GPU-Accelerated Computing with Gibbs Sampler for the 2PNO IRT Model
Item response theory (IRT) is a popular approach used for addressing large-scale statistical problems in psychometrics as well as in other fields. The fully Bayesian approach for estimating IRT models is usually memory and computational expensive due to the large number of iterations. This limits the use of the procedure in many applications. In an effort to overcome such restrictions, previous studies proposed to tackle the problem using massive core-based graphic processing units (GPU), and demonstrated the advantage of this approach over the message passing interface (MPI) by showing that a single GPU card could achieve a speedup of up to 50×. Given that GPU is practical, cost-effective, and convenient, this study aims to seek further improvements using a single GPU card.
KeywordsItem response theory Bayesian estimation MCMC Two-parameter IRT model High performance computing CUDA Optimization
- Baker, F. B., & Kim, S. H. (2004). Item response theory: Parameter estimation techniques. (2nd ed.). New York: Dekker.Google Scholar
- DiMarco, J., & Taufer. M. (2013). Performance impact of dynamic parallelism on different clustering algorithms. In SPIE Defense, Security, and Sensing, International Society for Optics and Photonics.Google Scholar
- Georganas, E. (2013). High performance parallel Gibbs sampling for IRT models. Poster session presented at ParLab Winter Retreat, Berkeley, USA.Google Scholar
- Harris, M. (2007). Optimizing parallel reduction in CUDA. Presentation packaged with CUDA Toolkit, NVIDIA Corporation.Google Scholar
- Hoberock, J., & Bell, N. (2010). Thrust: A parallel template library. http://thrust.github.io/.
- Karunadasa, N. P., & Ranasinghe, D. N. (2009). Accelerating high performance applications with CUDA and MPI. In 2009 International Conference on Industrial and Information Systems (ICIIS) (pp. 331–336).Google Scholar
- Kirk, D. B., & Hwu, W. W. (2013). Programming massively parallel processors: A hands-on approach (2nd ed.). Burlington, MA: Addison-Wesley.Google Scholar
- Luitjens, J. (2014). Faster parallel reduction on Kepler. Retrieved from http://devblogs.nvidia.com/parallelforall/faster-parallel-reductions-kepler/.Google Scholar
- NVIDIA. (2010). CUDA CURAND library. Santa Clara, CA: NVIDIA Corporation.Google Scholar
- Oancea, B., & Andrei, T. (2013). Developing a high performance software library with MPI and CUDA for matrix computations. Computational Methods in Social Sciences, 1(2), 1–10.Google Scholar
- Sheng, Y., & Headrick, T. C. (2007). An algorithm for implementing Gibbs sampling for 2PNO IRT models. Journal of Modern Applied Statistical Methods, 6(1), 341–349.Google Scholar
- Sheng, Y., Welling, W. S., & Meng, M. M. (2014). A GPU-based Gibbs sampler for a unidimensional IRT model. ISRN Computational Mathematics, 2014 (368149), 1–11.Google Scholar
- Tsutsumi, A., Iwata, N., Watanabe, N., de Jonge, J., Pikhart, H., Fernndez-Lpez, J. A., et al. (2009). Application of item response theory to achieve cross-cultural comparability of occupational stress measurement. International Journal of Methods in Psychiatric Research, 18(1), 58–67.CrossRefGoogle Scholar