How GPU Computing literally saved me at work!
Download as PPTX, PDF1 like66 views
The document discusses how GPU computing significantly optimized a recommendation system at Walmart, reducing computation time from 2 days to 20.5 seconds for identifying similar items. It details the infrastructure used, including NVIDIA GPUs and Python, as well as the complexity of the task, which involves calculating cosine similarity among items. The findings emphasize the importance of parallelism in tasks for effective GPU utilization and suggest a hybrid approach using both CPU and GPU based on task requirements.
How GPU Computing literally saved me at work!
- 1. 1© 2018 Walmart International | Confidential | For Internal Use Only How GPU Computing literally saved me at work! Abhishek Mungoli, Data Scientist, MTech(IIIT, Hyd.) Date: June 11th, 2019 Prepared for Fifth Elephant
- 2. 2© 2018 Walmart International | Confidential | For Internal Use Only Content • Problem Description • Solutions for run time optimizations • Focus on GPU- Infra used • Task Complexity details • Solution Framework • CPU v/s GPU-Theory + Implementation • Assumptions • Take Away
- 3. 3© 2018 Walmart International | Confidential | For Internal Use Only Identifying Similar Items • Recommendation systems • Item alternatives • Assortment • Customer basket customization
- 4. 4© 2018 Walmart International | Confidential | For Internal Use Only EngineeringTools Distributed processing Spark GPU
- 5. 5© 2018 Walmart International | Confidential | For Internal Use Only Infrastructure Used • Nvidia Volta V100 16GB GPU • Python • Numba
- 6. 6© 2018 Walmart International | Confidential | For Internal Use Only Task Complexity Computation TimeSize • Number of Items • 105 • Dimension of each item • 64-D vector • Task • Identify top 3 similar items to each item in set • Cosine Similarity • For finding items with highest similarity, for one item • O(n*k) • For finding items with highest similarity, for all items • O(n*n*k) • On a subset of 103 items • 17 seconds • 3.7 *10⁶ operations per second • On a subset of 104 items • 1700 seconds = ~28 minutes • On 105 item set • 1700 * 100 seconds = 2834 minutes = 47.2 hours ~ 2 days
- 7. 7© 2018 Walmart International | Confidential | For Internal Use Only CPU v/s GPU Few Cores Thousands of Cores Single-thread optimization Multiple concurrent threads Low LatencyTolerance High LatencyTolerance • When to Use GPU????
- 8. 8© 2018 Walmart International | Confidential | For Internal Use Only Solution Framework • Embeddings can be obtained at different level of hierarchies i.e. Category, Sub-category, Fineline, UPC or Item level. • Similar items have similar embeddings. Assumptions Work Flow
- 9. 9© 2018 Walmart International | Confidential | For Internal Use Only Take Away • The CPU estimated time of 2 days was brought down to 20.5 seconds with the use of GPU. • This was possible only because of the nature of the task. Finding top-3 similar items to Item ‘A’ is independent of finding top-3 similar items to Item ‘B’. • GPU’s can be used for fast text searching in document, fast searching of a node in graph, etc.We need to identify the parallelism in task and exploit GPU for speed up. • We can identify the components of the system/module parallel in nature and speed them up.This way we can have a system/module with some components running in CPU and some in GPU as per need and necessity.
- 10. 10© 2018 Walmart International | Confidential | For Internal Use Only Q&A
- 11. 11© 2018 Walmart International | Confidential | For Internal Use Only References • https://en.wikipedia.org/wiki/Time_complexity • https://en.wikipedia.org/wiki/Graphics_processing_unit • https://blogs.nvidia.com/blog/2009/12/16/whats-the-difference-between-a-cpu-and-a-gpu/ • https://www.datascience.com/blog/cpu-gpu-machine-learning • https://qr.ae/TWIuic • https://numba.pydata.org/numba-doc/latest/index.html • https://en.wikipedia.org/wiki/CUDA • https://www.nvidia.in/object/cuda-parallel-computing-in.html • https://llpanorama.wordpress.com/2008/06/11/threads-and-blocks-and-grids-oh-my/ • https://qr.ae/TWIwEW
- 12. 12© 2018 Walmart International | Confidential | For Internal Use Only ThankYou Abhishek Mungoli, Data Scientist, Walmart LinkedIn - https://www.linkedin.com/in/abhishek-mungoli-39048355/ Medium - https://medium.com/@mungoliabhishek81 Instagram - https://www.instagram.com/simplyspartanx/
Editor's Notes
- #4: In the retail domain, finding similar or closest entities is a common task. Provided with a list of items where each item is represented with k latent attributes. Task is to find top-3 similar items for every item in the list.