I didn’t start out to make a slower JS MsgPack implementation, but that’s what I ended up doing. By taking Tim Caswell’s msgpack-js-browser and combining it with Joshua Bell’s stringencoding shim, I slowed everything down by about 15x.
The idea was simple: I just wanted to use typed arrays to aid in decoding MsgPack payloads. MsgPack has a few advantages on the server side of things. It’s generally faster to both decode and encode, and it’s a more compact format than JSON. This makes it tempting to use on the client side as well, although it’s not actually smaller over the wire than gzipped JSON in many cases. (The data set I experimented on was about the same size gzipped in both MsgPack and JSON formats.) If MsgPack decoding could be implemented on the client side to be almost as fast as JSON, then the gains in encoding speed on the server side mide make up for the slower decode times.
There are a few implementations of MsgPack in the browser. The first was slow, hard to read, and hard to hack on. They used a flash fallback to make it faster, but flash is out of vogue these days and doesn’t even ship on the Macs that so many of us developers have made our default machines.
Luckily we don’t actually need MsgPack, which opens up more options. For browsers that don’t support fancy new browser features, we can always just continue to use JSON (or XML, I guess).
So, many months ago, I started implementing a version of a msgpack encoder and decoder that would use typed arrays in order to make it much more efficient to work with binary data in JS. Instead of going through every byte and munging it from a string into a character you can work with, you can read bytes of differing sizes in a sane way using a DataView.// In MsgPack, the first byte indicates the type var type = this.view.getUint8(this.offset); // later on, you can pull out the length if it has one var length = this.view.getUint32(this.offset + 1);
This is way better than trying to do the equivalent in JS.
I quickly gave up my initial experiments when I found that decoding and encoding strings in the browser is all but impossible. Tim Caswell, however, did not give up and actually implemented UTF-8 decoding and encoding in his library. This was great, but still really slow. Around 10x slower to encode than JSON and about 4x slower to decode. You can replicate these experiments yourself by checking out my fork.
I got the bright idea at some point to make Tim’s library future compatible, so
that it would magically become faster as browsers start supporting the encoding
spec. It turned out to be really quite easy, but the results were quite
TextDecoder implemented in Joshua Bell’s
shim are an order of magnitude slower than the one-off implementations
in Tim’s original msgpack encoder/decoder. We’re talking 70x slower than JSON
to decode and 143x slower than JSON to encode.
I spent a small amount of time trying to make the shim faster, but it honestly needs totally reworked (and probably support for a lot of encodings removed) in order to get it anywhere near the performance desired.
I think we just need to wait for the spec to actually get some implementations and then run the measurements again.