Efficient algorithms for collecting the statistics of large-scale IP address data

TL;DR

This paper introduces two efficient algorithms for large-scale IP address statistics collection, optimizing time and memory use by leveraging the sparse nature of IP data and dynamic hash indexing, with a parallel scheme for faster processing.

Contribution

The paper presents novel algorithms that avoid hash collisions and adapt hash index length, significantly improving efficiency in large-scale IP data analysis.

Findings

Outperforms baseline methods in time efficiency

Uses less memory than existing solutions

Supports parallel processing for faster computation

Abstract

Compiling the statistics of large-scale IP address data is an essential task in network traffic measurement. The statistical results are used to evaluate the potential impact of user behaviors on network traffic. This requires algorithms that are capable of storing and retrieving a high volume of IP addresses within time and memory constraints. In this paper, we present two efficient algorithms for collecting the statistics of large-scale IP addresses that balance time efficiency and memory consumption. The proposed solutions take into account the sparse nature of the statistics of IP addresses while building the hash function and maintain a dynamic balance among layered memory blocks. There are two layers in the first proposed method, each of which contains a limited number of memory blocks. Each memory block contains 256 elements of size $256 \times 8$ bytes for a 64-bit system. In contrast to built-in hash mapping functions, the proposed solution completely avoids expensive hash collisions while retaining the linear time complexity of hash-based solutions. Moreover, the mechanism dynamically determines the hash index length according to the range of IP addresses, and can balance the time and memory constraints. In addition, we propose an efficient parallel scheme to speed up the collection of statistics. The experimental results on several synthetic datasets show that the proposed method substantially outperforms the baselines with respect to time and memory space efficiency.