Empirical Evaluation of Hash Functions for Multipoint Measurements. C. Henke, C. Schmoll, T. Zseby, CCR vol. 38, No. 3, July 2008.
This paper is exactly what the title says it is: it uses 25 different hash functions, and assess the performance of these functions with respect to four criteria: performance (how long it takes to compute the hash); non-linearity (how different/correlated are the output given that the input are similar); unbiasedness (does the hash function select packets independently); representativeness of the selected subset (is the hashed output representative of the input).
Of the 25 hash functions, they show that 8 perform well under these criteria. Also, they compare non-crypto hashes and crypto hashes. For non-crypto hashes, it is possible to create an attack so that the packet of an adversary are disproportionally selected by the hash function (this is not a result of this paper, but is referenced here).
The paper is a good way to know which hash function to use for network measurement. It is also nicely written and easy to read.