Background: SNPs are the most abundant polymorphism type, and have been explored in many crop genomic studies, including rice and maize. SNP discovery in allotetraploid cotton genomes has lagged behind that of other crops due to their complexity and polyploidy. In this study, genome-wide SNPs are detected systematically using next-generation sequencing and efficient SNP genotyping methods, and used to construct a linkage map and characterize the structural variations in polyploid cotton genomes. Results: We construct an ultra-dense inter-specific genetic map comprising 4,999,048 SNP loci distributed unevenly in 26 allotetraploid cotton linkage groups and covering 4,042 cM. The map is used to order tetraploid cotton genome scaffolds for accurate assembly of G. hirsutum acc. TM-1. Recombination rates and hotspots are identified across the cotton genome by comparing the assembled draft sequence and the genetic map. Using this map, genome rearrangements and centromeric regions are identified in tetraploid cotton by combining information from the publicly-available G. raimondii genome with fluorescent in situ hybridization analysis. Conclusions: We report the genotype-by-sequencing method used to identify millions of SNPs between G. hirsutum and G. barbadense. We construct and use an ultra-dense SNP map to correct sequence mis-assemblies, merge scaffolds into pseudomolecules corresponding to chromosomes, detect genome rearrangements, and identify centromeric regions in allotetraploid cottons. We find that the centromeric retro-element sequence of tetraploid cotton derived from the D subgenome progenitor might have invaded the A subgenome centromeres after allotetrapolyploid formation. This study serves as a valuable genomic resource for genetic research and breeding of cotton.