Accelerating the XGBoost algorithm using GPU computing
Accelerating the XGBoost algorithm using GPU computing
复制标题
DOI：
10.7717/peerj-cs.127
复制
发表时间：
2017-07-24
影响因子：
3.8
通讯作者：
Frank, Eibe
中科院分区：
文献类型：
--
作者：
Mitchell, Rory;Frank, Eibe
我们在梯度增强库XGBoost中提出了一种基于cuda的决策树构建算法的实现。树形构造算法完全在图形处理单元（GPU）上执行，并且在各种数据集和设置（包括稀疏输入矩阵）下显示出高性能。单独的提升迭代是并行的，结合了两种方法。对于浅树，使用交错方法，对于深度较大的树，切换到更传统的基于基数排序的方法。我们展示了与4核i7 CPU相比，Titan x的速度提高了3到6倍，与2核Xeon CPU（24核）相比，Titan x的速度提高了1.2倍。我们表明，完全在GPU内存中处理希格斯数据集（1000万个实例，28个特征）是可能的。该算法作为XGBoost库中的插件提供，并完全支持所有XGBoost功能，包括分类、回归和排序任务。
We present a CUDA-based implementation of a decision tree construction algorithm within the gradient boosting library XGBoost. The tree construction algorithm is executed entirely on the graphics processing unit (GPU) and shows high performance with a variety of datasets and settings, including sparse input matrices. Individual boosting iterations are parallelised, combining two approaches. An interleaved approach is used for shallow trees, switching to a more conventional radix sort-based approach for larger depths. We show speedups of between 3 x and 6 x using a Titan X compared to a 4 core i7 CPU, and 1.2 x using a Titan X compared to 2 x Xeon CPUs (24 cores). We show that it is possible to process the Higgs dataset (10 million instances, 28 features) entirely within GPU memory. The algorithm is made available as a plug-in within the XGBoost library and fully supports all XGBoost features including classification, regression and ranking tasks.