GPU used : NVIDIA GTX280 (has about 30 multiprocessors each with 8 processors, frequency is 1.29 GHz)
CPU used : Intel i5D, 4 cores, frequency of 2.67 GHz.
GPU memory, bandwidth : 1 GB, 141.7GB/s
To get a more clear picture speedup calculated only after the I/O file is completed. Results that are obtained from the proposed differential (data size dependent) approach are compared with other approaches like HP_k_means (for smaller hence low-dimension data), UV_k-means , GMiner (for large data sets) and then fialy the performance is compared with CPU.
A. Small data sets (Low –dimension)
For this a data set of sizes 2 million and 4 million with varying values of “k” (number of the distinct sets/groups) and “d” …show more content…
And right for low dimens.)[1]
Finally a comparison has been made between in the way the algorithm works on the CPU and GPU architecture. The amount of data used exceeds the local memory of both the cases. Table 4: Comparison of GPU and CPU[1]
From these observations, it can be said that there was a linear relation between the dimensions and the time consumed. In the parallel implantation the number of operations performed were significantly smaller the combining of intermediate results will take a much smaller time in GPU. For working on even bigger chunks of data author has proposed future work as shown in fig 5. Fig 5: Master/slave model
IV. ANALYSIS OF PAPER AND FUTURE WORK The main highlight of the paper is suggesting method of handling data of varied dimensions in clustering applications. With clustering becoming increasingly popular in various applications, the size of data is relevant and very important concern. For example application like data mining deal with thousands of data points/sets at a time. So, it extremely necessary to improve efficiency of the processing. So, the proposed method helps in meeting the challenge of computational complexity along with efficient data handling. This can help in finding application in the low power devices like laptops, phones and other embedded …show more content…
And one of the most significant challenges in the field of computer architecture is the memory hierarchy and the corresponding data movement between different levels with varying bandwidth and access times. This paper has suggested a smart approach of moving data in the form of “tiles” and hence ensuring that in case of high dimensional data, the global memory is accessed the minimum number of times. The organization and the approach of the paper consistently keeps the GPU architecture in mind and suggest parallelization steps