Performance Evaluation of P2P and Cloud Computing Applications - A Module for SimGRID Bogdan Cornea, Julien Bourgeois, Vaidy Sunderam 14 June 2012, Valpré, Ecully
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 2/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 3/25
Origins of an idea ● Alternative to grids and clusters – Computing ● Infinite ressources of P2P ● Lower cost ● Existence of – ChronosMix tool – Real computing applications – Budget ● P2P computing frame ● P2P performance prediction tool 4/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 5/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 6/25
dPerf for performance prediction Input code C , C++, Fortran / Trace Automatic P2P-SAP, MPI replay code translator Scaling of Network scaling N° nodes ROSE compiler PAPI SimGrid MSG Scaling the N° nodes HW counters Prediction 7/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 8/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 9/25
dPerf and P2P apps. ● Target: P2P computing – Heterogeneous – P2P communication protocol ● Adapted to P2P performance computing ● Developed at LAAS (France) – Real code ported to P2P computing ● By LAAS (France) – Scaling 10/25
dPerf and P2P apps. ● How ? – Automatic instrumentation + execution ● Simple block benchmarking ● Take in account compiler optimization NAS IS ; Gcc optimization level : 0. Time [s] 11/25 Procs.
dPerf and P2P apps. ● How ? ● optimized block benchmarking*** NAS IS ; Gcc optimization level : 0. Time [s] Procs. 12/25
dPerf and P2P apps. ● How ? – Scaling of results ● Trace-based simulation with SimGrid MSG – Network configurations Obstacle problem ; P2P implementation Gcc optimization level : 0. Time [s] 13/25 Procs.
dPerf and P2P apps. ● How ? – Scaling of results ● Trace-based simulation with SimGrid MSG – Number of nodes 14/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 15/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 16/25
Module for SimGrid ● Initially ● After integration Input code Input code dPerf SMPI MSG dPerf SimGrid SimGrid 17/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 18/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 19/25
dPerf for Cloud apps. ● Ongoing ● Many perspectives 20/25
dPerf for Cloud apps. ● Phase 1 MPI HPC code Static analysis, dPerf Static communication topology identification Emory University adapter MPI to any Cloud Cloud 21/25
dPerf for Cloud apps. ● Phase 2 – Performance MPI HPC code prediction of applications on Cloud dPerf – Fine tune dPerf SMPI MSG adapter – Compare original SimGrid performance to adapted performance Cloud performance prediction 22/25 tuning
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 23/25
● Origins ● Performance prediction – dPerf ● dPerf and P2P apps. ● Module for SimGrid ● dPerf for Cloud apps. ● Future work 24/25
Future work ● Support for C++, Fortran ● Multi-core ● Memory ● SMPI ? dPerf ? MSG – Helps the integration process ● Compare to other tools based on MSG 25/25
Recommend
More recommend