This repository contains an implementation of uint32-to-uint32 map which is ~20-50% faster than Go standard map for the same types (see benchmarks below). The code was based of Brent Pedersen's intintmap and the main logic remains intact, with some bug fixes and improvements of the API itself. The map is backed by a single array which interleaves keys and values to improve data locality.
// Create a new map with capacity of 1024 (resizeable) and 90% desired fill rate
m := intmap.New(1024, 0.90)
// Store a few key/value pairs
m.Store(1, 100)
m.Store(2, 200)
// Load them
v, ok := m.Load(1)
v, ok := m.Load(2)
// Delete keys
m.Delete(1)
m.Delete(2)Looking at the benchmarks agains the standard Go map, this map should perform roughly 20-50% better depending on the conditions.
cpu: Intel(R) Core(TM) i7-9700K CPU @ 3.60GHz
BenchmarkStore/intmap-8 14422464 81.09 ns/op 0 B/op 0 allocs/op
BenchmarkStore/sync-8 10812642 93.62 ns/op 0 B/op 0 allocs/op
BenchmarkStore/stdmap-8 10179489 114.8 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/intmap-0%-8 10622955 108.6 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/sync-0%-8 9956158 115.7 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/stdmap-0%-8 10334260 108.4 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/intmap-10%-8 10721809 105.1 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/sync-10%-8 9463400 111.8 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/stdmap-10%-8 10433720 110.7 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/intmap-50%-8 12272485 88.29 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/sync-50%-8 11352820 91.41 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/stdmap-50%-8 10265517 111.1 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/intmap-90%-8 15370150 66.75 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/sync-90%-8 17687307 69.72 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/stdmap-90%-8 10236068 112.2 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/intmap-100%-8 20562958 58.48 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/sync-100%-8 18913777 62.92 ns/op 0 B/op 0 allocs/op
BenchmarkLoad/stdmap-100%-8 9638715 109.7 ns/op 0 B/op 0 allocs/op