Cultivated alfalfa derives from a taxonomic group called the Medicago sativa-falcata complex. The complex consists of several species and subspecies that do not have any hybridization barriers. Morphological traits such as flower color, pod shape, and ploidy have traditionally being used for taxonomic classification. Cultivated alfalfa is tetraploid, but a significant amount of diversity is present among diploid germplasm. A collection of 374 individual genotypes derived from 120 unimproved diploid accessions from the National Plant Germplasm System was selected to represent the diploid M. sativa-falcata complex, including M. sativa subspecies caerulea, falcata, and hemicycla. The accessions were screened with a set of 89 polymorphic SSR loci in order to estimate genetic diversity, infer the genetic bases of current morphology-based taxonomy, and determine population structure. High levels of variation were detected. A model-based clustering analysis of the genomic data identified the morphologically defined subspecies falcata and caerulea. The hybrid nature of subspecies hemicycla has also been confirmed based on its genome composition. Subsequent hierarchical population structures indicated that two distinct subpopulations exist within subspecies caerulea and subspecies falcata. We also evaluated performance of selected genotypes for cell wall constituents, total biomass yield, and other related agronomic traits and found a high amount of genetic variation in the diploid gene pool for agronomic traits and also for cell wall constituents. Large variation was present in this material, exceeding that observed in the tetraploid alfalfa core collection. Understanding patterns of linkage disequilibrium (LD) decay in alfalfa is necessary to determine the ability of association mapping to identify quantitative trait loci of important agronomic traits. We used SSR markers and sequence polymorphism in a lignin biosynthesis gene (F5H) to infer genomewide and within gene estimates of LD. We found extensive LD among SSR markers, extending over 10 Mb. In contrast, within gene LD extends over about 200 bp and sharply declined for longer distances. These results indicate that either more markers or more candidate genes are necessary in order to identify effective marker-trait associations.