Computational Sprinting on a Hardware/Software Testbed Arun Raghavan *, Laurel Emurian *, Lei Shao...
-
Upload
blake-lawson -
Category
Documents
-
view
214 -
download
0
Transcript of Computational Sprinting on a Hardware/Software Testbed Arun Raghavan *, Laurel Emurian *, Lei Shao...
Computational Sprinting on a Hardware/Software Testbed
Arun Raghavan*, Laurel Emurian*, Lei Shao#, Marios Papaefthymiou+, Kevin P. Pipe+#, Thomas F. Wenisch+, Milo M. K. Martin*
University of Pennsylvania, Computer and Information Science*
University of Michigan, Electrical Eng. and Computer Science+
University of Michigan, Mechanical Engineering#
This work licensed under the Creative Commons Attribution-Share Alike 3.0 United States License
• You are free:• to Share — to copy, distribute, display, and perform the work• to Remix — to make derivative works
• Under the following conditions:• Attribution. You must attribute the work in the manner specified by the author or
licensor (but not in any way that suggests that they endorse you or your use of the work). • Share Alike. If you alter, transform, or build upon this work, you may distribute the
resulting work only under the same, similar or a compatible license.
• For any reuse or distribution, you must make clear to others the license terms of this work. The best way to do this is with a link to:
http://creativecommons.org/licenses/by-sa/3.0/us/ • Any of the above conditions can be waived if you get permission from the
copyright holder.• Apart from the remix rights granted under this license, nothing in this
license impairs or restricts the author's moral rights.
2
3
Overview
• Computational sprinting [HPCA’12]• Targets responsiveness in thermally constrained environments• Far exceed sustainable power for short bursts of computation• Simulation based feasibility study
• This work: what can we learn with today’s hardware?• Engineer hardware/software testbed for sprinting
• Reduce heat venting capacity• Sustain only lowest power mode
• Can sprint on today’s system• Longer with phase-change material• Sprinting improves energy-efficiency
• Even for sustained computations
5
Tmaxp
ow
erte
mp
erat
ure
Effect of thermal capacitance
Computational Sprinting Using Dark Silicon [HPCA’12]
6
Tmaxp
ow
erte
mp
erat
ure
Effect of thermal capacitance
Computational Sprinting Using Dark Silicon [HPCA’12]
7
Tmaxp
ow
erte
mp
erat
ure
Effect of thermal capacitance
Computational Sprinting Using Dark Silicon [HPCA’12]
8
Tmaxp
ow
erte
mp
erat
ure
State of the art:Turbo Boost 2.0
exceeds sustainable power
with DVFS (~25%)
Our goal: 10x
Effect of thermal capacitance
Computational Sprinting Using Dark Silicon [HPCA’12]
9
Evaluating Sprinting
• Simulation-based feasibility study [HPCA’12]
• Thermal models: buffer heat using thermal capacitance• Electrical models: stabilize voltage with gradual core activation• Architectural models:
• Large responsiveness improvements • Little dynamic energy overheads
• Next steps: understanding sprinting on a real system• Build a real chip? • Sprint on today’s mobile chips?
Our approach: study sprinting on hardware available today
10
This Work: Testbed for Computational Sprinting
• How long can the testbed sprint?• How to select sprint intensity?• How can we extend sprint duration?• How does sprinting impact energy?
sprinting
sustainable
Quad-core Intel i7-2600With heatsink and fan: 95W
12
Remove heatsink, slow fan;10W thermal design (TDP)
Cores Freq. PowerNormalized
Power
PeakSpeedu
p
1 core 1.6 GHz 10 W 1x 1x
4 cores 1.6 GHz 20 W ~2x 4x
4 cores 3.2 GHz 50 W ~5x 8x
3 operating modes:
Sprinting Performance
0
1
2
3
4
5
6
7
8
norm
aliz
ed s
peed
up
sobel disparity segment kmeans feature texture
Cores + Frequency (3.2GHz): 6.3x speedup Cores only (1.6 GHz): 3.5x speedup
Max 4 core, 3.2 GHz
Max 4 core, 1.6 GHz
Baseline(no sprint)
3.2GHz
3.2GHz
3.2GHz
3.2GHz
3.2GHz
3.2GHz
1.6GHz
1.6GHz
1.6GHz
1.6GHz
1.6GHz
1.6GHz
13
Testbed Thermal Response40
5060
700
2040
60
Po
wer
(W
)
time (s)
Tem
p (
°C)
time (s)
Tmax
15
-5 0 5 10 15 20 25 300
1 0
2 0
3 0
4 0
5 0
6 0
sustainedsprint-3.2GHz
-5 0 5 10 15 20 25 30
sustainedsprint-3.2GHz
020
4060
4050
6070
sustained
sustained
-5 0 5 10 15 20 25 300
1 0
2 0
3 0
4 0
5 0
6 0
sustainedsprint-3.2GHz
-5 0 5 10 15 20 25 30
sustainedsprint-3.2GHz
Testbed Thermal Response40
5060
700
2040
60
Po
wer
(W
)
time (s)
Tem
p (
°C)
time (s)16
5x
3s
-5 0 5 10 15 20 25 300
1 0
2 0
3 0
4 0
5 0
6 0
sustainedsprint-3.2GHz
-5 0 5 10 15 20 25 30
sustainedsprint-3.2GHz
020
4060
4050
6070
sustainedsprint (3.2 GHz)
sustainedsprint (3.2 GHz)
20g copper
Δ 25oC, ~188J 3.5s @ 50W
Heat spreader
-5 0 5 10 15 20 25 30
sustainedsprint-3.2GHzsprint-1.6GHz
-5 0 5 10 15 20 25 300
1 0
2 0
3 0
4 0
5 0
6 0
sustainedsprint-3.2GHzsprint-1.6GHz
Testbed Thermal Response40
5060
700
2040
60
Po
wer
(W
)
time (s)
Tem
p (
°C)
time (s)17
5x
3s
-5 0 5 10 15 20 25 30
sustainedsprint-3.2GHz
020
4060
4050
6070
-5 0 5 10 15 20 25 300
1 0
2 0
3 0
4 0
5 0
6 0
sustainedsprint-3.2GHz
21ssustainedsprint (3.2 GHz)sprint (1.6 GHz)
sustainedsprint (3.2 GHz)sprint (1.6 GHz)
2x
20g copper
Δ 25oC, ~188J 3.5s @ 50W
Heat spreader
Truncated Sprint Performance
19
012345678
4 cores, 1.6 GHz4 cores, 3.2 GHz
computation length
nor
ma
lize
d s
pe
edu
p
Little benefit
Near-peak performance for short tasks
Truncated Sprint Performance
012345678
4 cores, 1.6 GHz4 cores, 3.2 GHz
computation length
nor
ma
lize
d s
pe
edu
p
20
Little benefit
Near-peak performance for short tasks
Lower peak performance; benefits longer tasks
• Best sprint intensity depends on task size• How to sprint when task size is unknown?• Sprint pacing
• Max intensity sprint for half thermal capacitance• Cores-only sprinting for other half
Truncated Sprint Performance
012345678
4 cores, 1.6 GHz4 cores, 3.2 GHz
computation length
nor
ma
lize
d s
pe
edu
p
21
• Best sprint intensity depends on task size• How to sprint when task size is unknown?• Sprint pacing
• Max intensity sprint for half thermal capacitance• Cores-only sprinting for other half
Truncated Sprint Performance
012345678
adaptive4 cores, 1.6 GHz4 cores, 3.2 GHz
22
• Best sprint intensity depends on task size• How to sprint when task size is unknown?• Sprint pacing
• Max intensity sprint for half thermal capacitance• Cores-only sprinting for other half
computation length
nor
ma
lize
d s
pe
edu
p
24
Two Ways of Adding Thermal Capacitance
• Specific heat capacity: introduce thermal mass
• Latent heat: absorb heat to change phase (e.g. melting)
tem
per
atu
rete
mp
era
ture
time (s)
time (s)
Phase-change absorbs heat while holding temperature constant
Baselinesprinting
More specific heat takes longer to heat
20g copper, Δ 25oC ~188J
1g of wax ~200J
25 Computational Sprinting on a Hardware-Software Testbed
Extending Sprints with Phase Change Material
4g of wax, 1g of aluminum foam
Copper shim
26
0 50 100 150 200 250 300 350 40040
50
60
70
80
Impact of Phase Changete
mp
erat
ure
(°C
)air
time (s)
27
0 50 100 150 200 250 300 350 40040
50
60
70
80
tem
per
atu
re (
°C)
Small extension from heat capacity of encasement…
foamair
time (s)
Impact of Phase Change
28
0 50 100 150 200 250 300 350 40040
50
60
70
80
tem
per
atu
re (
°C)
Small extension from heat capacity of encasement……6x increase in sprint duration with phase change
foam waxair
time (s)
Impact of Phase Change
29
0 50 100 150 200 250 300 350 40040
50
60
70
80
tem
per
atu
re (
°C)
Small extension from heat capacity of encasement……6x increase in sprint duration with phase change
due to phase change
foamairwater
time (s)
Impact of Phase Changewax
32
Energy Impact of Sprinting
0
0.5
1
1.5
norm
aliz
ed e
nerg
y
sobel disparity segment kmeans feature texture3.2GHz
3.2GHz
3.2GHz
3.2GHz
3.2GHz
3.2GHz
1.6GHz
1.6GHz
1.6GHz
1.6GHz
1.6GHz
1.6GHz
Race-to-idle: 7% energy savings!
Sprint 3.2GHz
Sprint 1.6GHz
Idle
0 50 100 150 200 250 300 350 400 450 50005
10152025
sustainedsprint-rest
Sprint-and-Rest
seconds
pow
er (
W)
0 50 100 150 200 250 300 350 400 450 500304050607080
sustainedsprint-rest
seconds
tem
pera
ture
(o C
)
0 50 100 150 200 250 300 350 400 450 5000
50100150200250300
sustainedsprint-rest
seconds
cum
ulat
ive
wor
k
sprint-and-rest
sprint-and-rest
sprint-and-rest
34
35% faster
5s @ 20W sprint, 12s @ 5W rest < 10W average
Conclusions
• Testbed confirms sprinting improves responsiveness
• Sprint pacing can extend benefits of sprinting
• Exploiting phase change allows longer sprints
• Sprinting can save energy in thermally limited systems