mirror of https://github.com/svaarala/duktape.git
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
1048 lines
37 KiB
1048 lines
37 KiB
=======================
|
|
Low memory environments
|
|
=======================
|
|
|
|
Overview
|
|
========
|
|
|
|
This document describes suggested feature options for reducing Duktape
|
|
memory usage for memory-constrained environments, which are one important
|
|
portability target for Duktape.
|
|
|
|
The default Duktape options are quite memory conservative, and significant
|
|
Ecmascript programs can be executed with e.g. 1MB of memory. Currently
|
|
realistic memory targets are roughly:
|
|
|
|
* 256-384kB system flash memory (code) and 256kB system RAM
|
|
|
|
- Duktape compiled with default options is feasible
|
|
|
|
- Duktape compiles to around 200-210kB of code (x86), so 256kB is
|
|
technically feasible but leaves little space for user bindings,
|
|
hardware initialization, communications, etc; 384kB is a more
|
|
realistic flash target
|
|
|
|
* 256-384kB system flash memory (code) and 128kB system RAM
|
|
|
|
- Duktape feature options are needed to reduce memory usage
|
|
|
|
- A custom pool-based memory allocation with manually tuned pools
|
|
may be required
|
|
|
|
- Aggressive measures like lightweight functions, 16-bit fields for
|
|
various internal structures (strings, buffers, objects), pointer
|
|
compression, external strings, etc may need to be used
|
|
|
|
* 192-256kB system flash memory (code) and 96kB system RAM
|
|
|
|
- Requires a bare metal system, possibly a custom C library, etc.
|
|
|
|
- http://pt.slideshare.net/seoyounghwang77/js-onmicrocontrollers
|
|
|
|
* 160-192kB system flash memory (code) and 64kB system RAM
|
|
|
|
- Requires a bare metal system, possibly a custom C library, etc.
|
|
|
|
- Requires use of ROM strings and objects to reduce Duktape startup
|
|
RAM usage (which drops to around 2-3kB with ROM strings/objects).
|
|
|
|
* 128kB system flash memory (code) and 32kB system RAM
|
|
|
|
- Requires the above, and removing built-in bindings like the global
|
|
Object (DUK_USE_OBJECT_BUILTIN), Array (DUK_USE_ARRAY_BUILTIN), etc
|
|
bindings. See stripped configuration examples, and add back whatever
|
|
bindings are absolutely necessary.
|
|
|
|
There are four basic goals for low memory optimization:
|
|
|
|
1. Reduce Duktape code (flash) footprint. This is currently a low priority
|
|
item because flash size doesn't seem to be a bottleneck for most users.
|
|
Techniques includes dropping optional Duktape functionality and compiler
|
|
options.
|
|
|
|
2. Reduce initial memory usage of a Duktape heap. This provides a baseline
|
|
for memory usage which won't be available for user code (technically some
|
|
memory can be reclaimed by deleting some built-ins after heap creation).
|
|
Techniques include pointer compression, external strings, ROM-based
|
|
strings and objects, and reducing built-in object and property count.
|
|
|
|
3. Minimize the growth of the Duktape heap relative to the scope and
|
|
complexity of user code, so that as large programs as possible can be
|
|
compiled and executed in a given space. Important contributing factors
|
|
include the footprint of user-defined Ecmascript and Duktape/C functions,
|
|
the size of compiled bytecode, etc. Techniques include reducing object
|
|
and property count for e.g. function objects.
|
|
|
|
4. Make remaining memory allocations as friendly as possible for the memory
|
|
allocator, especially a pool-based memory allocator. Concretely, prefer
|
|
small chunks over large contiguous allocations.
|
|
|
|
The following genconfig option file template enables most low memory related
|
|
option: ``config/examples/low_memory.yaml``. It doesn't enable pointer
|
|
compression because that always requires some application specific code.
|
|
More aggressive feature stripping examples are in
|
|
``config/examples/low_memory_strip.yaml``.
|
|
|
|
Optimizing code footprint
|
|
=========================
|
|
|
|
The best options for reducing code footprint depend obviously on your compiler.
|
|
The footprint difference between different options can be quite large, sometimes
|
|
over 20%, so for targets with code footprint limitations it's worth it to
|
|
investigate the best compiler specific options.
|
|
|
|
Gcc size optimization
|
|
---------------------
|
|
|
|
Based on: https://software.intel.com/en-us/blogs/2013/01/17/x86-gcc-code-size-optimizations.
|
|
|
|
Default ``Makefile.cmdline`` on x64 which uses ``-O2 -fomit-frame-pointer``::
|
|
|
|
text data bss dec hex filename
|
|
231549 1184 56 232789 38d55 duk
|
|
|
|
Adding ``-flto``::
|
|
|
|
text data bss dec hex filename
|
|
219825 1160 56 221041 35f71 duk
|
|
|
|
Adding ``-flto -fno-asynchronous-unwind-tables``::
|
|
|
|
text data bss dec hex filename
|
|
186745 1160 56 187961 2de39 duk
|
|
|
|
Adding ``-flto -fno-asynchronous-unwind-tables -ffunction-sections -Wl,--gc-sections``::
|
|
|
|
text data bss dec hex filename
|
|
186666 1144 56 187866 2ddda duk
|
|
|
|
Adding an explicit ``-fno-stack-protector -fno-stack-check`` may also have an effect.
|
|
|
|
Stripping unused API functions
|
|
------------------------------
|
|
|
|
If you compile and link your application and Duktape statically, you can
|
|
often strip away any Duktape API functions which are not actually used by
|
|
your application or Duktape. Doing so requires compiler specific steps,
|
|
but see for example:
|
|
|
|
- http://stackoverflow.com/questions/6687630/how-to-remove-unused-c-c-symbols-with-gcc-and-ld
|
|
|
|
Example using GCC: compile the Duktape command line utility without removing
|
|
unused API symbols::
|
|
|
|
$ gcc -o duk -Os -pedantic -std=c99 -Wall -fstrict-aliasing \
|
|
-fomit-frame-pointer -I./src src/duktape.c \
|
|
examples/cmdline/duk_cmdline.c -lm
|
|
$ size duk
|
|
text data bss dec hex filename
|
|
231079 1184 56 232319 38b7f duk
|
|
|
|
Add GCC specific options to remove unused symbols::
|
|
|
|
# With -fdata-sections -ffunction-sections -Wl,--gc-sections:
|
|
$ gcc -o duk -Os -pedantic -std=c99 -Wall -fstrict-aliasing \
|
|
-fomit-frame-pointer -I./src -fdata-sections \
|
|
-ffunction-sections -Wl,--gc-sections \
|
|
src/duktape.c examples/cmdline/duk_cmdline.c -lm
|
|
$ size duk
|
|
text data bss dec hex filename
|
|
222743 1152 48 223943 36ac7 duk
|
|
|
|
Here the difference is ~8kB on x64. For the dist Makefile.hello example,
|
|
which uses very few public API calls, the difference is ~12kB.
|
|
|
|
Miscellaneous
|
|
-------------
|
|
|
|
* On some low memory targets only libc features which are actually used get
|
|
pulled into the binary being built. In such cases it may be useful to
|
|
avoid calling platform sprintf/sscanf because they may be surprisingly
|
|
large (>20 kB). You can use ``extras/minimal-printf`` instead.
|
|
|
|
* Math functions can sometimes have a relatively large footprint (15-20 kB),
|
|
usually from trigonometric and other transcendental functions. You can
|
|
stub out unnecessary functions in ``duk_config.h``. Note, however, that
|
|
Duktape internals at present depend on a few Math functions like ``DUK_FMOD()``.
|
|
|
|
Suggested options
|
|
=================
|
|
|
|
NOTE: This list is not exhaustive, see ``config/examples/low_memory.yaml``.
|
|
|
|
* Use the default memory management settings: although reference counting
|
|
increases heap header size, it also reduces memory usage fluctuation
|
|
which is often more important than absolute footprint.
|
|
|
|
* If the target has a shallow C stack, you may want to limit C stack
|
|
recursion, see:
|
|
|
|
- ``config/examples/shallow_c_stack.yaml``
|
|
|
|
* Reduce error handling footprint with one or more of:
|
|
|
|
- ``#undef DUK_USE_AUGMENT_ERRORS``
|
|
|
|
- ``#undef DUK_USE_TRACEBACKS``
|
|
|
|
- ``#undef DUK_USE_VERBOSE_ERRORS``
|
|
|
|
- ``#undef DUK_USE_VERBOSE_EXECUTOR_ERRORS``
|
|
|
|
- ``#undef DUK_USE_PC2LINE``
|
|
|
|
* Use slower but more compact lexer algorithm (saves on code footprint):
|
|
|
|
- ``#undef DUK_USE_LEXER_SLIDING_WINDOW``
|
|
|
|
* Disable JSON fast paths (saves on code footprint):
|
|
|
|
- ``#undef DUK_USE_JSON_STRINGIFY_FASTPATH``
|
|
|
|
- ``#undef DUK_USE_JSON_QUOTESTRING_FASTPATH``
|
|
|
|
- ``#undef DUK_USE_JSON_DECSTRING_FASTPATH``
|
|
|
|
- ``#undef DUK_USE_JSON_DECNUMBER_FASTPATH``
|
|
|
|
- ``#undef DUK_USE_JSON_EATWHITE_FASTPATH``
|
|
|
|
* If you don't need Node.js Buffer and ES2015 typed array support, use:
|
|
|
|
- ``#undef DUK_USE_BUFFEROBJECT_SUPPORT``
|
|
|
|
* If you don't need the Duktape-specific additional JX/JC formats, use:
|
|
|
|
- ``#undef DUK_USE_JX``
|
|
|
|
- ``#undef DUK_USE_JC``
|
|
|
|
* Features borrowed from Ecmascript ES2015 can usually be disabled
|
|
(not exhaustive):
|
|
|
|
- ``#undef DUK_USE_ES6_OBJECT_SETPROTOTYPEOF``
|
|
|
|
- ``#undef DUK_USE_ES6_OBJECT_PROTO_PROPERTY``
|
|
|
|
- ``#undef DUK_USE_ES6_PROXY``
|
|
|
|
* If you don't need regexp support, use:
|
|
|
|
- ``#undef DUK_USE_REGEXP_SUPPORT``
|
|
|
|
* Disable unnecessary parts of the C API:
|
|
|
|
- ``#undef DUK_USE_BYTECODE_DUMP_SUPPORT``
|
|
|
|
* Duktape debug code uses a large, static temporary buffer for formatting
|
|
debug log lines. If you're running with debugging enabled, use e.g.
|
|
the following to reduce this overhead:
|
|
|
|
- ``#define DUK_USE_DEBUG_BUFSIZE 2048``
|
|
|
|
* If strict Unicode support is not critical in your application, you can:
|
|
|
|
- Strip the ``UnicodeData.txt`` and ``SpecialCasing.txt`` files manually.
|
|
There are example files in the distributable for Unicode data limited
|
|
to 8-bit codepoints.
|
|
|
|
- Provide the stripped files to ``configure.py`` to reduce Unicode
|
|
table size.
|
|
|
|
- Possible footprint savings are about 2-3kB.
|
|
|
|
More aggressive options
|
|
=======================
|
|
|
|
The following may be needed for very low memory environments (e.g. 96-128kB
|
|
system RAM):
|
|
|
|
* Consider using lightweight functions for your Duktape/C bindings and to
|
|
force Duktape built-ins to be lightweight functions:
|
|
|
|
- ``#define DUK_USE_LIGHTFUNC_BUILTINS``
|
|
|
|
* If code footprint is a significant issue, disabling reference counting
|
|
reduces code footprint by several kilobytes at the cost of more RAM
|
|
fluctuation:
|
|
|
|
- ``#undef DUK_USE_REFERENCE_COUNTING``
|
|
|
|
- ``#undef DUK_USE_DOUBLE_LINKED_LIST``
|
|
|
|
* Enable other 16-bit fields to reduce header size; these are typically
|
|
used together (all or none):
|
|
|
|
- ``#define DUK_USE_REFCOUNT16`` (and ``#undef DUK_USE_REFCOUNT32``)
|
|
|
|
- ``#define DUK_USE_STRHASH16``
|
|
|
|
- ``#define DUK_USE_STRLEN16``
|
|
|
|
- ``#define DUK_USE_BUFLEN16``
|
|
|
|
- ``#define DUK_USE_OBJSIZES16``
|
|
|
|
- ``#undef DUK_USE_HSTRING_CLEN``
|
|
|
|
* Enable heap pointer compression, assuming pointers provided by your allocator
|
|
can be packed into 16 bits:
|
|
|
|
- ``#define DUK_USE_HEAPPTR16``
|
|
|
|
- ``#define DUK_USE_HEAPPTR_ENC16(udata,p) ...``
|
|
|
|
- ``#define DUK_USE_HEAPPTR_DEC16(udata,x) ...``
|
|
|
|
- Note: you cannot currently enable Duktape debug prints (DUK_USE_DEBUG)
|
|
when heap pointer compression is enabled.
|
|
|
|
* Enable data pointer compression if possible. Note that these pointers can
|
|
point to arbitrary memory locations (outside Duktape heap) so this may not
|
|
be possible even if Duktape heap pointers can be compressed:
|
|
|
|
- ``#define DUK_USE_DATAPTR16``
|
|
|
|
- ``#define DUK_USE_DATAPTR_ENC16(udata,p) ...``
|
|
|
|
- ``#define DUK_USE_DATAPTR_DEC16(udata,x) ...``
|
|
|
|
- **UNIMPLEMENTED AT THE MOMENT**
|
|
|
|
* Enable C function pointer compression if possible. Duktape compiles to
|
|
around 200kB of code, so assuming an alignment of 4 this may only be
|
|
possible if there is less than 56kB of user code:
|
|
|
|
- ``#define DUK_USE_FUNCPTR16``
|
|
|
|
- ``#define DUK_USE_FUNCPTR_ENC16(udata,p) ...``
|
|
|
|
- ``#define DUK_USE_FUNCPTR_DEC16(udata,x) ...``
|
|
|
|
- **UNIMPLEMENTED AT THE MOMENT**
|
|
|
|
* Configure string table parameters. Often in low memory targets it's
|
|
preferable to use a fixed size, i.e. set the minimum and maximum size
|
|
to the same value. For example:
|
|
|
|
- ``#define DUK_USE_STRTAB_MINSIZE 128``
|
|
|
|
- ``#define DUK_USE_STRTAB_MAXSIZE 128``
|
|
|
|
- ``#define DUK_USE_STRTAB_PTRCOMP``
|
|
|
|
* Use "external" strings to allocate most strings from flash (there are
|
|
multiple strategies for this, see separate section):
|
|
|
|
- ``#define DUK_USE_EXTERNAL_STRINGS``
|
|
|
|
- ``#define DUK_USE_EXTSTR_INTERN_CHECK(udata,ptr,len) ...``
|
|
|
|
- ``#define DUK_USE_EXTSTR_FREE(udata,ptr) ...``
|
|
|
|
- As of Duktape 1.5 an alternative to external strings is to move strings
|
|
(including the string heap header) to ROM, see below.
|
|
|
|
* Enable struct packing in compiler options if your platform doesn't have
|
|
strict alignment requirements, e.g. on gcc/x86 you can:
|
|
|
|
- ``-fpack-struct=1`` or ``-fpack-struct=2``
|
|
|
|
The following may be appropriate when even less memory is available
|
|
(e.g. 64kB system RAM):
|
|
|
|
* Consider moving built-in strings and objects into ROM (a read-only data
|
|
section):
|
|
|
|
- ``DUK_USE_ROM_STRINGS`` and ``DUK_USE_ROM_OBJECTS`` (define both).
|
|
See: ``config/examples/rom_builtins.yaml``.
|
|
|
|
- ``DUK_USE_ROM_GLOBAL_CLONE`` or ``DUK_USE_ROM_GLOBAL_INHERIT`` if
|
|
a writable global object is needed. ``DUK_USE_ROM_GLOBAL_INHERIT``
|
|
is more memory efficient: it creates a writable (empty) global object
|
|
which inherits from the ROM global object.
|
|
|
|
- Rerun ``configure.py`` with ``--rom-support`` to create prepared
|
|
sources with support for ROM builtins. ROM builtin support is not
|
|
enabled by default because it increases the size of ``duktape.c``
|
|
considerably. Add the option ``--rom-auto-lightfunc`` to convert
|
|
built-in function properties into lightfuncs to reduce ROM footprint.
|
|
(See repo Makefile ``ajduk-rom`` target for some very simple examples.)
|
|
|
|
- Moving built-ins into ROM makes them read-only which has some side
|
|
effects. Some side effects are technical compliance issues while
|
|
others have practical impact and may prevent running some existing
|
|
scripts. The following testcases illustrate some of the issues:
|
|
|
|
+ ``tests/ecmascript/test-dev-rom-builtins-1.js``
|
|
|
|
+ ``tests/api/test-dev-rom-builtins-1.c``
|
|
|
|
- When using pointer compression you need to add support for compressing
|
|
ROM strings, see ``doc/objects-in-code-section.rst`` and a concrete
|
|
example in ``examples/cmdline/duk_cmdline_ajduk.c``.
|
|
|
|
- See ``doc/objects-in-code-section.rst`` for technical details and
|
|
current limitations.
|
|
|
|
* Consider also moving your own built-in objects and strings into ROM:
|
|
|
|
- User strings and objects can also be moved into ROM. You can also
|
|
modify default Duktape built-ins, adding and removing properties, etc.
|
|
For more details, see:
|
|
|
|
+ ``util/example_user_builtins1.yaml``: examples of user builtins
|
|
|
|
+ ``src-input/builtins.yaml``: documents some more format details
|
|
|
|
+ Repo Makefile ``ajduk-rom`` target: illustrates how to run
|
|
``configure.py`` with user builtins
|
|
|
|
* Consider using lightfuncs for representing function properties of ROM
|
|
built-ins.
|
|
|
|
- For custom built-ins you can use a "lightfunc" type for a property
|
|
value directly.
|
|
|
|
- You can also request automatic lightfunc conversion for built-in
|
|
function properties using ``--rom-auto-lightfunc``. This saves
|
|
around 15kB for Duktape built-ins.
|
|
|
|
Notes on pointer compression
|
|
============================
|
|
|
|
Pointer compression can be applied throughout (where it matters) for three
|
|
pointer types:
|
|
|
|
* Compressed 16-bit Duktape heap pointers, assuming Duktape heap pointers
|
|
can fit into 16 bits, e.g. max 256kB memory pool with 4-byte alignment
|
|
|
|
* Compressed 16-bit function pointers, assuming C function pointers can
|
|
fit into 16 bits
|
|
|
|
* Compressed 16-bit non-Duktape-heap data pointers, assuming C data
|
|
pointers can fit into 16 bits
|
|
|
|
Pointer compression can be quite slow because often memory mappings are not
|
|
linear, so the required operations are not trivial. NULL also needs specific
|
|
handling.
|
|
|
|
When ROM object/string support is enabled, pointer compression and
|
|
decompression must support ROM pointer compression. This is done by
|
|
reserving a range of 16-bit compressed pointer values to represent
|
|
ROM pointers, and to use a ROM pointer table to compress/decompress
|
|
ROM pointers. See ``examples/cmdline/duk_cmdline_ajduk.c`` for an
|
|
example.
|
|
|
|
External string strategies (DUK_USE_EXTSTR_INTERN_CHECK)
|
|
========================================================
|
|
|
|
The feature can be used in two basic ways:
|
|
|
|
* You can anticipate a set of common strings, perhaps extracted by parsing
|
|
source code, and build them statically into your program. The strings will
|
|
then be available in the "text" section of your program. This works well
|
|
if the set of common strings can be estimated well, e.g. if the program
|
|
code you will run is mostly known in advance.
|
|
|
|
* You can write strings to memory mapped flash when the hook is called.
|
|
This is less portable but can be effective when the program you will run
|
|
is not known in advance.
|
|
|
|
Note that:
|
|
|
|
* Using an external string pointer for short strings (e.g. 3 chars or less)
|
|
is counterproductive because the external pointer takes more room than the
|
|
character data.
|
|
|
|
The Duktape built-in strings are available from prepared source metadata:
|
|
|
|
* For example, ``dist/src/duk_source_meta.json``, the
|
|
``builtin_strings_base64`` contains the byte exact strings used, encoded
|
|
with base-64.
|
|
|
|
Strings used by application C and Ecmascript code can be extracted with
|
|
various methods. The Duktape main repo contains an example script for
|
|
scraping strings from C and Ecmascript code using regexps:
|
|
|
|
* ``util/scan_strings.py``
|
|
|
|
There are concrete examples for some external string strategies in:
|
|
|
|
* ``dist/examples/cmdline/duk_cmdline_ajduk.c``
|
|
|
|
Tuning pool sizes for a pool-based memory allocator
|
|
===================================================
|
|
|
|
The memory allocations used by Duktape depend on the architecture and
|
|
especially the low memory options used. So, the safest approach is to
|
|
select the options you want to use and then measure actual allocation
|
|
patterns of various programs.
|
|
|
|
The memory allocations needed by Duktape fall into two basic categories:
|
|
|
|
* A lot of small allocations (roughly between 16 and 128 bytes) are needed
|
|
for strings, buffers, objects, object property tables, etc. These
|
|
allocation sizes constitute most of the allocation activity, i.e. allocs,
|
|
reallocs, and frees. There's a lot churn (memory being allocated and
|
|
freed) even when memory usage is nearly constant.
|
|
|
|
* Much fewer larger allocations with much less activity are needed for
|
|
Ecmascript function bytecode, large strings and buffers, value stacks,
|
|
the global string table, and the Duktape heap object.
|
|
|
|
The ``examples/alloc-logging`` memory allocator can be used to write out
|
|
an allocation log file. The log file contains every alloc, realloc, and
|
|
free, and will record both new and old sizes for realloc. This allows you
|
|
to replay the allocation sequence so that you can simulate the behavior of
|
|
pool sizes offline.
|
|
|
|
The ``examples/allog-logging/pool_simulator.py`` simulates pool allocator
|
|
behavior for a given allocation log, and provides a lot of detailed graphs
|
|
of pool usage, allocated bytes, waste bytes, etc. It also provides some
|
|
tools to optimize pool counts for one or multiple application "profiles".
|
|
See detailed description below.
|
|
|
|
You can also get a dump of Duktape's internal struct sizes by enabling debug
|
|
prints; Duktape will debug print struct sizes when a heap is created. The
|
|
struct sizes will give away the minimum size needed by strings, buffers,
|
|
objects, etc. They will also give you ``sizeof(duk_heap)`` which is a large
|
|
allocation that you should handle explicitly in pool tuning.
|
|
|
|
Finally, you can look at existing projects and what kind of pool tuning
|
|
they do. AllJoyn.js has a manually tuned pool allocator which may be a
|
|
useful starting point:
|
|
|
|
* https://git.allseenalliance.org/cgit/core/alljoyn-js.git/
|
|
|
|
Tuning pool sizes using pool_simulator.py
|
|
=========================================
|
|
|
|
Overview
|
|
--------
|
|
|
|
The pool simulator replays allocation logs, simulates the behavior of a
|
|
pool-based memory allocator, and provides several useful commands:
|
|
|
|
* Replay an allocation log and provide statistics and graphs for the pool
|
|
performance: used bytes, wasted bytes, by-pool breakdowns, etc.
|
|
|
|
* Optimize pool counts based on a high-water-mark measurement, when given
|
|
pool byte sizes (a base pool configuration) and an allocation log.
|
|
|
|
* Optimize pool counts based on a more complex algorithm which takes pool
|
|
borrowing into account (see discussion below).
|
|
|
|
* Generate a pool configuration for a given total memory target, given the
|
|
tight pool configuration for Duktape and a set of representative
|
|
applications.
|
|
|
|
These operations are discussed in more detail below.
|
|
|
|
Important notes
|
|
---------------
|
|
|
|
* Before optimizing pools, you should select Duktape feature options
|
|
(especially low memory options) carefully.
|
|
|
|
* It may be useful to use DUK_USE_GC_TORTURE to ensure that there is no
|
|
slack in memory allocations; reference counting frees unreachable values
|
|
but does not handle loops. When GC torture is enabled, Duktape will run
|
|
a mark-and-sweep for every memory allocation. High-water-mark values
|
|
will then reflect the memory usage achievable in an emergency garbage
|
|
collect.
|
|
|
|
* The pool simulator provides pool allocator behavior matching AllJoyn.js's
|
|
ajs_heap.c allocator. If your pool allocator has different basic features
|
|
(for example, splitting and merging of chunks) you'll need to tweak the
|
|
pool simulator to get useful results.
|
|
|
|
Basics
|
|
------
|
|
|
|
The Duktape command line tool writes out an allocation log when requested::
|
|
|
|
# Log written to /tmp/duk-alloc-log.txt
|
|
$ make clean duk
|
|
$ ./duk --alloc-logging tests/ecmascript/test-dev-mandel2-func.js
|
|
|
|
The "ajduk" command line tool is a variant with AllJoyn.js pool allocator,
|
|
and a host of low memory optimizations. It represents a low memory target
|
|
quite well and it can also be requested to write out an allocation log::
|
|
|
|
# Log written to /tmp/ajduk-alloc-log.txt
|
|
$ make clean ajduk
|
|
$ ./ajduk --ajsheap-log tests/ecmascript/test-dev-mandel2-func.js
|
|
|
|
Allocation logs are represented in examples/alloc-logging format::
|
|
|
|
...
|
|
A 0xf7541c38 16
|
|
R 0xf754128c -1 0xf754125c 6
|
|
A 0xf7541c24 16
|
|
...
|
|
|
|
The pool simulator doesn't need to know the "previous size" for a realloc
|
|
entry, so it can be written out as -1 (like ajduk does).
|
|
|
|
Pool configurations are expressed in JSON::
|
|
|
|
{
|
|
"pools": [
|
|
{ "size": 8, "count": 10, "borrow": true },
|
|
{ "size": 12, "count": 10, "borrow": true },
|
|
{ "size": 16, "count": 200, "borrow": true },
|
|
...
|
|
]
|
|
}
|
|
|
|
The "size" (entry size, byte size) of a pool is the byte-size of individual
|
|
chunks in that pool. The "count" (entry count) is the number of chunks
|
|
preallocated for that pool. Above, the second pool has entry size of 12
|
|
bytes and a count of 10, for a total of 120 bytes.
|
|
|
|
The pool simulator matches AllJoyn.js ajs_heap.c behavior:
|
|
|
|
* Allocations are taken from smallest matching pool. Borrowing is enabled
|
|
or disabled for each pool individually.
|
|
|
|
* Reallocation tries to shrink the allocation to a previous pool size if
|
|
possible.
|
|
|
|
"High-water-mark" (hwm) over an entire allocation log means simulating the
|
|
allocation log against a certain pool configuration, and recording the
|
|
highest number of used entries for each pool. There are two variants for
|
|
this measurement:
|
|
|
|
* Without borrowing: ignore the "count" for each pool in the configuration
|
|
and autoextend the pool as needed. This provides a high-water-mark
|
|
without a need to borrow from larger pools.
|
|
|
|
* With borrowing: respect the "count" in the pool configuration and borrow
|
|
as needed.
|
|
|
|
Tight pool counts using high water mark (hwm)
|
|
---------------------------------------------
|
|
|
|
To find out the high water mark for each pool size without borrowing::
|
|
|
|
$ rm -rf /tmp/out; mkdir /tmp/out
|
|
$ python examples/alloc-logging/pool_simulator.py \
|
|
--out-dir /tmp/out \
|
|
--alloc-log /tmp/duk-alloc-log.txt \
|
|
--pool-config examples/alloc-logging/pool_config_1.json \
|
|
--out-pool-config /tmp/tight_noborrow.json \
|
|
tight_counts_noborrow
|
|
|
|
The hwm records the maximum count for each pool size::
|
|
|
|
^ pool entry count
|
|
|
|
|
| ##
|
|
| #####
|
|
| ######
|
|
| ######
|
|
| ########
|
|
+---------> pool entry size
|
|
|
|
As described above, this command ignores the pool counts in the pool config
|
|
and autoextends each pool to find its hwm. The resulting pool configuration
|
|
with updated counts is written out.
|
|
|
|
Tight pool counts taking borrowing into account
|
|
-----------------------------------------------
|
|
|
|
The high water marks for each pool entry size don't necessarily happen
|
|
at the same time. Let's use the example above::
|
|
|
|
^ pool entry count
|
|
|
|
|
| ##
|
|
| #####
|
|
| ######
|
|
| ######
|
|
| ########
|
|
+---------> pool entry size
|
|
|
|
As an example, when the hwm for the third pool size (highlighted below)
|
|
happens, the allocation state might be::
|
|
|
|
^ pool entry count
|
|
|
|
|
| #
|
|
| :#
|
|
| ::#::
|
|
| ::#:::
|
|
| ::#:::::
|
|
+---------> pool entry size
|
|
|
|
This means that we can often reduce the hwm-based pool counts and still
|
|
allow the application to run; the application will be able to borrow
|
|
allocations from larger pool entry sizes.
|
|
|
|
As an extreme example, if Duktape were to allocate and free one entry
|
|
from each pool entry size (but so that only one allocation would be
|
|
active at a time), the hwm counts would look like::
|
|
|
|
^ pool entry count
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| ########
|
|
+---------> pool entry size
|
|
|
|
However, the allocations can all be satisfied by having just one pool
|
|
entry of the largest allocated size: all other allocation requests
|
|
will just borrow from that (assuming borrowing is allowed)::
|
|
|
|
^ pool entry count
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| #
|
|
+---------> pool entry size
|
|
|
|
The pool simulator optimizes for tight pool counts with borrowing effects
|
|
taken into account using a pretty simple brute force algorithm:
|
|
|
|
* Get the basic hwm profile with no borrowing.
|
|
|
|
* Start from the largest pool entry size and loop downwards:
|
|
|
|
- Reduce pool entry count for that pool entry size in question and rerun
|
|
the allocation log.
|
|
|
|
- If allocation requests can be still satisfied through borrowing, continue
|
|
to reduce the allocation.
|
|
|
|
- When the pool entry count can no longer be reduced, move on to the next
|
|
pool size.
|
|
|
|
The basic observation in the algorithm is as follows:
|
|
|
|
* The pool entry counts above the current one are optimal: they can't be
|
|
reduced further.
|
|
|
|
* The pool entry counts below the current one never borrow from any of the
|
|
higher pool counts (yet) because they were optimized for their hwm.
|
|
|
|
* We reduce the current pool entry count, hoping that some of the allocations
|
|
needed for its hwm can be borrowed from the larger pool entry sizes. This
|
|
is possible if the hwm of the current pool entry size doesn't coincide with
|
|
the hwm of the larger pool entry sizes.
|
|
|
|
This algorithm leads to reasonable pool entry counts, but:
|
|
|
|
* The counts may not be an optimal balance for other applications.
|
|
|
|
* The pool entry sizes are assumed to be given and are not optimized for
|
|
automatically.
|
|
|
|
Use the following command to run the optimization::
|
|
|
|
$ rm -rf /tmp/out; mkdir /tmp/out
|
|
$ python examples/alloc-logging/pool_simulator.py \
|
|
--out-dir /tmp/out \
|
|
--alloc-log /tmp/duk-alloc-log.txt \
|
|
--pool-config examples/alloc-logging/pool_config_1.json \
|
|
--out-pool-config /tmp/tight_borrow.json \
|
|
tight_counts_borrow
|
|
|
|
This may take a lot of time, so be patient.
|
|
|
|
As a concrete example, for test-dev-mandel2-func.js on x86 with low memory
|
|
optimizations, the tight pool configuration based on hwm is::
|
|
|
|
total 31564:
|
|
8=91 12=25 16=373 20=56 24=2 28=58 32=1 40=32 48=4 52=27 56=1 60=5 64=0
|
|
128=20 256=9 512=8 1024=4 1360=1 2048=2 4096=0 8192=0 16384=0 32768=0
|
|
|
|
and after borrow optimization::
|
|
|
|
total 28532:
|
|
8=91 12=20 16=370 20=53 24=2 28=58 32=0 40=10 48=3 52=26 56=1 60=4 64=0
|
|
128=16 256=8 512=8 1024=3 1360=1 2048=2 4096=0 8192=0 16384=0 32768=0
|
|
|
|
The more dynamic an application's memory usage is, the more potential there
|
|
is for borrowing.
|
|
|
|
Optimizing for multiple application profiles
|
|
--------------------------------------------
|
|
|
|
Run hello world with alloc logging for Duktape baseline::
|
|
|
|
# Using "duk", writes log to /tmp/duk-alloc-log.txt
|
|
$ ./duk --alloc-logging tests/ecmascript/test-dev-hello-world.js
|
|
|
|
# Using "ajduk", writes log to /tmp/ajduk-alloc-log.txt
|
|
$ ./ajduk --ajsheap-log tests/ecmascript/test-dev-hello-world.js
|
|
|
|
Extract a "tight" pool configuration for the hello world baseline,
|
|
pool entry sizes (but not counts) need to be known in advance::
|
|
|
|
$ rm -rf /tmp/out; mkdir /tmp/out
|
|
$ python examples/alloc-logging/pool_simulator.py \
|
|
--out-dir /tmp/out \
|
|
--alloc-log /tmp/duk-alloc-log.txt \
|
|
--pool-config examples/alloc-logging/pool_config1.json \
|
|
--out-pool-config /tmp/config_tight_duktape.json \
|
|
tight_counts_borrow
|
|
|
|
Run multiple test applications and extract tight pool configurations for
|
|
each (includes Duktape baseline but that is subtracted later) using the
|
|
same method::
|
|
|
|
$ ./duk --alloc-logging tests/ecmascript/test-dev-mandel2-func.js
|
|
$ rm -rf /tmp/out; mkdir /tmp/out
|
|
$ python examples/alloc-logging/pool_simulator.py \
|
|
--out-dir /tmp/out \
|
|
--alloc-log /tmp/duk-alloc-log.txt \
|
|
--pool-config examples/alloc-logging/pool_config1.json \
|
|
--out-pool-config /tmp/config_tight_app1.json \
|
|
tight_counts_borrow
|
|
|
|
$ ./duk --alloc-logging tests/ecmascript/test-bi-array-proto-push.js
|
|
$ rm -rf /tmp/out; mkdir /tmp/out
|
|
$ python examples/alloc-logging/pool_simulator.py \
|
|
--out-dir /tmp/out \
|
|
--alloc-log /tmp/duk-alloc-log.txt \
|
|
--pool-config examples/alloc-logging/pool_config1.json \
|
|
--out-pool-config /tmp/config_tight_app2.json \
|
|
tight_counts_borrow
|
|
|
|
# ...
|
|
|
|
Select a target memory amount (here 200kB) and optimize pool entry
|
|
counts for that amount::
|
|
|
|
$ python examples/alloc-logging/pool_simulator.py \
|
|
--out-pool-config /tmp/config_200kb.json \
|
|
--out-ajsheap-config /tmp/ajsheap_200kb.c \
|
|
pool_counts_for_memory \
|
|
204800 \
|
|
/tmp/config_tight_duktape.json \
|
|
/tmp/config_tight_app1.json \
|
|
/tmp/config_tight_app2.json \
|
|
... \
|
|
/tmp/config_tight_appN.json
|
|
|
|
# /tmp/config_200kb.json is the pool config in JSON
|
|
|
|
# /tmp/ajsheap_200kb.c is the pool config as an ajs_heap.c initializer
|
|
|
|
The optimization algorithm is based on the following basic idea:
|
|
|
|
* Pool entry byte sizes are kept fixed throughout the process.
|
|
|
|
* Application pool counts are normalized by subtracting Duktape baseline
|
|
pool counts, yielding application memory usage on top of Duktape. These
|
|
pool counts can be scaled meaningfully to estimate memory demand if the
|
|
"application size" (function count, statement count, etc) were to grow
|
|
or shrink.
|
|
|
|
* The resulting pool count profiles are normalized to a fixed total memory
|
|
usage (any value will do, 1MB is used now). The resulting pool counts
|
|
are fractional.
|
|
|
|
* A pool count profile representing all the applications is computed as
|
|
follows. For each pool entry size, take the maximum of the normalized,
|
|
scaled pool counts over the application profiles. This profile represents
|
|
the the memory usage of a mix of applications.
|
|
|
|
* Allocate pool counts for Duktape baseline. This allocation is independent
|
|
of application code and doesn't grow in relation to application memory
|
|
usage.
|
|
|
|
* Scale the representative pool count profile to fit the remaining memory,
|
|
using fractional counts.
|
|
|
|
* Round pool counts into integers, ensuring the total memory usage is as
|
|
close to the target (without exceeding it).
|
|
|
|
Summary of potential measures
|
|
=============================
|
|
|
|
Heap headers
|
|
------------
|
|
|
|
* Compressed 16-bit heap pointers
|
|
|
|
* 16-bit field for refcount
|
|
|
|
* Move one struct specific field (e.g. 16-bit string length) into the unused
|
|
bits of the ``duk_heaphdr`` 32-bit flags field
|
|
|
|
Objects
|
|
-------
|
|
|
|
* Tweak growth factors to keep objects always or nearly always compact
|
|
|
|
* 16-bit field for property count, array size, etc.
|
|
|
|
* Drop hash part entirely: it's rarely needed in low memory environments
|
|
and hash part size won't need to be tracked
|
|
|
|
* Compressed pointers
|
|
|
|
Strings
|
|
-------
|
|
|
|
* Use an indirect string type which stores string data behind a pointer
|
|
(same as dynamic buffer); allow user code to indicate which C strings
|
|
are immutable and can be used in this way
|
|
|
|
* Allow user code to move a string to e.g. memory-mapped flash when it
|
|
is interned or when the compiler interns its constants (this is referred
|
|
to as "static strings" or "external strings")
|
|
|
|
* Memory map built-in strings (about 2kB bit packed) directly from flash
|
|
|
|
* 16-bit fields for string char and byte length
|
|
|
|
* 16-bit string hash
|
|
|
|
* Compressed pointers
|
|
|
|
Duktape/C function footprint
|
|
----------------------------
|
|
|
|
* Lightweight functions, converting built-ins into lightweight functions
|
|
|
|
* Lightweight functions for user Duktape/C binding functions
|
|
|
|
* Magic value to share native code cheaply for multiple function objects
|
|
|
|
* Compressed pointers
|
|
|
|
Ecmascript function footprint
|
|
-----------------------------
|
|
|
|
* Motivation
|
|
|
|
- Small lexically nested callbacks are often used in Ecmascript code,
|
|
so it's important to keep their size small
|
|
|
|
* Reduce property count:
|
|
|
|
- _pc2line: can be dropped, lose line numbers in tracebacks
|
|
|
|
- _formals: can be dropped for most functions (affects debugging)
|
|
|
|
- _varmap: can be dropped for most functions (affects debugging)
|
|
|
|
* Reduce compile-time maximum alloc size for bytecode: currently each
|
|
instruction takes 8 bytes, 4 bytes for the instruction itself and 4 bytes
|
|
for line number. Change this into two allocations so that the maximum
|
|
allocation size is not double that of final bytecode, as that is awkward
|
|
for pool allocators.
|
|
|
|
* Improve property format, e.g. ``_formals`` is now a regular array which
|
|
is quite wasteful; it could be converted to a ``\xFF`` separated string
|
|
for instance.
|
|
|
|
* Spawn ``.prototype`` on demand to eliminate one unnecessary object per
|
|
function
|
|
|
|
* Use virtual properties when possible, e.g. if ``nargs`` equals desired
|
|
``length``, use virtual property for it (either non-writable or create
|
|
concrete property when written)
|
|
|
|
* Write bytecode and pc2line to flash during compilation
|
|
|
|
* Compressed pointers
|
|
|
|
Contiguous allocations
|
|
----------------------
|
|
|
|
Unbounded contiguous allocations are a problem for pool allocators. There
|
|
are at least the following sources for these:
|
|
|
|
* Large user strings and buffers. Not much can be done about these without
|
|
a full rework of the Duktape C programming model (which assumes string and
|
|
buffer data is available as plain ``const char *``).
|
|
|
|
* Bytecode/const buffer for long Ecmascript functions:
|
|
|
|
- Bytecode and constants can be placed in separate buffers.
|
|
|
|
- Bytecode could be "segmented" so that bytecode would be stored in chunks
|
|
(e.g. 64 opcodes = 256 bytes). An explicit JUMP to jump from page to page
|
|
could make the executor impact minimal.
|
|
|
|
- During compilation Duktape uses a single buffer to track bytecode
|
|
instructions and their line numbers. This takes 8 bytes per instruction
|
|
while the final bytecode takes 4 bytes per instruction. This is easy to
|
|
fix by using two separate buffers.
|
|
|
|
* Value stacks of Duktape threads. Start from 1kB and grow without
|
|
(practical) bound depending on call nesting.
|
|
|
|
* Catch and call stacks of Duktape threads. Also contiguous but since these
|
|
are much smaller, they're unlikely to be a problem before the value stack
|
|
becomes one.
|
|
|
|
Notes on function memory footprint
|
|
==================================
|
|
|
|
Normal function representation
|
|
------------------------------
|
|
|
|
In Duktape 1.0.0 functions are represented as:
|
|
|
|
* A ``duk_hcompfuncn`` (a superset of ``duk_hobject``): represents an
|
|
Ecmascript function which may have a set of properties, and points to
|
|
the function's data area (bytecode, constants, inner function refs).
|
|
|
|
* A ``duk_hnatfunc`` (a superset of ``duk_hobject``): represents a
|
|
Duktape/C function which may also have a set of properties. A pointer
|
|
to the C function is inside the ``duk_hnatfunc`` structure.
|
|
|
|
In Duktape 1.1.0 a lightfunc type is available:
|
|
|
|
* A lightfunc is an 8-byte ``duk_tval`` with no heap allocations, and
|
|
provides a cheap way to represent many Duktape/C functions.
|
|
|
|
RAM footprints for each type are discussed below.
|
|
|
|
Ecmascript functions
|
|
--------------------
|
|
|
|
An ordinary Ecmascript function takes around 300-500 bytes of RAM. There are
|
|
three objects involved:
|
|
|
|
- a function template
|
|
- a function instance (multiple instances can be created from one template)
|
|
- automatic prototype object allocated for the function instance
|
|
|
|
The function template is used to instantiate a function. The resulting
|
|
function is not dependent on the template after creation, so that the
|
|
template can be garbage collected. However, the template often remains
|
|
reachable in callback style programming, through the enclosing function's
|
|
inner function templates table.
|
|
|
|
The function instance contains a ``.prototype`` property while the prototype
|
|
contains a ``.constructor`` property, so that both functions require a
|
|
property table. This is the case even for the majority of user functions
|
|
which will never be used as constructors; built-in functions are oddly exempt
|
|
from having an automatic prototype.
|
|
|
|
Duktape/C functions
|
|
-------------------
|
|
|
|
A Duktape/C function takes about 70-80 bytes of RAM. Unlike Ecmascript
|
|
functions, Duktape/C function are already stripped of unnecessary properties
|
|
and don't have an automatic prototype object.
|
|
|
|
Even so, there are close to 200 built-in functions, so the footprint of
|
|
the ``duk_hnatfunc`` objects is around 14-16kB, not taking into account
|
|
allocator overhead.
|
|
|
|
Duktape/C lightfuncs
|
|
--------------------
|
|
|
|
Lightfuncs require only a ``duk_tval``, 8 bytes. There are no additional heap
|
|
allocations.
|
|
|