Korali a High-Performance Multi-Intrusive Bayesian Inference Software for Large-Scale Scientific Models G. Arampatzis, S. Martin, D. Wälchli, and P. Koumoutsakos # 1 y ax 1 y tr 1 y ax,new 1 y tr,new 1 # 2 y ax 2 y tr 2 y ax,new 2 y tr,new 2 # 3 y sh 3 y sh,new 3 # 7 y sh 7 y sh,new 7 ... ... # new y new 1 2 stretching shear # i =( Q 1,i ,Q 2,i ,μ 0,i , σ st,i ),i =1, 2 # i =( Q 1,i ,Q 2,i ,μ 0,i , Q 3,i ,Q 4,i ,, σ sh,i ),i =3,..., 7 # new =( Q 1 ,Q 2 ,μ 0 , Q 3 ,Q 4 , σ sh , σ st ) 5 6 9 11 12 13 14 15 16 17 18 19 20 1 2 3 4 7 8 10 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 • Optimal Sensor Placement • Hierarchical Bayesian Infernce p (ϑ 1 | d 1 , ℳ 1 ) + U (s)= Z Y Z R log p(y |r , s) p(y |s) p(r ) p(y |r , s)dr dy s ? = argmax s ˆ U (s) Korali Problem SolYer EYalXaWion E[ecXWion GaXVVian ProceVV Model GaXVVian ProceVV Ba\eVian DirecW Ζnference Hierarchical Reference CXVWom PVi TheWa E[ecXWor OpWimiVer Sampler CMA-ES CCMA-ES LM-CMA-ES DEA Rprop CondXiW ConcXrenW DiVWribXWed SeqXanWial MCMC DRAM TMCMC Extensible Ease of Use Motivation Workflow 5orker 9 Rank 0 _Core 0` Rank 1 _Core 1` 5orker : Rank 0 _Core 2` Rank 1 _Core <` 5orker ; Rank 0 _Core =` Rank 1 _Core >` 5orker < Rank 0 _Core ?` Rank 1 _Core @` #periment ' 1olver .roblem !onduit waiting for sample Generate Samples Preprocess Samples Postprocess Results Distribute Samples Collect Results #periment '' 1olver .roblem Preprocess Samples Postprocess Results Update State Generate Samples #ngine finsihed busy Update State SWaUW GeneUaWion RXn E[SeUimenW Check TeUminaWion waiting for sample Korali Supervisor Supercomputer p(# | d)= p(d | #) p(#) p(d) • Bayesian Inference Load Balance Unbalanced Balanced Scalable Computational Model Relaxation Wime (hRXUV) JRb 1 ChecN PRLQW 1 JRb 2 ReVXme O(1000) NRdeV ChecN PRLQW 2 ChecN PRLQW N Fault Tolerant system error