How Oj Just Got Faster

The original Oj parser is a performant parser that supports several modes. As of this writing Oj is almost 10 years old. A dinosaur by coding standards. It was time for an upgrade. Dealing with issues over the years it became clear that a few things could have been done better. The new Oj::Parser is a response that not only attempts to address some of the issues but also give the Oj parser a significant boost in performance. Oj::Parser takes a different approach to JSON parsing than the now legacy Oj parser. Not really a legacy parser yet since the Oj::Parser is not a drop-in replacement for the JSON gem but it is as much 3 times or more faster than the previous parser in some modes.

Address Issues

There are a few features of theOj.load parser that continue to be the reason for many of the issue on the project. The most significant area is compatibility with both Rails and the JSON gem as they battle it out for which behavior will win out in any particular situation. Most of the issues are on the writing or dumping side of the JSON packages but some are present on the parsing as well. Conversion of decimals is one area where the Rails and the JSON gem vary. The Oj::Parser addresses this by allowing for completely separate parser instances. Create a parser and configure it for the situation and leave the others parsers on their own.

The Oj::Parser is mostly compatible with the JSON gem and Rails but no claims are made that the behavior will be the same as either.

The most frequent issues that can addressed with the new parser are around the handling of options. For Oj.load there is a set of default options that can be set and the same options can be specified for each call to parse or load. This approach as a couple of downsides. One the defaults are shared across all calls to parse no matter what the desire mode is. The second is that having to provide all the options on each parse call incurs a performance penalty and is just annoying to repeat the same set of options over may calls.

By localizing options to a specific parser instance there is never any bleed over to other instances.

How

It’s wonderful to wish for a faster parser that solves all the annoyances of the previous parser but how was it done is a much more interesting question to answer.

At the core, the API for parsing was changed. Instead of a sinle global parser any number of parsers can be created and each is separate from the others. The parser itself is able to rip through a JSON string, stream, or file and then make calls to a delegate to process the JSON elements according to the delegate behavior. This is similar to the Oj.load parser but the new parser takes advantage of character maps, reduced conditional branching, and calling function pointers.

Options

As mentioned, one way to change the options issues was to change the API. Instead of having a shared set of default options a separate parser is created and configured for each use case. Options are set with methods on the parser so no more guessing what options are available. With options isolated to individual parsers there is no unintended leakage to other parse use cases.

Structure

A relative small amount of time is spent in the actual parsing of JSON in Oj.load. Most of the time is spent building the Ruby Objects. Even cutting the parsing time in half only gives a 10% improvement in performance but 10% is still an improvement.

The Oj::Parser is designed to reduce conditional branching. To do that it uses character maps for the various states that the parser goes through when parsing. There is no recursion as the JSON elements are parsed. The use of a character maps for each parser state means the parser function can and is re-entrant so partial blocks of JSON can be parsed and the results combined.

There are no Ruby calls in the parser itself. Instead delegates are used to implement the various behaviors of the parser which are currently validation (validate), callbacks (SAJ), or building Ruby objects (usual). The delegates are where all the Ruby calls and related optimizations take place.

Considering JSON file parsing, Oj.load_file is able to read a file a block at a time and the new Oj::Parser does the same. There was a change in how that is done though. Oj.load_file sets up a reader that must be called for each character. Basically a buffered reader. Oj::Parser drops down a level and uses a re-entrant parser that takes a block of bytes at a time so there is no call needed for each character but rather just iterating over the block read from the file.

Reading a block at a time also allows for an efficient second thread to be used for reading blocks. That feature is not in the first iteration of the Oj::Parser but the stage is set for it in the future. The same approach was used successfully in OjC which is where the code for the parser was taken from.

Delegates

There are three delegates; validate, SAJ, and usual.

Validate

The validate delegate is trivial in that does nothing other than let the parser complete. There are no options for the validate delegate. By not making any Ruby calls other than to start the parsing the validate delegate is no surprise that the validate delegate is the best performer.

SAJ (Simple API for JSON)

The SAJ delegate is compatible with the SAJ handlers used with Oj.saj_parse so it needs to keep track of keys for the callbacks. Two optimizations are used. The first is a reuseable key stack while the second is a string cache similar to the Ruby intern function.

When parsing a Hash (JSON object) element the key is passed to the callback function if the SAJ handler responds to the method. The key is also provided when closing an Array or Hash that is part of a parent Hash. A key stack supports this.

If the option is turned on a lookup is made and previously cached key VALUEs are used. This avoids creating the string for the key and setting the encoding on it. The cache used is a auto expanding hash implementation that is limited to strings less than 35 characters which covers most keys. Larger strings use the slower string creation approach. The use of the cache reduces object creation which save on both memory allocation and time. It is not appropriate for one time parsing of say all the keys in a dictionary but is ideally suited for loading similar JSON multiple times.

Usual

By far the more complex of the delegates is the ‘usual’ delegate. The usual delegate builds Ruby Objects when parsing JSON. It incorporates many options for configuration and makes use of a number of optimizations.

Reduce Branching

In keeping with the goal of reducing conditional branching most of the delegate options are implemented by changing a function pointer according to the option selected. For example when turning on or off :symbol_keys the function to calculate the key is changed so no decision needs to be made during parsing. Using this approach option branching happens when the option is set and not each time when parsing.

Cache

Creating Ruby Objects whether Strings, Array, or some other class is expensive. Well expensive when running at the speeds Oj runs at. One way to reduce Object creation is to cache those objects on the assumption that they will most likely be used again. This is especially true of Hash keys and Object attribute IDs. When creating Objects from a class name in the JSON a class cache saves resolving the string to a class each time. Of course there are times when caching is not preferred so caching can be turned on or off with option methods on the parser which are passed down to the delegate..

The Oj cache implementation is an auto expanding hash. When certain limits are reached the hash is expanded and rehashed. Rehashing can take some time as the number of items cached increases so there is also an option to start with a larger cache size to avoid or reduce the likelihood of a rehash.

The Oj cache has an advantage over the Ruby intern function (rb_intern()) in that several steps are needed for some cached items. As an example Object attribute IDs are created by adding an @ character prefix to a string and then converting to a ID. This is done once when inserting into the cache and after that only a lookup is needed.

Bulk Insert

The Ruby functions available for C extension functions are extensive and offer many options across the board. The bulk insert functions for both Arrays and Hashes are much faster than appending or setting functions that set one value at a time. The Array bulk insert is around 15 times faster and for Hash it is about 3 times faster.

To take advantage of the bulk inserts arrays of VALUEs are needed. With a little planning there VALUE arrays can be reused which leads into another optimization, the use of stacks.

Stacks

Parsing requires memory to keep track of values when parsing nested JSON elements. That can be done on the call stack making use of recursive calls or it can be done with a stack managed by the parser. The Oj.load method maintains a stack for Ruby object and builds the output as the parsing progresses.

Oj::Parser uses three different stacks. One stack for values, one for keys, and one for collections (Array and Hash). By postponing the creation of the collection elements the bulk insertions for Array and Hash can be used. For arrays the use of a value stack and creating the array after all elements have been identified gives a 15x improvement in array creation.

For Hash the story is a little different. The bulk insert for Hash alternates keys and values but there is a wrinkle to consider. Since Ruby Object creation is triggered by the occurrence of an element that matches a creation identifier the creation of a collection is not just for Array and Hash but also Object. Setting Object attributes uses an ID and not a VALUE. For that reason the keys should not be created as String or Symbol types as they would be ignored and the VALUE creation wasted when setting Object attributes. Using the bulk insert for Hash gives a 3x improvement for that part of the object building.

Looking at the Object creation the JSON gem expects a class method of #json_create(arg). The single argument is the Hash resulting from the parsing assuming that the parser parsed to a Hash first. This is less than ideal from a performance perspective so Oj::Parser provides an option to take that approach or to use the much more efficient approach of never creating the Hash but instead creating the Object and then setting the attributes directly.

To further improve performance and reduce the amount of memory allocations and frees the stacks are reused from one call to #parse to another.

Results

The results are even better than expected. Running the perf_parser.rb file shows the improvements. There are four comparisons all run on a MacBook Pro with Intel processor.

Validation

Without a comparible parser that just validates a JSON document the Oj.saj_parse callback parser with a nil handler is used for comparison to the new Oj::Parser.new(:validate). In that case the comparison is:

System  time (secs)  rate (ops/sec)
-------------------  -----------  --------------
Oj::Parser.validate       0.101      494369.136
       Oj::Saj.none       0.205      244122.745

The Oj::Parser.new(:validate) is 2.03 times faster!

Callback

Oj has two callback parsers. One is SCP and the other SAJ. Both are similar in that a handler is provided that implements methods for processing the various element types in a JSON document. Comparing Oj.saj_parse to Oj::Parser.new(:saj) with a all callback methods implemented handler gives the following raw results:

System  time (secs)  rate (ops/sec)
--------------  -----------  --------------
Oj::Parser.saj       0.783       63836.986
   Oj::Saj.all       1.182       42315.397

The Oj::Parser.new(:saj) is 1.51 times faster.

Parse to Ruby primitives

Parsing to Ruby primitives and Array and Hash is possible with most parsers including the JSON gem parser. The raw results comparing Oj.strict_load, Oj::Parser.new(:usual), and the JSON gem are:

System  time (secs)  rate (ops/sec)
----------------  -----------  --------------
Oj::Parser.usual       0.452      110544.876
 Oj::strict_load       0.699       71490.257
       JSON::Ext       1.009       49555.094

The Oj::Parser.new(:saj) is 1.55 times faster than Oj.load and 2.23 times faster than the JSON gem.

Object

Oj supports two modes for Object serialization and deserialization. Comparing to the JSON gem compatible mode Oj.compat_load, Oj::Parser.new(:usual), and the JSON gem yields the following raw results:

System  time (secs)  rate (ops/sec)
----------------  -----------  --------------
Oj::Parser.usual       0.071      703502.033
 Oj::compat_load       0.225      221762.927
       JSON::Ext       0.401      124638.859

The Oj::Parser.new(:saj) is 3.17 times faster than Oj.compat_load and 5.64 times faster than the JSON gem.

Summary

With a performance boost of from 1.5x to over 3x over the Oj.load parser the new Oj::Parser is a big win in the performance arena. The isolation of options is another feature that should make life easier for developers.