1 / 25

Weekly Report Start learning GPU

Weekly Report Start learning GPU . Ph.D. Student: Leo Lee date: Sep. 18, 2009. Outline. References CUDA Work plan. Outline. References CUDA Work plan. References. Frequent itemset mining on graphics. Introduction Two representative algorithms: Apriori and FP-growth;

gryta
Download Presentation

Weekly Report Start learning GPU

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Weekly ReportStart learning GPU Ph.D. Student: Leo Leedate: Sep. 18, 2009

  2. Outline • References • CUDA • Work plan

  3. Outline • References • CUDA • Work plan

  4. References

  5. Frequent itemset mining on graphics • Introduction • Two representative algorithms: Apriori and FP-growth; • FP-growth were generally faster than Apriori; • Apriori-borgelt was slightly faster when the support was high; • No prior work focuses on studying the GPU acceleration for FIM algorithms. • Challenge: the data structure is not aligned and access patterns are not regular (pointer-chasing).

  6. Frequent itemset mining on graphics • Background and related work-GPGPU • The parallel primitives [19] are a small set of common operations exploiting the architectural features of GPUs. We utilize map, reduce, and prefix sum primitives in our two FIM implementations. • Improvement - Memory optimizations: • Local memory optimization for temporal locality • Coalesced access optimization of device memory for spatial locality • The built-in vector data type to reduce the number of memory access. • Difference • we study the GPU acceleration of Apriori for FIM, which incurs much more complex control fows and memory accesses than performing database joins or maintaining quantiles from data streams.

  7. Frequent itemset mining on graphics • Implementation

  8. Frequent itemset mining on graphics • Implementation

  9. Frequent itemset mining on graphics • Implementation-Pure Bitmap Implementation

  10. Frequent itemset mining on graphics Given m frequent (K ¡1)-itemsets, and n items. In order to check whether one (K ¡ 1)-itemset is frequent, we need to access (logm*(n/128)*16) bytes of data, where logm is the cost of performing a binary search, and (n/128)*16 is the size of a row (in bytes) in the bitmap of (K¡1)-itemsets. Typically, if m = 10000 and n = 10000, we need to access about 16 KB for checking only one (K ¡ 1)-subset. This problem in our pure bitmap- based solution triggers us to consider adopting another data structure in the Candidate Generation procedure in the presence of a large number of items. • Implementation-PBI

  11. Frequent itemset mining on graphics • Implementation-Trie based Implemetation The candidate generation based on trie traversal is implemented on the CPU. This decision is based on the fact that, the trie is an irregular structure and difficult to share among SIMD threads. Thus, we store the trie representing itemsets in the CPU memory, and the bitmap representation of transactions in the GPU device memory.

  12. Frequent itemset mining on graphics • Implementation-TBI

  13. Frequent itemset mining on graphics • Experiments

  14. Frequent itemset mining on graphics • Experiments

  15. Frequent itemset mining on graphics • Results

  16. Frequent itemset mining on graphics • Results

  17. Frequent itemset mining on graphics • Results

  18. Frequent itemset mining on graphics • Results

  19. Outline • References • CUDA • Work plan

  20. CUDA • Review the code of K-means • CPU: 1101 S (10 S) • GPU: still need debug, no results right now

  21. Outline • References • CUDA • Work plan

  22. Work Plan • Summary this month • Make plan for next month • Try to implement a data mining algorithm • Homework

  23. References

  24. Thanks for your listening

More Related