Multiadaptive Sampling for Lightweight Network Measurements

Facing the huge traffic volumes involved in today's networks it is of utmost importance to deploy efficient network measurement solutions to assist network management and traffic engineering tasks correctly, without interfering with normal network operation. Sampling techniques contribute effectively for this purpose as the amount of traffic processed is reduced, ideally without endangering the accuracy of network statistical behavior estimation. Although recent proposals of sampling techniques tend to improve the correctness of the estimation process, their underlying overhead is yet considerably when handling high traffic volumes. This paper proposes a new traffic sampling technique for performing lightweight network measurements. This technique, based on linear prediction, is multiadaptive regarding the packet sampling process, allowing to reduce significantly the amount of traffic under analysis while maintaining the representativeness of network samples for accurate network parameters' estimation. The performance evaluation of the sampling technique demonstrates the effectiveness and versatility of the proposal when considering real traces representing distinct traffic load scenarios. The statistical analysis provided evinces that the present solution outperforms classic sampling techniques, both in accuracy and amount of data involved in the measurement process.