Page 31: Predicting Performance Impact of DVFS for Realistic Memory Systems](https://reader035.vdocuments.site/reader035/viewer/2022070503/568161d3550346895dd1d05d/html5/thumbnails/31.jpg)
31
Final Thought
Performance predictors need realistic evaluation