14th ACM-IEEE International Conference On Formal Methods and Models For System Design
Indian Institute of Technology, Kanpur
November 18-20, 2016
As in previous years, The IPM HPC’s team won the honor of Memocode design contest 2016. Our team not only defended its previous years rank in performance/cost class, but also we won the performance class.
IPM-HPC Team members: Saeid Rahmani, Armin Ahmadzadeh, Omid Hajihassani, SeyedPooya Mirhosseini, and Saeid Gorgin
Institute for Research in Fundamental Sciences (IPM), Iran
Contest Problem: The MEMOCODE’16 will include a design contest, which will pose a computational challenge that participants may solve using hardware or software on FPGAs, GPUs, and CPUs. The conference will sponsor at least one prize with a monetary award for the contest winners. The 2016 challenge is K-means clustering that is an unsupervised method for clustering multidimensional data points, aiming to partition the points into “K” subgroups (clusters) that are similar. This is used in a variety of applications such as data mining, image segmentation, medical imaging, and bioinformatics.
IPM-HPC Solution: Our method makes exhaustive use of four High throughput GPU and hides memory latency. The 2016 MEMOCODE Design Contest was to efficiently compute k-means clustering on a large multidimensional data set. We performed effective optimizations involving the algorithmic structure and parallelism. We implemented our design using Intel Xeon E5 CPUs and NVIDIA GTX 980 GPUs. Our overall best result computed the solution in 106ms using four GPUs. In terms of cost normalized results, our best solution was the 2x GPU implementation, which was only 1.5x slower than the 4x GPU solution, at half the cost. The IPM team’s implementation strategy involved careful parallelization of the problem across available platforms, as well as optimization of the arithmetic required by the problem. Moreover, The solution was based on Lloyd’s algorithm.
our scientific article will be published soon.