1. A modeling approach for estimating execution time of long-running Scientific Applications Seyed Masoud Sadjadi 1, Shu Shimizu 2, Javier Figueroa 1,3, Raju Rangaswami 1, Javier Delgado 1, Hector Duran 4, Xabriel J. Collazo-Mojica 5 Presented by: Xabriel J. Collazo-Mojica 5 1: Florida International University (FIU), Miami, Florida, USA; 2: IBM Tokyo Research Laboratory, Tokyo, Japan; 3: University of Miami, Coral Gables, Florida, USA; 4: University of Guadalajara, CUCEA, Mexico; 5: University of Puerto Rico, Mayagüez Campus, Puerto Rico Miami, Florida – April 2008

2. Presentation Outline Motivation Research Approach Research Validation Related Work Concluding Remarks Future Research HPGC '08 - April 14 - LA Grid2

3. Motivation The impact of hurricanes is devastating The Weather Research and Forecasting (WRF) model Most popular It is computational and storage intensive We need higher resolution and more precise forecast Many organizations are willing to share resources But these resources are dynamic and unpredictable HPGC '08 - April 14 - LA Grid3

4. Motivation At the time of a hurricane, we need to act fast What resources should we allocate? We need to finish in a strict deadline (i.e. on time for hurricane forecast) In the order of seconds, we need to make a decision We need to model execution time of WRF based on target resources In our case: clusters with different parameters HPGC '08 - April 14 - LA Grid4

5. Approach to Modeling Resource Usage WRF HPGC '08 - April 14 - LA Grid5

6. Approach to Modeling Execution Parallelism Platform heterogeneity We assume identical individual resource characteristics of computation, communication and storage power. Execution scale We add a parameter to model the number of nodes utilized during execution. 123N … HPGC '08 - April 14 - LA Grid6

7. Application Resource Usage Model Characterize Applications according to their resource usage characteristics (i.e. application "profiles”) Assumptions: Execution time is based on contributors Product of contributors determines total execution time Computation nodes are homogeneous (e.g. Beowulf cluster) Non-ad-hoc application characteristics HPGC '08 - April 14 - LA Grid7

8. Application Resource Usage Model - Contributors Model aims to allow as many contributors as necessary This paper focus: 2 contributors First contributor: Parallelism P para = degree of parallelism α 0 = constant contribution α 1 = variable contribution Second contributor: CPU Performance P clock = clock speed of compute node ß 0 = constant contribution related to CPU performance ß 1 = variable contribution related to CPU performance HPGC '08 - April 14 - LA Grid8

9. Experimental Approach - Environment GCB cluster: Rocks ver. 4.0, 8 nodes, each containing 32-bit x86 Intel 3.0 GHz processors, 1GB of main memory and uses a gigabit network connection Mind cluster: Rocks ver. 4.0, 16 nodes, each containing dual Xeon 3.6GHz processors, 2GB of main memory and uses gigabit network connection CPU vs. #-of-NODES:100% to 10% CPU percentages with intervals of 10% We use CPULimit HPGC '08 - April 14 - LA Grid9

10. Experimental Approach - Monitoring and Prediction Two tools were used Amon – A Monitoring Tool Daemon-like application that collects and reports exploratory variables Aprof – A Profiling Tool Statistical Prediction Program Listens to Amon reports from compute nodes Stores collected data as matrix for each application HPGC '08 - April 14 - LA Grid10

11. Experimental Approach - Monitoring and Prediction HPGC '08 - April 14 - LA Grid11

12. Application Resource Usage Model - Validation Intuitive Assumption that execution time decreases linearly with the inverse of total computational power. Predictions within a cluster (i.e. GCB to GCB) GCB - FE 5.34% ME 5.86% Mind - FE 5.66% ME 3.80% Predictions across clusters GCB to Mind - FE 9.97% ME 5.86% Mind to GCB - FE 5.83% ME 4.13% This results validate our simple model. HPGC '08 - April 14 - LA Grid12

13. Application Resource Usage Model - Mind to GCB prediction HPGC '08 - April 14 - LA Grid13

14. Concluding Remarks We've proposed a new approach for modeling resource usage and execution time of a distributed application Experimental results using WRF execution on two different clusters show good accuracy - within 10% from across cluster predictions Using only two parameters - CPU speed and number of nodes. WRF specific, we are one step closer to devising a complete solution for our goal of higher-resolution weather predictions and simulations. HPGC '08 - April 14 - LA Grid14

15. Related Work S. Shimizu, R. Rangaswami, and H. A. Duran-Limon. "Platform-independent Modeling and Prediction of Application Resource Usage Characteristics” Basis for prediction model It is limited to one node D. M. Swany and R. Wolski. “Multivariate Resource Performance Forecasting In the Network Weather Service.” High-accuracy prediction model They emphasize latency and bandwidth HPGC '08 - April 14 - LA Grid15

16. Related Work R. Badia, F. Escale, E. Gabriel, J. Gimenez, R. Keller, J. Labarta, M. S. Müller, Perf. “Prediction in a Grid Environment.” Offline prediction Need to link their library to the application to be profiled HPGC '08 - April 14 - LA Grid16

17. Future Research Extend our parallelism model to address heterogeneous resources. Include more resource parameters to the model Started joint research with Barcelona Supercomputing Center We acknowledge that Amon & Aprof have limitations We will integrate our tools with their simulation application - DIMEMAS HPGC '08 - April 14 - LA Grid17

18. Acknowledgements National Science Foundation –REU Grant # IIS-0552555 –PIRE Grant # OISE-0730065 –CREST Grant # HRD-0317692 –GCB Grant # OCI-0636031 IBM Research LA Grid FIU SCIS HPGC '08 - April 14 - LA Grid18

19. Questions?