partitions.txt

For sparse and uniform distributions, the algorithm performed ok, <amount to compara here> better than naive solution.
For sparse and non uniform distributions, it depends on the structure of the distribution. In general, a distribution that is continuous in segments (e.g. has gaps in between continous chunks of ids) works very well, especially when the number of gaps is approximately equal to the budget. 
If the distribution is completely non continuous, almost discrete, it tends to approximate the naive distribution.