Frequent Items Mining Acceleration Exploiting Fast Parallel Sorting on the GPU

Abstract

In this paper, we show how to employ Graphics Processing Units (GPUs) to provide an effcient and highperformance solution for finding frequent items in data streams. We discuss several design alternatives and present an implementation that exploits the great capability of graphics processors in parallel sorting. We provide an exhaustive evaluation of performances, quality results and several design trade-offs. Onanoff-the-shelf GPU, the fastest of our implementations can process over 200 million items per second, which is better than the best known solution based on Field Programmable Gate Arrays (FPGAs) and CPUs. Moreover, in previous approaches, performances are directly related to the skewness of the input data distribution, while in our approach, the high throughput is independent from this factor

An efficient framework for parallel and continuous frequent item monitoring

“…The hardware‐based methods can achieve very high performance, for example, the FPGA‐based approach can process more than 80 million items per second . The main disadvantage is that the hardware‐based solutions are often more expensive and complex than the CPU‐based solutions.Porting frequent item mining algorithms to special purpose processors . The basic idea is to exploit the great parallel processing capabilities of special purpose processors, such as the graphic processing unit (GPU) , network processing unit (NPU) , and imagine stream processor , to accelerate frequent item mining.…”

“…(2) Porting frequent item mining algorithms to special purpose processors [26][27][28][29][30]. The basic idea is to exploit the great parallel processing capabilities of special purpose processors, such as the graphic processing unit (GPU) [28,29], network processing unit (NPU) [26,27], and imagine stream processor [30], to accelerate frequent item mining. Govindaraju et al [29] used the GPU as a general co-processor for frequency estimation and achieved a double throughput (about five million items per second) compared with the CPU-based implementation.…”

“…Govindaraju et al [29] used the GPU as a general co-processor for frequency estimation and achieved a double throughput (about five million items per second) compared with the CPU-based implementation. Similarly, Erra and Frola [28] used radix sorting on the GPU to compute frequent items and achieved better performance (over 200 million items per second) compared with the solutions based on FPGAs and CPUs. Both Bandi et al [26] and Das et al [27] leveraged the constant time lookups of content addressable memories in NPUs to accelerate frequent item mining and achieved better performance than the CPU-based solutions.…”

“…As far as we know, all the related work [24][25][26][27][28][29][30][31][32][33][34][35] have only dealt with the traditional unweighed items, whereas this work explores ways to improve the performance of weighted frequent item mining (because the IP packet length in network monitoring is not fixed, usually ranging from 46 to 1500 bytes). Moreover, most published studies focus on using specialized hardware circuit (e.g., FPGA [24,25]) or special processors with high parallel processing capabilities (e.g., GPU [28,29], NPU [26,27] or imagine stream processor [30]), whereas this study tries to leverage the hardware advances in traditional CPU, that is, multi-core processors. Currently, there has been some research [31][32][33][34][35] on utilizing the multi-core processors in frequent item mining as well; however, they all suffer from some serious limitations.…”

Fast and accurate mining of correlated heavy hitters

“…Novel shared-memory parallel algorithms for frequent items were recently proposed in [28]. Accelerator based algorithms for frequent items exploiting a GPU (Graphics Processing Unit) include [14,16].…”