A major challenge in next-generation genome sequencing (NGS) is to assemble massive overlapping short reads that are randomly sampled from DNA fragments. To complete assembling, one needs to finish a fundamental task in many leading assembly algorithms: counting the number of occurrences of k-mers (length-k substrings in sequences). The counting results are critical for many components in assembly (e.g. variants detection and read error correction). For large genomes, the k-mer counting task can easily consume a huge amount of memory, making it impossible for large-scale parallel assembly on commodity servers. In this paper, we develop MSPKmerCounter, a disk-based approach, to efficiently perform k-mer counting for large genomes using a small amount of memory. Our approach is based on a novel technique called Minimum Substring Partitioning (MSP). MSP breaks short reads into multiple disjoint partitions such that each partition can be loaded into memory and processed individually. By leveraging the overlaps among the k-mers derived from the same short read, MSP can achieve astonishing compression ratio so that the I/O cost can be significantly reduced. For the task of k-mer counting, MSPKmerCounter offers a very fast and memory-efficient solution. Experiment results on large real-life short reads data sets demonstrate that MSPKmerCounter can achieve better overall performance than state-of-the-art k-mer counting approaches. MSPKmerCounter is available at http://www.cs.ucsb.edu/~yangli/MSPKmerCounter