Context
Passive gathering of information
Integrated behavioral dvfs method
Energy efficient low level dvfs for HPC
applications
Georges Da Costa, Jean-Marc Pierson
IRIT, Toulouse University
Green Days @ Paris
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Context
Passive gathering of information
Plan
1 Context
2 Passive gathering of information
3 Integrated behavioral dvfs method
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
Integrated behavioral dvfs method
Context
To optimize a computing center:
Gather insight on running applications
Choose how to act
depends on application
More precise : phase of application
Act (change frequency, switch on/off parts of nodes,...)
Optimize : reduce energy consumption at the same performance
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Context
Passive gathering of information
Integrated behavioral dvfs method
Ignorance is bliss, really?
*-AAS (PAAS, IAAS,...) leads to ignorance
Ignorance leads to errors
Errors lead to inefficiency
Focus :
How to optimize a computing center while knowing nothing?
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Context
Passive gathering of information
Know your enemy
What we know
HPC applications
Goal: Save energy
No impact on performance (SLA,...)
Name your weapon (constraints)
Minimum impact of monitoring
Closed application, no source
Even full OS freedom (Grid’5000, VMs)
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
Is it so important?
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
Plan
1 Context
2 Passive gathering of information
3 Integrated behavioral dvfs method
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
Integrated behavioral dvfs method
Who’s Who
Which application is running?
Ask the developer, but
Depends on library
Can cheat (if accounting is related to it)
Computers work for us, not the opposite
Application is not important, its behavior is!
Two different apps can have the same impact
System changes can have the same impact
We need Run-Time Behavioral Detection
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Context
Passive gathering of information
Integrated behavioral dvfs method
BigBrother is watching
Run-time detection
Behavioral pattern
Extract information
Fine grained : performance counters, system values
Coarse grained : network, disk, energy consumption
Classical remark: impact of the monitoring infrastructure
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Context
Passive gathering of information
Integrated behavioral dvfs method
Finding patterns, NPB example, CG
1.2e+08
Number of bytes sent per second
1e+08
8e+07
6e+07
4e+07
2e+07
0
0
10
20
30
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
40
time (s)
50
60
70
80
Context
Passive gathering of information
Integrated behavioral dvfs method
Finding patterns, NPB example, FT
1.2e+08
Number of bytes sent per second
1e+08
8e+07
6e+07
4e+07
2e+07
0
0
20
40
60
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
80
100
time (s)
120
140
160
180
Context
Passive gathering of information
Integrated behavioral dvfs method
Finding patterns, NPB example, SP
1.2e+08
Number of bytes sent per second
1e+08
8e+07
6e+07
4e+07
2e+07
0
0
50
100
150
200
time (s)
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
250
300
350
Context
Passive gathering of information
Approach
Once we have a model
Step 1: Measure some characteristics
Step 2: Compare to reference
Step 3: Categorize application (or phase)
Usecase : Nas Parallel Benchmark
Different type of workload
Representative of HPC applications
Seven benchmarks
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
Integrated behavioral dvfs method
Median sent bytes
6e+07
5e+07
4e+07
2e+07
Number of sent bytes (med value)
7e+07
6e+07
FT
SP
BT
EP
LU
IS
CG
8e+07
4e+07
9e+07
8e+07
Network and energy (left measured, right computed by R)
3e+07
0e+00
2e+07
1e+07
240
0
235
240
245
250
255
Power consumption (W)
260
265
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
270
245
250
255
Mean power (W)
260
●
● ●●
●
● ●●
265
Context
Passive gathering of information
Integrated behavioral dvfs method
Energy efficient dvfs
Categorize application (phase)
Apply a rule-based algorithm
Change processor speed to min or max
Can take into account several objectives depending on rules
Energy only
Energy with taking into account performance
...
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Context
Passive gathering of information
Plan
1 Context
2 Passive gathering of information
3 Integrated behavioral dvfs method
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
With more control comes more efficiency
Governors
Needs control over OS
Take low level information
Take decision on frequency
Ongoing fellowship
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
Preliminary results
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications
Integrated behavioral dvfs method
Context
Passive gathering of information
Integrated behavioral dvfs method
Conclusion
Application characterization is possible with no impact!
Using it we can optimize resources usages and reduce energy
(J)
Also possible to use at the kernel level.
Georges Da Costa, Jean-Marc Pierson IRIT, Toulouse University
Energy efficient low level dvfs for HPC applications