readex runtime exploitation of application dynamism for ... · dta management dta process...
TRANSCRIPT
![Page 1: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/1.jpg)
READEX – Runtime Exploitation of Application Dynamism for Energy-efficient eXascale computingEnA-HPC @ ISC’17
Robert Schöne – TUD
![Page 2: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/2.jpg)
Energy Efficiency Tuning Types
1. Static or dynamic tuning?• Uniform or changing behavior of programs
• Dynamic: sampling or instrumentation
2. Reducing power or runtime?• Power:
• Frequencies
• C-States
• Speculative execution (e.g., prefetchers)
• Runtime:• Frequencies (Turbo, various resources share single power budget)
• Select optimal code paths
• Optimize code
3. Tackling regions or synchronization?
EnA-HPC - 22.06.2017 2
![Page 3: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/3.jpg)
Energy Efficiency Tuning Types
EnA-HPC - 22.06.2017 3
![Page 4: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/4.jpg)
Energy Efficiency Tuning with READEX
EnA-HPC - 22.06.2017 4
Balancing based tuning
Region based tuning
Power based tuning Runtime based tuning
READEX
![Page 5: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/5.jpg)
Project Motivation
Applications exhibit dynamic behaviour
• Changing resource requirements
• Computational characteristics
• Changing load on processors over time
EnA-HPC - 22.06.2017 5
![Page 6: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/6.jpg)
Overview
READEX creates a tools-aided methodology for automatic tuning of parallel applications
• Dynamically adjust system parameters to actual resource requirements
Join technologies from embedded systems and HPC
• HPC: PTF, Score-P, and HDEEM
• ES: System scenario methodology
EnA-HPC - 22.06.2017 6
![Page 7: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/7.jpg)
Overview
Co-design approach
• Manual tuning for energy efficiency as a baseline
• Automatic tuning for comparison
• Applications• PERMON and ESPRESO (FETI tools from IT4Innovations)
• Indeed (GNS)
• CORAL benchmark suite
• ProxyApps
EnA-HPC - 22.06.2017 7
![Page 8: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/8.jpg)
Project Partners
• Grant agreement No 671657
• Officially started September 1st, 2015
• Technische Universität Dresden/ZIH (Coordinator)
• Norwegian University of Science and Technology
• Technische Universität München
• IT4Innovations, VSB-Technical University of Ostrava
• NUI Galway, Irish Centre for High-End Computing
• Intel France
• Gesellschaft für numerische Simulation mbH
EnA-HPC - 22.06.2017 8
![Page 9: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/9.jpg)
Terminology: Region and Region Instance
External Advisory Board Online Meeting, December 12th 2016
Phase
Phase region
Significant regionRuntime situation
9
![Page 10: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/10.jpg)
Terminology: Tuning Parameter and Intra-Phase Dynamism
External Advisory Board Online Meeting, December 12th 2016
Tuning Parameter FREQ=2 GHz
Tuning Parameter FREQ=1.5 GHz
10
![Page 11: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/11.jpg)
Workflow
1. Instrument applicationScore-P provides different kinds of instrumentation
2. Detect dynamismCheck whether runtime situations could benefit from tuning
3. Detect energy saving potential and configurations (DTA)
Use tuning plugin and power measure-ment infrastructure to search for opti-mal configurationCreate tuning model
4. Runtime application tuning (RAT)Apply tuning model, use optimal configuration
EnA-HPC - 22.06.2017 11
![Page 12: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/12.jpg)
Instrumentation via Score-P
EnA-HPC - 22.06.2017 12
![Page 13: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/13.jpg)
Instrumentation via Score-P
• HPC performance measurement infrastructure• Creates CUBEx profiles or OTF2 traces
• Instrumentation and sampling
• Supports most HPC programming paradigms
• Mechanism for online usage of data – Periscope
• Efficient implementation
• Power measurement plugins (see talk by T. Ilsche)
• Re-use and extend existing infrastructure• Parse CUBEx profiles to find significant regions
• Score-P Substrate Plugins
• Use READEX Runtime Library (RRL) to change parameters
• Support tools to lower measurement overhead via filtering
EnA-HPC - 22.06.2017 13
![Page 14: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/14.jpg)
Toggling Parameters
EnA-HPC - 22.06.2017 14
![Page 15: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/15.jpg)
Toggling Parameters
EnA-HPC - 22.06.2017 15
• Hardware parameters• Core frequency, uncore frequency
• Clock modulation, Energy Performance Bias, prefetchers
• Runtime parameters• Message Passing Interface, e.g., message size threshold
• OpenMP parameters, e.g., loop scheduler, number of threads
• Application tuning parameters
![Page 16: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/16.jpg)
Application Tuning Parameters (Work in Progress)
16EnA-HPC - 22.06.2017
// C example
// register parameters at READEX
ATP_PARAM_DECLARE("PARAMETER1", ATP_PARAM_TYPE_RANGE, 1, "Domain1");
// declare set of possible values for the parameter
ATP_PARAM_ADD_VALUES("PARAMETER1", values_array, num_values, "Domain1")
// getting parameter setting from READEX, store in variable app_param
ATP_PARAM_GET("PARAMETER1",&app_param,"Domain1")
// … usage of app_param, e.g., switch (app_param) {
![Page 17: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/17.jpg)
Application Tuning Parameters (Work in Progress)
Application parameters example: different preconditioners in ESPRESO solver
• Full Dirichlet preconditioner is usually the preferred one (the best numerical properties)
• Depends on input dataset / problem that is solved
• All preconditioners have been evaluated with the optimal hardware parameter settings
17
Preconditioner typeNumber of iterations
Single iteration costTime and energy
Total solution costTime and energy
No preconditioner 172 130 + 0 ms 32.3 + 0.00 J 21.4 s 5.50 kJ
Weight function 100 130 + 2 ms 32.3 + 0.53 J 12.9 s 3.28 kJ
Lumped 45 130 + 10 ms 32.3 + 3.86 J 6.3 s 1.64 kJ
Light Dirichlet 39 130 + 10 ms 32.3 + 3.74 J 5.5 s 1.41 kJ
Full Dirichlet (default) 30 130 + 80 ms 32.3 + 20.6 J 6.3 s 1.59 kJ
Note: 130 ms and 32.3 J – is a baseline for single iteration cost without preconditioner
Nu
mer
ical
eff
icie
ncy
Co
mp
ute
de
man
d
11.3% energy savings against the default full Dirichlet preconditioners
EnA-HPC - 22.06.2017
![Page 18: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/18.jpg)
Design Time Analysis
EnA-HPC - 22.06.2017 18
![Page 19: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/19.jpg)
Design Time Analysis
• Periscope Tuning Framework
• Pre-computation of dynamicity and significant regions
• Different objectives (e.g., runtime, energy, EDP)
• Different search strategies (complete, random, genetic)
• Uses RRL to switch parameters
• Determine best configuration for significant regions
• Current work in progress:
• Cluster regions in scenarios
• Application tuning parameters
EnA-HPC - 22.06.2017 19
![Page 20: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/20.jpg)
Tuning Model Visualization
Force graph for scenarios
EnA-HPC - 22.06.2017 20
Vampir visualization of parameter changes
![Page 21: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/21.jpg)
Runtime Tuning
EnA-HPC - 22.06.2017 21
![Page 22: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/22.jpg)
Runtime Tuning
• READEX Runtime Library
• Reads and applies Tuning Model
• Sets and resets configuration at runtime
• Current work in progress:
• Application tuning parameters
• Online calibration mechanism
• Advanced switching decision making
EnA-HPC - 22.06.2017 22
![Page 23: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/23.jpg)
Tuning Potential
23
orange regions are called just once per iteration and therefore are used only for
intra-phase dynamism evaluation
green regions denotes
iterative solver (conjugate
gradient (CG)) and provides
opportunity for inter-phase dynamism
regions with names highlighted in bold are
called only if Hybrid Total FETI method is
used
white regions are ignored because there are other significant regions nested in them
…
ESPRESO: 12.3% + 9.1% = 20.3%
• Structural mechanics code
• Finite element + sparse FETI solver
static dynamic total
Review Meeting, 23.05.2017, Brussels – WP5
![Page 24: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/24.jpg)
Discussion
EnA-HPC - 22.06.2017 24
![Page 25: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/25.jpg)
Backup
EnA-HPC - 22.06.2017 25
![Page 26: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/26.jpg)
Tuning ParametersREADEX READEX EAB Meeting
Robert Schöne – TUD
![Page 27: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/27.jpg)
WP1 Implementation
Review Meeting, 23.05.2017, Brussels - WP1 27
![Page 28: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/28.jpg)
Hardware Parameters
TUD, INTEL (M01 – M12)
Existing Hardware parameters
• Dynamic Voltage and Frequency Scaling (DVFS)
• Uncore Frequency
• Energy Performance Bias (EPB)
Discarded parameters
• Dynamic Duty Cycle Management (DDCM, T-states)
28
DRAM
Review Meeting, 23.05.2017, Brussels - WP1
![Page 29: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/29.jpg)
Application Parameters
29Review Meeting, 23.05.2017, Brussels - WP1
// C example
// register parameters at READEX
ATP_PARAM_DECLARE("PARAMETER1", ATP_PARAM_TYPE_RANGE, 1, "Domain1");
// declare set of values for the parameter
ATP_PARAM_ADD_VALUES("PARAMETER1", values_array, num_values, "Domain1")
// getting parameter setting from READEX, store in variable app_param
ATP_PARAM_GET("PARAMETER1",app_param,"Domain1")
![Page 30: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/30.jpg)
System-software Parameters Investigation
TUD, IT4I, INTEL (M01-12)
• Message Passing Interface • Short message size threshold
• SMP-awareness
• Relevant for MPI_AlltoAll and MPI_Reduce
• OpenMP Threading• Dynamic Concurrency Throttling (change number of threads)
• Workload scheduling algorithm
• Chunksize
30Review Meeting, 23.05.2017, Brussels - WP1
Sch
edu
lers
![Page 31: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/31.jpg)
Design Time AnalysisREADEX READEX EAB Meeting
Robert Schöne – TUD
![Page 32: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/32.jpg)
WP2 Implementation
32Review Meeting, 23.05.2017, Brussels – WP2
![Page 33: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/33.jpg)
Terminology: Region and Region Instance
Phase
Phase region
Significant regionRuntime situation
33
FREQ=2 GHz
FREQ=1.5 GHz
Scenario
Review Meeting, 23.05.2017, Brussels – WP2
![Page 34: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/34.jpg)
Design Time Analysis
34
Instrumentation
Dynamism Detection
Design Time Analysis
READEXConfiguration File
Tuning Model
Task 2.2
Task 2.3
Task 2.1
Task 2.4
scorep_autofilter
readex_dyn_detect
PTF and RRL
TM visualizationVampir
Review Meeting, 23.05.2017, Brussels – WP2
![Page 35: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/35.jpg)
Periscope Tuning Framework
Automatic application analysis & tuning• Tune performance and energy (statically)
• Plug-in-based architecture
• Evaluate alternatives online
• Scalable and distributed framework
Support variety of parallel paradigms• MPI, OpenMP, OpenCL, Parallel pattern
Developed in the AutoTune EU-FP7 project
35Review Meeting, 23.05.2017, Brussels – WP2
![Page 36: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/36.jpg)
Readex_dyn_detect
36
readex-dyn-detect
Significant Region Table
Configuration Serializer
Cu
be
Inte
rface
PTF
Configuration Deserializer
SignificantRegion List
Application
Score-P
profile.cubex
readex-config.xml
GranularityAnalysis
DynamismAnalysis
Review Meeting, 23.05.2017, Brussels – WP2
![Page 37: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/37.jpg)
rts detection
Periscope Tuning Framework
Analysis
Plugin Control
PerformanceDatabase
SearchAlgorithms Experiments
Engine
READEXTuning Plugin
DTA Management
DTA ProcessManagement
RTSManagement
RTS
Dat
abas
eScenario
Identification
ApplicationTuning Model
Score-P READEX Runtime Library
OnlineAccess
Interface
Substrate Plugin
Interface
Instrumen-tation
Metric Plugin
Interface
Energy Measurements
(HDEEM)
37Review Meeting, 23.05.2017, Brussels – WP2
![Page 38: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/38.jpg)
Pre-Computation of Configurations
Periscope Tuning Framework
Analysis
Plugin Control
PerformanceDatabase
SearchAlgorithms Experiments
Engine
READEXTuning Plugin
DTA Management
DTA ProcessManagement
RTSManagement
RTS
Dat
abas
eScenario
Identification
ApplicationTuning Model
Score-P READEX Runtime Library
OnlineAccess
Interface
Substrate Plugin
Interface
Instrumen-tation
Metric Plugin
Interface
Energy Measurements
(HDEEM)
38Review Meeting, 23.05.2017, Brussels – WP2
![Page 39: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/39.jpg)
Runtime TuningREADEX READEX EAB Meeting
Robert Schöne – TUD
![Page 40: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/40.jpg)
WP3 Implementation
40
![Page 41: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/41.jpg)
Task 3.1: Scalable Runtime Library Architecture … (2)
41
![Page 42: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/42.jpg)
Task 3.2: Runtime Scenario Detection … (2)
• During DTA
• Tuning request from PTF as JSON string
• Parsed by OA Event Receiver
• Stored by TMM
42
![Page 43: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/43.jpg)
Task 3.2: Runtime Scenario Detection … (3)
• During RAT
• Identifier value detection
• Scenario classification
43
![Page 44: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/44.jpg)
Task 3.4: Efficient Switching Decision making … (2)
Switching decision component
Manipulation of tuning parameters
44
![Page 45: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/45.jpg)
Task 3.3: Runtime Calibration Mechanism … (2)
Identifier value detection
Unseen rts
Calibrate
Configure
Update ATM
45
![Page 46: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/46.jpg)
IntegrationREADEX READEX EAB Meeting
Robert Schöne – TUD
![Page 47: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/47.jpg)
Initial ResultsREADEX READEX EAB Meeting
Robert Schöne – TUD
![Page 48: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/48.jpg)
WP5 Implementation
Review Meeting, 23.05.2017, Brussels – WP5 48
![Page 49: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/49.jpg)
WP5 Objectives
49Review Meeting, 23.05.2017, Brussels – WP5
Focus of WP5 till M18
![Page 50: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/50.jpg)
Task 5.1: Evaluating dynamism in HPC applications
Tools for Manual Evaluation
• MERIC tool • Based on manual annotation of significant regions• Search the space of tuning parameters to find optimal settings for each significant region• Support tool for energy measurements from HDEEM and RAPL
• READEX Application Dynamism Analysis Report (RADAR) & RADAR generator• Evaluates and reports the dynamic behaviour of the application
The dynamism observed in applications can be caused by the following factors:
• Floating point computations – variation in computational intensity (example - next slide)
• Memory read/write access patterns - variation in the sparsity of matrices in sparse linear algebra
• Inter-process communication patterns
• I/O operations performed during the application's execution
• Different inputs to regions in the application
50Review Meeting, 23.05.2017, Brussels – WP5
![Page 51: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/51.jpg)
Task 5.1: Evaluating dynamism in HPC applications
Goal: Investigate techniques to detect and evaluate dynamic behaviour in HPC applications
Hardware tuning parameters:
• core frequency (DVFS), uncore frequency (UFS)
• number of OpenMP threads
Example: Effect of Computational Intensity (CI)
51
Low CI (DGEMV) High CI10 threads 2.2 GHz UCF 1.2 GHz CF
12 threads 1.2 GHz UCF2.5 GHz CF
Static tuning for both kernels12 threads2.2 GHz UCF 2.4 GHz CF
Co
mp
ute
no
de
ener
gy c
on
sum
pti
on
[J]
CPU core frequency [GHz] CPU core frequency [GHz] CPU core frequency [GHz]
Co
mp
ute
no
de
ener
gy c
on
sum
pti
on
[J]
Co
mp
ute
no
de
en
erg
y co
nsu
mp
tio
n [
J]
Two kernels with1:1 workload ratio
Energy consumption
Energy savings
Default settings 2017J - -
Static optimal 1833J 179J 9%
Dynamic optimal 1612J 221J 12%
Total savings - 400J 20%
Note: runtime of both kernels was equal for default settings
Review Meeting, 23.05.2017, Brussels – WP5
![Page 52: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/52.jpg)
Task 5.1: Evaluating dynamism in HPC applications
52Review Meeting, 23.05.2017, Brussels – WP5
Investigation of application parameters
![Page 53: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/53.jpg)
Task 5.2 Manually exploiting application dynamism
Approach and methodology for manual dynamism evaluation
1. Identify significant regions as the most time consuming (profiler – Alinea MAP)
2. Manually annotate the significant regions in the code – no compiler instrumentation overhead
3. Apply tools developed in T5.1 to detect potential savings of an application
4. Using MERIC run application for all possible configurations of tuning parameters (parameters are set statically before each execution) and measure energy consumption
5. Using RADAR find the best configuration for entire application – static tuning potential
6. Using RADAR find the best configuration for each significant region – dynamic tuning potential
7. Combine both static and dynamic savings to define the potential for total energy savings
53Review Meeting, 23.05.2017, Brussels – WP5
![Page 54: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/54.jpg)
Task 5.2 Manually exploiting application dynamism
54
OpenFOAM: 15.9% + 1.8% = 17.4% energy savings
• Computational fluid dynamics
• Finite volume + multigrid solver
static dynamic total
Review Meeting, 23.05.2017, Brussels – WP5
![Page 55: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/55.jpg)
Task 5.2 Manually exploiting application dynamism
55
orange regions are called just once per iteration and therefore are used only for
intra-phase dynamism evaluation
green regions denotes
iterative solver (conjugate
gradient (CG)) and provides
opportunity for inter-phase dynamism
regions with names highlighted in bold are
called only if Hybrid Total FETI method is
used
white regions are ignored because there are other significant regions nested in them
…
ESPRESO: 12.3% + 9.1% = 20.3%
• Structural mechanics code
• Finite element + sparse FETI solver
static dynamic total
Review Meeting, 23.05.2017, Brussels – WP5
![Page 56: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/56.jpg)
Task 5.2 Manually exploiting application dynamism
Indeed: 17.6% + 3.9% = 20.9%
• Structural mechanics code (Simulation of sheet metal forming)
• Finite Element solver
56
static dynamic total
Review Meeting, 23.05.2017, Brussels – WP5
![Page 57: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/57.jpg)
Task 5.2 Manually exploiting application dynamism
57
Indeed: Specific look at solver region – intra-phase dynamism
Review Meeting, 23.05.2017, Brussels – WP5
Total sum of values from dynamic savings from all phases
Energy consumption: from 50.7kJ to 47.6kJ
6.3% energy savings
![Page 58: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/58.jpg)
M18 - Deliverable 5.1: Report on application dynamism
Evaluation of HPC codes ranging from basic kernels to very complex applications
Key results
• Highly optimized applications tend to provide higher static and lower dynamic savings
• Complex applications, such as ESPRESO, which contains variation on workload (not only compute) shows opportunity for dynamic tuning
Review Meeting, 23.05.2017, Brussels – WP5 58
![Page 59: READEX Runtime Exploitation of Application Dynamism for ... · DTA Management DTA Process Management RTS Management S e Scenario Identification Application Tuning Model Score -P READEX](https://reader035.vdocument.in/reader035/viewer/2022071117/60034ec931d0d66ed11bea16/html5/thumbnails/59.jpg)
WP5: New applications
BEM4I - Boundary element method: 12.4% + 11.7% = 22.6%
• High CI for matrix assembling & Low CI for linear solver
assemble_K
assemble_V
gmres_solve
Low CI - 17%
High CI isdominant83% of RT
Ener
gy [
J]
CF [GHz]
static dynamic total
Ener
gy [
J]
737J
1788J
1833J
Ener
gy [
J]
CF [GHz]
CF [GHz]
Review Meeting, 23.05.2017, Brussels – WP5 59