Investigate hashing methods

[object88]

Current performance for Get is between 2 and 7 times slower than native implementation for maps (depending on map size). There are many possible reasons, including the performance of the hashing mechanism.

The current code uses the built-in FNV32a hash methods, as supplied in the hash package. Using the benchmark tools, we can see that...

for i := 0; i < max; i++ {
  hasher := fnv.New32a()
  binary.Write(hasher, binary.LittleEndian, uint32(i))
  hash32 = hasher.Sum32()
}

...takes about 0.02ns / op, where max is 500,000.

Although it seems that there is little to gain here, we might find some slightly better performance by using a custom hashing mechanism.

[object88]

On further investigation, after fixing the benchmarking and performing some profiling, it looks like the hashing method may be more at fault than initially presumed. When using the default hash function (FNV32a), the hash function consumed roughly 30% of CPU time.

[object88]

First release, #36 . Very limited in scope; required custom hashing on all Key types. Need to move to something shared or more generic, reusable.

[object88]

Have been pointed to the xxhash library; may be sufficient for hashing needs.