This implements a memory backed multimap intended for use where: * your keys are integers, or can be mapped to dense range of integers, * the memory mapped file is on fast storage, like an SSD (although this is not a requirement), * you have arbitrary values of fixed size (e.g. structs, other POD types) that can be sorted, * you don't need dynamic updates of the table, * and you are likely to run out of memory of you use a traditional map or hash table, * but you can handle approximately 1 bit per record in RAM.

These may seem to be very specific, but many problems can be mapped into a dense integer set. mmmulti::map developed first as a data structure to support seqwish, which uses it to generate precise variation graphs from pairwise alignments between collections of sequences. As this multimap forms a key data processing kernel in the algorithm, it can scale to extremely large problem sizes, limited only by available disk space. Although performance is much slower than an in-memory structure, it is virtually guaranteed to be able to complete the compute.