xpdt is (yet another) language for defining data-types and generating code for serializing and deserializing them. It aims to produce code with little or no overhead, especially in the case where some fields aren't required, and is based on fixed-length representations allowing for branch-free zero-copy deserialization and (at-most-)one-copy writes (source to buffer).
The generated C code, in particular, is highly optimized and often permits the elimination of data-copying for writes and enables optimizations such as loop-unrolling when deserializing fixed-length objects. This can lead to read speeds in excess of 500 million objects per second (~1.8 nsec per object).
The xpdt source language looks similar to C struct definitions:
struct timestamp {
u32 tv_sec;
u32 tv_nsec;
};
struct point {
i32 x;
i32 y;
i32 z;
};
struct line {
timestamp time;
point line_start;
point line_end;
bytes comment;
};
Fixed width integer types from 8 to 128 bit are supported, along with the
bytes type, which is a variable-length sequence of bytes.
The following target languages are currently supported:
- C
- Python
The C code is very highly optimized.
The Python code is about as well optimized for CPython as I can make it. It
uses typed NamedTuple for objects, which has some small overhead over regular
tuples, and it uses struct.Struct to do the packing/unpacking. I have also
code-golfed the generated bytecodes down to what I think is minimal given the
design constraints. As a result, performance of the pure Python code is
comparable to a JSON library implemented in C or Rust.
For better performance in Python, it may be desirable to develop a Cython target. In some instances CFFI structs may be more performant since they can avoid the creation/destruction of an object for each record.
Target languages are implemented purely as jinja2 templates.
The serialization format for fixed-length objects is simply a packed C struct, with little-endian fields.
For any object which contains variable length fields (eg. bytes or utf8):
- a 32bit unsigned record length is prepended to the struct, this allows efficient skipping of the whole record
- all variable-length fields are converted to
u32and contain the length, in bytes, of the data - all variable-length contents are appended after the struct in the order in which they appear
For example, following the example above, the serialization would be:
u32 tot_len # = 41
u32 time.tv_sec
u32 time.tv_usec
i32 line_start.x
i32 line_start.y
i32 line_start.z
i32 line_end.x
i32 line_end.y
i32 line_end.z
u32 comment # = 5
u8 'H'
u8 'e'
u8 'l'
u8 'l'
u8 'o'
Placing variable-length fields at the end of the struct provides significant performance benefits:
- xpdt is optimized for fast zero-copy deserialization, it especially tries to avoid adding overhead in the case where only a subset of fields are required. If variable length fields could come in the middle of a struct, reading the length and skipping the field would incur a cost even when the field is being skipped.
- Potentially large strings and payloads isolated: Only when the variable-length field is needed do you pay the cost of calculating their offsets. This is ideal for descriptive strings or optional metadata that is rarely accessed.
The feature-set is, as of now, pretty slim.
There are no array / sequence / map types, and no keyed unions.
Support for such things may be added in future provided that suitable implementations exist. An implementation is suitable if:
- It admits a zero (or close to zero) overhead implementation
- it causes no overhead when the feature isn't being used
The compiler is released under the GPLv3.
The C support code/headers are released under the MIT license.
The generated code is yours.