Parallel Cartographic Modeling 1 This template can be used as a starter file to give updates for project milestones. Sections Right-click on a slide to add sections. Sections can help to organize your slides or facilitate collaboration between multiple authors. Notes Use the Notes section for delivery notes or to provide additional details for the audience. View these notes in Presentation View during your presentation. Keep in mind the font size (important for accessibility, visibility, videotaping, and online production) Coordinated colors Pay particular attention to the graphs, charts, and text boxes. Consider that attendees will print in black and white or grayscale. Run a test print to make sure your colors work when printed in pure black and white and grayscale. Graphics, tables, and graphs Keep it simple: If possible, use consistent, non-distracting styles and colors. Label all graphs and tables. Parallel Cartographic Modeling Kiumars Soltani 20th July 2012

Outline What is the project about? Define the goal of this project 2 What is the project about? Define the goal of this project Is it similar to projects in the past or is it a new effort? Define the scope of this project Is it an independent project or is it related to other projects? * Note that this slide is not necessary for weekly status meetings Outline Debugging the Code Model 1 Model 2 Future Challenges

* If any of these issues caused a schedule delay or need to be discussed further, include details in next slide. 3 Debugging the Code A mistake in indexing the DirectionRaster was giving us wrong result! Count = … //offset For(j=0; j < 600; j++) { /* Calculating DirectionRaster */ DirectionRaster[count] = … } DirectionRaster[count+j] = …

4 Duplicate this slide as necessary if there is more than one issue. This and related slides can be moved to the appendix or hidden if necessary. The results The result from the previous version that was presented in the last presentation was for the 900k dataset, not 90k dataset. Therefore, that result should have been compared to our result for the largest dataset. That might lead to dramatic difference in the final efficiency as we have issues in scalability. All the result are captured using 32 cores on Trestles.

5 Duplicate this slide as necessary if there is more than one issue. This and related slides can be moved to the appendix or hidden if necessary. Model 1 Taking advantage of locality Store the local result of each core in a privately defined array. Do the global sum on the local_result arrays. Using static scheduling with the chunk size of 1. Using nowait clause on the main loop

Model 1 – Result for 90k dataset 6 Duplicate this slide as necessary if there is more than one issue. This and related slides can be moved to the appendix or hidden if necessary. Model 1 – Result for 90k dataset Reading File Euclidean Direction and Distance Computation Writing Output Old Version 0.620 5.05 1057.61 1 Version 2.1 0.234 0.018 72.258 0.402

Model 1 – Result for 900k dataset 7 Duplicate this slide as necessary if there is more than one issue. This and related slides can be moved to the appendix or hidden if necessary. Model 1 – Result for 900k dataset Reading File Euclidean Direction and Distance Computation Writing Output Old Version 7.140 5.120 9515.070 1.060 Version 2.1 2.434 0.034 736.340 0.465

Profiling – Result for 900k dataset

Result for 900k dataset

10 Duplicate this slide as necessary if there is more than one issue. This and related slides can be moved to the appendix or hidden if necessary. Model 2 Was considerably faster for smaller values. Doesn’t scale well! The initial inspection shows that the most time consuming part is writing the output. Writing the output in this model is much less than the model 1! Using static scheduling with the chunk size of 1. Using nowait clause on the main loop

Model 2 – Result for 90k dataset 11 Duplicate this slide as necessary if there is more than one issue. This and related slides can be moved to the appendix or hidden if necessary. Model 2 – Result for 90k dataset Reading File Euclidean Direction and Distance Computation Writing Output Old Version 0.62 5.05 1052.39 1 Version 2.1 0.234 0.030 91.350 0.392

Model 2 – Result for 900k dataset 12 Duplicate this slide as necessary if there is more than one issue. This and related slides can be moved to the appendix or hidden if necessary. Model 2 – Result for 900k dataset Reading File Euclidean Direction and Distance Computation Writing Output Old Version 7.140 5.120 9515.070 1.060 Version 2.1 2.327 0.022 2450.467 423.309

Profiling – Result for 900k dataset

Future works Running the code for the dataset provided by April. She is interested in investigating if the 12 hours period data for a single day produce similar result. Doing more advanced profiling (with Babak Behzad) Profiling the I/O related issues for Model 2(Cache miss, … ) This code’s improvement is basically done. What we are expecting to get out of this implementation. Domain Specific Language CUDA, MPI, …

15 Prepare slides for the appendix in the event that more details or supplemental slides are needed. The appendix is also useful if the presentation is distributed later. Thank You! Got Question?