Colibri

ANNOUNCE: Colibri version 0.12

As part of my ongoing work on the Cloverfield project, I (FB) am pleased to announce the twelfth version of colibri.

What is Cloverfield?

Wiki page: https://wiki.tcl-lang.org/Cloverfield

What is Colibri?

Colibri is a string and data type infrastructure. It features:

  • Rope-based strings (see https://wiki.tcl-lang.org/20690 ) :
    • ropes are immutable
      ... but ...
    • extraction, insertion, deletion, concatenation... are fast
    • limited compatibility with native C strings allocated in static space (or guaranteed to survive the whole application lifetime)
    • support for the full Unicode range (up to 32-bit codepoints)
    • 1/2/4-byte fixed-width encodings as well as variable-width UTF-8 with transparent conversions and access
    • custom rope types allows for lazy or procedural representations
    • iterator interface for easy access
  • Extensible data type system dubbed "words"
    • similar to Tcl_Obj
    • minimum payload is identical to that of Tcl_Obj, i.e. 8 bytes, but can take more space if needed
    • words can have synonyms, i.e. words of any type
    • synonyms can form chains of arbitrary length, so a word can have several synonyms with different representations — no more shimmering!
    • predefined types such as ints, single chars and small strings (up to 3 chars on 32-bit systems) are represented as immediate values: the value is stored directly in the pointer rather than in an allocated structure
    • several high level datastructures are provided, such as vectors (flat arrays), lists (binary trees of vectors), and sequences (potentially cyclic and unlimited linked lists of lists and references to other sequences), along with an easy to use iterator interface
  • Fast and efficient cell-based allocator
    • page-based allocation for optimal locality of reference and cache use
    • 16-byte cells on 32-bit systems fit most needs, but elements can allocate more cells if needed (up to 63 cells on 32-bit systems)
    • overhead is small: 2 bits per 16-byte cell.
    • raw alloc performances are competitive with the stock system malloc, and in many cases much faster (up to 5 times as fast on small strings and on words vs. Tcl_Obj-like structs)
    • single cell allocation (the most frequent case) is very fast
  • Automatic memory management thanks to an exact (AKA accurate or precise), generational, copying, mark-and-sweep, garbage collector
    • exact GC implies that roots (externally referenced elements) and parent-child relations must be declared by the application, using a simple API
    • custom types can define a finalizer; one of the applications is to plug in externally managed structures (e.g. malloc/free)
    • the GC process is fully controllable (pause/resume) so that the application don't get interrupted unexpectedly
    • generational GC limits the overhead by restricting the collection to newer elements, which are typically short-lived
    • longer-living elements are collected less often as they get promoted to older generations
    • promotion is done by moving whole pages between memory pools, optionally performing compaction when fragmentation exceeds a certain threshold, thus limiting the overhead and improving the cache-friendliness over time
    • contrary to reference-counting schemes (RC), GCs allow circular references without memory leaks; word synonym chains take advantage of this, being implemented as circular lists
    • studies show that GCs consistently outperform RC in raw performances and real-world cases, because the cost (both space and time) of maintaining reference counters outweights the amortized GC overhead, especially with generational GCs

Colibri is written in plain C and is free of any dependency besides system libraries. The compiled binary DLL on Windows is about 39kB. The source code is heavily commented and follows the Tcl quality standards.

What does Colibri stand for?

Colibris, known in English as hummingbirds, are a family of birds known for their small size and high wing speed. The bee hummingbird (Mellisuga helenae), found in Cuba, is the smallest of all birds, with a mass of 1.8 g and a total length of about 5cm. They are renown for their stationary and backward flight abilities on par with flying insects, which allow them to feed on the nectar of plants and flowers in-flight.

I've chosen this name for this project because its goal is to be fast and lightweight, and to follow the feather and bird theme shared with Tcl and many related projects.

How does Colibri relate to Tcl?

From the Cloverfield announcement:

“The last major point of the project is related to implementation and low level issues. The existing Tcl implementation is notoriously solid, however some changes are too radical to be performed on the current code base, so a reimplementation is certainly needed on vast portions. For example, the string representations, object structures, and virtual machines will certainly require complete rewrite to accommodate with the needed changes, which means that a lot of code won't be reused. However vast portions of library code and algorithms certainly will (clock scan, bigint, regexp, ...), as well as the high coding standards and QA that are the trademarks of our beloved language.”

So Colibri is a candidate infrastructure for Tcl9 as an alternative to the current Tcl_Obj-based core implementation. I believe that the features provided by Colibri shall yield higher levels of performances than the current architecture, at the price of an ABI incompatibility (for which major versions are made anyway), but with a similar philosophy that should ease conversion of existing code.

Planned features for next versions

1. Mutable data types with circular references and graceful degradation to immutable types:

  • Mutable lists. Candidate models include unrolled linked lists and skip lists.
  • Maps (AKA dicts) with string keys. Candidate models include Patricia trees and Ternary Search Trees, but not hash tables.

2. Proper internal and user documentation

3. Better test suite

What needs to be done?

My main development platform is Windows, so the source archive primarily includes Microsoft Visual Studio project files. Microsoft provides a free edition of their development environment known as Visual Studio Express for those willing to compile and test the library without having to buy an expensive license. Other systems need a makefile and/or autoconf scripts.

I also uses a CentOS 5.2 Linux VMware image for Linux development, so the archive also includes minimalist GNU makefiles for building using the included GCC compiler. However it makes no use of other parts of the GNU toolchain (autoconf and the like).

The code is fairly portable on 32-bit systems. 64-bit support will need more work because all the internals are fine-tuned and optimized at the bit level; however porting to 64-bit should be rather straightforward: the algorithms will remain unchanged, structure access is abstracted behing macros, and cell size is proportional to the word size (a cell should be able to store 4 pointers, which add up to 16 bytes on 32-bit systems).

The only part that needs platform-specific code is the low-level page allocator. Colibri needs system calls that allocate boundary-aligned pages. At present both Windows and Unix (Linux) version is provided, the latter using mmap. Porting to other systems should require only minimal effort, as the platform-specific code is limited to a handful of functions gathered in a platform-specific source tree. Platform-specific peculiarities should not impact the overall architecture.

Exception handling is poor, especially because Colibri is meant to be part of a larger system that provides the appropriate infrastructure. "Exception handling" includes what to do upon allocation failures due to low-memory conditions. As Colibri uses a non-prehemptive model it cannot stop-the-world and GC like other implementations do to free memory. More generally, due to the nature of the model, such conditions would only occur when the memory is actually used or when the GC is not triggered often enough (for example during a long computation phase involving a large number of temporary allocations), both being unrecoverable anyway without explicit programming.

Tests have been run extensively during development, both on stability and performance. However Colibri lacks a real test suite (including unit testings). The source includes a test application that has to be hand-edited to run or customize specific tests. Note that some stress tests are configured in such a way that they require a system with a large memory size (2GB), so make sure to tune their parameters before running them.

Last, it lacks user and design documentation, although the source code is extensively commented.

Grand Unification Scheme

A great side project would be to reimplement Tcl over Colibri as a replacement for the current Tcl_Obj based code. Of course the resulting library would be incompatible on an ABI level, but this would provide a great testbed for Colibri in real-world cases (using pure script applications) as well as a comparison point with the official Tcl core, and will possibly open the path to Tcl9. This most likely involves a lot of work.

Where can it be found?

Wiki page: https://wiki.tcl-lang.org/Colibri
Project page @ SourceForge.net: http://sourceforge.net/projects/tcl9/
Mailing list: http://sourceforge.net/mailarchive/forum.php?forum_name=tcl9-colibri
Direct Download:

License

The license is the same as for Tcl.

Changes since version 0.3

1. Added sequence and reference word types.

  • potentially unlimited collections
  • self-referencing structures with efficient cycle detection

2. Type- and nil/NULL- related cleanup.

3. Improved source code documentation on the internal data structures.

Changes since version 0.2

1. Multithreading support.

  • apartment model, each thread has its own memory pool

2. Linux port and code portability improvements.

  • now uses C99 types such as int8_t and uintptr_t
  • moved platform-specific code to platform/* subtree
  • Linux version uses mmap and pthreads

3. Test app overhaul.

  • each test has its own C file
  • multithreading support