A technique known as maps can store the values for each attribute in a compact array. The knowledge which attribute name corresponds to which index is maintained in an auxiliary data structure (the eponymous map), so you don't immediately gain a performance benefit (though it does use memory more efficiently if many objects share a set of attributes). With a JIT compiler, you can make the map persistent and constant-fold lookups, so the final machine code can use constant offsets into the attributes array (for constant attribute names).
In an interpreter (I'll assume byte code), things are much harder because you don't have much opportunity to specialize code for specific objects. However, I have an idea myself for turning attribute names into integral keys. Maintain a global mapping assigning integral IDs to attribute names. When adding new byte code to the VM (loading from disk or compiling in memory), scan for strings used as attributes, and replace them with the associated ID, creating a new ID if the string hasn't been seen before. Instead of storing hash tables or similar mappings on each object - or in the map, if you use maps - you can now use sparse arrays, which are hopefully more compact and faster to operate on.
I haven't had a change to implement and test this, and you still need a sparse array. Unless you want to make all objects (or maps) take as many words of memory as there are distinct attribute names in the whole program, that is. At least you can replace string hash tables with integer hash tables.
Just by tuning a hash table for IDs as keys, you can make several optimizations: Don't invoke a hash function (use the ID as hash), remove some indirection and hence cache misses, save yourself the complexity of dealing with pathologically bad hash functions, etc.