Packet sampling is needed to measure network traffic scalably at high speeds. While many sampled-based measurement techniques have been developed in the recent past, most approaches select packets uniformly, without regard to their size. We discuss that this behavior can negatively impact the performance of tools, and present a sampling scheme that, taking into account packet sizes, overcomes several weaknesses of classic sampling methods. While the idea behind of our approach is conceptually not new, the way we propose to implement it yields a cost-effective solution that is suitable at very high speeds. To see the advantages of the approach, we study the problem of estimating traffic volumes. Our analysis and experimental evaluation with real traffic traces show that sampling considering packet sizes can improve the quality of measurements and make their accuracy less dependant on the properties of the traffic, at a small additional overhead compared to traditional methods.