Permalink
Switch branches/tags
Find file Copy path
a2cf928 Apr 28, 2018
1 contributor

Users who have contributed to this file

1050 lines (726 sloc) 36.7 KB

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
  • 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:

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.
  • By default Duktape uses natural alignment, which in practice translates to DUK_USE_ALIGN_BY=8. If the target can use lower alignment, consider forcing alignment to 4 or 1.
  • 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 duk-low-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_lowmem.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 duk-low-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_lowmem.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_lowmem.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.

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 quite simple pool allocator behavior. 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 "duk-low" command line tool is a variant with a simple 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/lowmem-alloc-log.txt
$ make clean duk-low
$ ./duk-low --lowmem-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 duk-low 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 has simplistic behavior:

  • Allocations are taken from smallest matching pool. Borrowing from a larger pool is allowed if the smallest matching pool is out of chunks.
  • 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 "duk-low", writes log to /tmp/lowmem-alloc-log.txt
$ ./duk-low --lowmem-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.