![Flow Diagram of
ppOpen‐APPL/FDM
13
),,,(
}],,,)
2
1
({},,)
2
1
({[
1
),,(
2/
1
zyxqp
zyxmxzyxmxc
x
zyx
dx
d
pqpq
M
m
m
pq
Space difference by FDM.
),,(,
12
1
2
1
zyxptf
zyx
uu n
p
n
zp
n
yp
n
xpn
p
n
p
Explicit time expansion by central
difference.
Initialization
Velocity Derivative (def_vel)
Velocity Update (update_vel)
Stress Derivative (def_stress)
Stress Update (update_stress)
Stop Iteration?
NO
YES
End
Velocity PML condition (update_vel_sponge)
Velocity Passing (MPI) (passing_vel)
Stress PML condition (update_stress_sponge)
Stress Passing (MPI) (passing_stress)](https://image.slidesharecdn.com/siampp16-hp-160528152937/85/Auto-Tuning-of-Hierarchical-Computations-with-ppOpen-AT-13-320.jpg)
Auto‐Tuning of Hierarchical Computations with ppOpen‐AT
- 1. Auto‐Tuning of Hierarchical Computations with ppOpen‐AT Takahiro Katagiri i),ii),iii) , Masaharu Matsumoto i),ii), Satoshi Ohshima i),ii) 1 17th SIAM Conference on Parallel Processing for Scientific Computing (SIAM PP16) Universite Pierre et Marie Curie, Cordeliers Campus, Paris, France Thursday, April 14, 2016, MS55:Auto‐Tuning for the Post Moore’s Era – Part I of II, 2:40‐3:00 i) Information Technology Center, The University of Tokyo ii) JST, CREST iii) Currently, Information Technology Center, Nagoya University
- 2. Outline 1. Background 2. ppOpen‐HPC and ppOpen‐AT 3. Code Selection in Seism3D and Its Implementation by ppOpen‐AT 4. Performance Evaluation 5. Conclusion
- 3. Outline 1. Background 2. ppOpen‐HPC and ppOpen‐AT 3. Code Selection in Seism3D and Its Implementation by ppOpen‐AT 4. Performance Evaluation 5. Conclusion
- 4. Source: http://www.engad get.com/2011/12/0 6/the‐big‐memory‐ cube‐gamble‐ibm‐ and‐micron‐stack‐ their‐chips/ Reconstruction of algorithms w.r.t. Increase memory bandwidth Increase local memory amounts (caches) Non‐Uniform memory accesses latency Issues in AT Technologies. Hierarchical AT Methodology Reducing Communication Algorithm. New algorithms and code (algorithm) selection utilizing high bandwidth ability. Rethink classical algorithms. Non‐Blocking Algorithms. From explicit method to implicit method. Out of Core algorithms (Out of main memory) Source: http://www.there gister.co.uk/2015/ 07/28/intel_micro n_3d_xpoint/ • It is expected that Moore’s Low is broken around end of 2020. – End of “One‐time Speedup”: Many Cores, Wiring miniaturization to reduce power. → It cannot increase FLOPS inside node. • However, memory bandwidth inside memory can increase by using “3D Stacking Memory” Technologies. • 3D Stacking Memory: – It can increase bandwidth for Z‐Direction (Stacking distance) , and keeping access latency (→ High performance) – It can be low access latency for X‐Y directions. • Access latency between nodes goes down, but bandwidth can be increased by optical interconnection technology. • We need to take care of new algorithms with respect to ability of data movements. HMC Intel 3D Xpoint AT Technologies in Post Moore’s Era
- 5. Outline 1. Background 2. ppOpen‐HPC and ppOpen‐AT 3. Code Selection in Seism3D and Its Implementation by ppOpen‐AT 4. Performance Evaluation 5. Conclusion
- 6. ppOpen‐HPC Project • Middleware for HPC and Its AT Technology – Supported by JST, CREST, from FY2011 to FY2015. – PI: Professor Kengo Nakajima (U. Tokyo) • ppOpen‐HPC – An open source infrastructure for reliable simulation codes on post‐peta (pp) scale parallel computers. – Consists of various types of libraries, which covers 5 kinds of discretization methods for scientific computations. • ppOpen‐AT – An auto‐tuning language for ppOpen‐HPC codes – Using knowledge of previous project： ABCLibScript Project 4). – Auto‐tuning language based on directives for AT. 6
- 7. Software Architecture of ppOpen‐HPC 7 FVM DEMFDMFEM Many‐core CPUs GPUs Multi‐core CPUs MG COMM Auto‐Tuning Facility Code Generation for Optimization Candidates Search for the best candidate Automatic Execution for the optimization ppOpen‐APPL ppOpen‐MATH BEM ppOpen‐AT User Program GRAPH VIS MP STATIC DYNAMIC ppOpen‐SYS FT Optimize memory accesses
- 9. Scenario of AT for ppOpen‐APPL/FDM 9 Execution with optimized kernels without AT process. Library User Set AT parameters, and execute the library (OAT_AT_EXEC=1) ■Execute auto-tuner: With fixed loop lengths (by specifying problem size and number of MPI processes) Time measurement for target kernels Store the best variant information. Set AT parameters, and execute the library (OAT_AT_EXEC=0) Store the fastest kernel information Using the fastest kernel without AT (except for varying problem size, number of MPI processes and OpenMP threads.) Specify problem size, number of MPI processes and OpenMP threads.
- 10. AT Timings of ppOpen‐AT (FIBER Framework) OAT_ATexec() … do i=1, MAX_ITER Target_kernel_k() … Target_kernel_m() … enddo Read the best parameter Is this first call? Yes Read the best parameter Is this first call? Yes One time execution (except for varying problem size and number of MPI processes ) AT for Before Execute‐time Execute Target_Kernel_k() with varying parameters Execute Target_Kernel_m() with varying parameters parameter Store the best parameter …
- 11. Outline 1. Background 2. ppOpen‐HPC and ppOpen‐AT 3. Code Selection in Seism3D and Its Implementation by ppOpen‐AT 4. Performance Evaluation 5. Conclusion
- 12. Target Application • Seism3D: Simulation for seismic wave analysis. • Developed by Professor T.Furumura at the University of Tokyo. –The code is re‐constructed as ppOpen‐APPL/FDM. • Finite Differential Method (FDM) • 3D simulation –3D arrays are allocated. • Data type: Single Precision (real*4) 12
- 13. Flow Diagram of ppOpen‐APPL/FDM 13 ),,,( }],,,) 2 1 ({},,) 2 1 ({[ 1 ),,( 2/ 1 zyxqp zyxmxzyxmxc x zyx dx d pqpq M m m pq Space difference by FDM. ),,(, 12 1 2 1 zyxptf zyx uu n p n zp n yp n xpn p n p Explicit time expansion by central difference. Initialization Velocity Derivative (def_vel) Velocity Update (update_vel) Stress Derivative (def_stress) Stress Update (update_stress) Stop Iteration? NO YES End Velocity PML condition (update_vel_sponge) Velocity Passing (MPI) (passing_vel) Stress PML condition (update_stress_sponge) Stress Passing (MPI) (passing_stress)
- 14. Original Implementation (For Vector Machines) call ppohFDM_pdiffx3_m4_OAT( VX,DXVX, NXP,NYP,NZP,NXP0,NXP1,NYP0,…) call ppohFDM_pdiffy3_p4_OAT( VX,DYVX, NXP,NYP,NZP,NXP0,NXP1,NYP0,…) call ppohFDM_pdiffz3_p4_OAT( VX,DZVX, NXP,NYP,NZP,NXP0,NXP1,NYP0,…) call ppohFDM_pdiffy3_m4_OAT( VY,DYVY, NXP,NYP,NZP,NXP0,NXP1,NYP0,… ) call ppohFDM_pdiffx3_p4_OAT( VY,DXVY, NXP,NYP,NZP,NXP0,NXP1,NYP0,… ) call ppohFDM_pdiffz3_p4_OAT( VY,DZVY, NXP,NYP,NZP,NXP0,NXP1,NYP0,…) call ppohFDM_pdiffx3_p4_OAT( VZ,DXVZ, NXP,NYP,NZP,NXP0,NXP1,NYP0,…) call ppohFDM_pdiffy3_p4_OAT( VZ,DYVZ, NXP,NYP,NZP,NXP0,NXP1,NYP0,… ) call ppohFDM_pdiffz3_m4_OAT( VZ,DZVZ, NXP,NYP,NZP,NXP0,NXP1,NYP0,…) if( is_fs .or. is_nearfs ) then call ppohFDM_bc_vel_deriv( KFSZ,NIFS,NJFS,IFSX,IFSY,IFSZ,JFSX,JFSY,JFSZ ) end if call ppohFDM_update_stress(1, NXP, 1, NYP, 1, NZP) Fourth‐order accurate central‐difference scheme for velocity. (def_stress) Process of model boundary. Explicit time expansion by leap‐frog scheme. (update_stress)
- 19. Outline 1. Background 2. ppOpen‐HPC and ppOpen‐AT 3. Code Selection in Seism3D and Its Implementation by ppOpen‐AT 4. Performance Evaluation 5. Conclusion
- 20. The Number of AT Candidates (ppOpen‐APPL/FDM) 20 Kernel Names AT Objects The Number of Candidates 1. update_stress ・Loop Collapses and Splits ：8 Kinds ・Code Selections : 2 Kinds 10 2. update_vel ・Loop Collapses, Splits, and re‐ ordering of statements: ：6 Kinds ・Code Selections: 2 Kinds 8 3. update_stress_sponge ・Loop Collapses：3 Kinds 3 4. update_vel_sponge ・Loop Collapses：3 Kinds 3 5. ppohFDM_pdiffx3_p4 Kernel Names：def_update、def_vel ・Loop Collapses：3 Kinds 3 6. ppohFDM_pdiffx3_m4 3 7. ppohFDM_pdiffy3_p4 3 8. ppohFDM_pdiffy3_m4 3 9. ppohFDM_pdiffz3_p4 3 10.ppohFDM_pdiffz3_m4 3 11. ppohFDM_ps_pack Data packing and unpacking ・Loop Collapses： 3 Kinds 3 12. ppohFDM_ps_unpack 3 13. ppohFDM_pv_pack 3 14. ppohFDM_pv_unpack 3 Total：54 Kinds Hybrid MPI/OpenMP: 7 Kinds 54×7 = 378 Kinds
- 21. Machine Environment (8 nodes of the Xeon Phi) The Intel Xeon Phi Xeon Phi 5110P (1.053 GHz), 60 cores Memory Amount：8 GB (GDDR5) Theoretical Peak Performance：1.01 TFLOPS One board per node of the Xeon phi cluster InfiniBand FDR x 2 Ports Mellanox Connect‐IB PCI‐E Gen3 x16 56Gbps x 2 Theoretical Peak bandwidth 13.6GB/s Full‐Bisection Intel MPI Based on MPICH2, MVAPICH2 Version 5.0 Update 3 Build 20150128 Compiler：Intel Fortran version 15.0.3 20150407 Compiler Options： ‐ipo20 ‐O3 ‐warn all ‐openmp ‐mcmodel=medium ‐shared‐intel –mmic ‐align array64byte KMP_AFFINITY=granularity=fine, balanced (Uniform Distribution of threads between sockets)
- 22. Execution Details • ppOpen‐APPL/FDM ver.0.2 • ppOpen‐AT ver.0.2 • The number of time step: 2000 steps • The number of nodes: 8 node • Native Mode Execution • Target Problem Size (Almost maximum size with 8 GB/node) – NX * NY * NZ = 512 x 512 x 512 / 8 Node – NX * NY * NZ = 256 * 256 * 256 / node (!= per MPI Process) • The number of iterations for kernels to do auto‐tuning: 100
- 23. Execution Details of Hybrid MPI/OpenMP • Target MPI Processes and OMP Threads on the Xeon Phi – The Xeon Phi with 4 HT (Hyper Threading) – PX TY: X MPI Processes and Y Threads per process – P8T240 : Minimum Hybrid MPI/OpenMP execution for ppOpen‐APPL/FDM, since it needs minimum 8 MPI Processes. – P16T120 – P32T60 – P64T30 – P128T15 – P240T8 – P480T4 – Less than P960T2 cause an MPI error in this environment. # 0 # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8 # 9 # 1 0 # 1 1 # 1 2 # 1 3 # 1 4 # 1 5 P2T8 # 0 # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8 # 9 # 1 0 # 1 1 # 1 2 # 1 3 # 1 4 # 1 5 P4T4 Target of cores for one MPI Process
- 26. Outline 1. Background 2. ppOpen‐HPC and ppOpen‐AT 3. Code Selection in Seism3D and Its Implementation by ppOpen‐AT 4. Performance Evaluation 5. Conclusion
- 27. RELATED WORK
- 28. Originality (AT Languages) AT Language / Items # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8 ppOpen‐AT OAT Directives ✔ ✔ ✔ ✔ ✔ None Vendor Compilers Out of Target Limited ‐ Transformation Recipes Recipe Descriptions ✔ ✔ ChiLL POET Xform Description ✔ ✔ POET translator, ROSE X language Xlang Pragmas ✔ ✔ X Translation, ‘C and tcc SPL SPL Expressions ✔ ✔ ✔ A Script Language ADAPT ADAPT Language ✔ ✔ Polaris Compiler Infrastructure, Remote Procedure Call (RPC) Atune‐IL Atune Pragmas ✔ A Monitoring Daemon PEPPHER PEPPHER Pragmas (interface) ✔ ✔ ✔ PEPPHER task graph and run-time Xevolver Directive Extension (Recipe Descriptions) (✔) (✔) (✔) ROSE, XSLT Translator #1: Method for supporting multi-computer environments. #2: Obtaining loop length in run-time. #3: Loop split with increase of computations6) ,and loop collapses to the split loops6),7),8) . #4: Re-ordering of inner-loop sentences8) . #5:Code selection with loop transformations (Hierarchical AT descriptions*) *This is originality in current researches of AT as of 2015. #6:Algorithm selection. #7: Code generation with execution feedback. #8: Software requirement. (Users need to define rules. )
- 30. Future Work • Improving Search Algorithm – We use a brute‐force search in current implementation. • This is feasible by applying knowledge of application. – We have implemented a new search algorithm based on black box performance models. • d‐Spline Model (interpolation and incremental addition based) ,collaborated with Prof. Tanaka (Kogakuin U.) • Surrogate Model (interpolation and probability based) collaborated with Prof. Wang (National Taiwan U.) • Off‐loading Implementation Selection (for the Xeon Phi) – If problem size is too small to do off‐loading, the target execution is performed on CPU automatically. • Adaptation of OpenACC for GPU computing – Selection of OpenACC directives with ppOpenAT by Dr. Ohshima. • gang, vector, parallel, etc.