Novel Parallel Algorithms for Fast Multi-GPU-Based Generation of Massive Scale-Free Networks

Abstract

A novel parallel algorithm is presented for generating random scale-free networks using the preferential attachment model. The algorithm, named cuPPA, is custom-designed for "single instruction multiple data (SIMD)" style of parallel processing supported by modern processors such as graphical processing units (GPUs). To the best of our knowledge, our algorithm is the first to exploit GPUs, and also the fastest implementation available today, to generate scale-free networks using the preferential attachment model. A detailed performance study is presented to understand the scalability and runtime characteristics of the cuPPA algorithm. Also another version of the algorithm called cuPPA-Hash tailored for multiple GPUs is presented. On a single GPU, the original cuPPA algorithm delivers the best performance, but is challenging to port to multi-GPU implementation. For multi-GPU implementation, cuPPA-Hash has been used as the parallel algorithm to achieve a perfect linear speedup up to 4 GPUs. In one of the best cases, when executed on an NVidia GeForce 1080 GPU, the original cuPPA generates a scale-free network of two billion edges in less than 3 s. On multi-GPU platforms, cuPPA-Hash generates a scale-free network of 16 billion edges in less than 7 s using a machine consisting of 4 NVidia Tesla P100 GPUs.

Multi-task learning based on question–answering style reviews for aspect category classification and aspect term extraction on GPU clusters

“…In recent years, more paralleled deep learning methods have been brought up [11]. In the aspect of algorithms, several algorithms have been brought up to accelerate multi-GPU implementation or make the inference more accurate [1,26] and faster [7,12]. Moreover, there are researches have been done to integrate DP and MP [8].…”

Scalable generation of random graphs

“…Alam et al [12] transfer the algorithm into the distributed memory model and show how dependency chains, which are short in practice, are resolved e ciently in parallel. Alternatively if multi-edges are acceptable, [294] can be adapted to multi-GPU scenarios [11] yielding an even more scalable approach.…”

“…In that spirit, we execute EM ES on an undirected version of the crawled ClueWeb12 graph's core [324] which we obtain by deleting all nodes corresponding to uncrawled URLs. 11 Performing k = m swaps on this graph with n ≈ 9.8 • 10 8 nodes and m ≈ 3.7 • 10 10 edges is feasible in less than 19.1 h on SysB. Bhuiyan et al propose a distributed edge switching algorithm and evaluate it on a compute cluster with 64 nodes each equipped with two Intel Xeon E5-2670 2.60GHz 8-core processors and 64GB RAM [43].…”

“…In this section we evaluate the quality of the proposed algorithms and analyze the runtime of our C++ implementations. 11 EM CB, IM CB, EM GCB are designed as modules of NetworKit [316]; due to their superior performance, only the latter two were added to the library and are available since release 4.6. EM PGCB's implementation is developed separately and facilitates external memory data structures and algorithms of STXXL [102].…”

“…To achieve comparability, we removed connectivity tests, xed memory management issues, and adopted the number of swaps. Further, we extended counters for edge ids and accumulated degrees to 64 bit integers in order to support experiments with more than 2 30 edges 11. We consider such vertices atypically simple as they have degree 1 and account for ≈84 % of nodes.…”