Misc #16160
openLazy init thread local storage
Description
References PR https://github.com/ruby/ruby/pull/2295
Why?¶
The local_storage
member of execution context is lazy initialized and drives the Thread#[]
and Thread#[]=
APIs, which are Fiber local and not Thread local storage. I think the same lazy init pattern should be applied to the APIs below as well - reduces one Hash
alloc per thread created that does not use thread locals.
Lazy allocates thread local storage for the following APIs¶
-
Thread#thread_variable_get
- early returnsnil
on locals Hash not initialised -
Thread#thread_variable_set
- forces allocation of the locals Hash if not initilalised -
Thread#thread_variables
- early returns the empty array AND saves on Hash iteration if locals Hash not initialised -
Thread#thread_variable?
- early returnsfalse
on locals Hash not initialised
Other notes¶
- Moved initial implementation from
internal.h
tothread.c
local to call sites. - Preferred
defs/id.def
for thelocals
ID (seeing this pattern used more often, but not sure if that is preferred to inlinerb_intern
yet. Either way there's quite a few different conventions around IDs in the codebase at the moment and happy to help converging to a standard instead. - Maybe a flag is overkill and
NIL_P
onlocals
ivar could also work ...
Thoughts?
Updated by nobu (Nobuyoshi Nakada) over 5 years ago
I'm positive about this, except for the performance.
Do you have any numbers?
Updated by methodmissing (Lourens Naudé) over 5 years ago
nobu (Nobuyoshi Nakada) wrote:
I'm positive about this, except for the performance.
Do you have any numbers?
Apologies for the delay in replying.
Using benchmark-driver
script (running set last as it would taint the others by initializing the locals table on the thread object initialized in prelude):
prelude: |
th = Thread.new {}
benchmark:
thread_variable_get: th.thread_variable_get('foo')
thread_variables: th.thread_variables
thread_variable_p: th.thread_variable?('foo')
thread_variable_set: th.thread_variable_set('foo', 'bar')
loop_count: 1000000
lourens@CarbonX1:~/src/ruby/ruby$ /usr/local/bin/ruby --disable=gems -rrubygems -I./benchmark/lib ./benchmark/benchmark-driver/exe/benchmark-driver --executables="compare-ruby::~/src/ruby/trunk/ruby --disable=gems -I.ext/common --disable-gem" --executables="built-ruby::./miniruby -I./lib -I. -I.ext/common -r./prelude --disable-gem" -v --repeat-count=10 $HOME/src/lazy_init_thread_locals.yml
compare-ruby: ruby 2.7.0dev (2019-09-22T01:11:51Z master a0ce0b6297) [x86_64-linux]
built-ruby: ruby 2.7.0dev (2019-09-22T01:21:06Z lazy-init-thread-l.. 24463b7252) [x86_64-linux]
Calculating -------------------------------------
compare-ruby built-ruby
thread_variable_get 11.305M 33.901M i/s - 1.000M times in 0.088456s 0.029498s
thread_variables 22.765M 40.344M i/s - 1.000M times in 0.043927s 0.024787s
thread_variable_p 19.260M 20.883M i/s - 1.000M times in 0.051921s 0.047886s
thread_variable_set 8.195M 8.543M i/s - 1.000M times in 0.122030s 0.117054s
Comparison:
thread_variable_get
built-ruby: 33900806.8 i/s
compare-ruby: 11305022.7 i/s - 3.00x slower
thread_variables
built-ruby: 40344251.1 i/s
compare-ruby: 22765106.8 i/s - 1.77x slower
thread_variable_p
built-ruby: 20882884.5 i/s
compare-ruby: 19260142.8 i/s - 1.08x slower
thread_variable_set
built-ruby: 8543090.8 i/s
compare-ruby: 8194725.4 i/s - 1.04x slower
A regression on thread_variable_set
, but improvement on others.
And with memory runner (although knowing ahead of time it's just the hash, 40 bytes with array table saved):
lourens@CarbonX1:~/src/ruby/ruby$ /usr/local/bin/ruby --disable=gems -rrubygems -I./benchmark/lib ./benchmark/benchmark-driver/exe/benchmark-driver --executables="compare-ruby::~/src/ruby/trunk/ruby --disable=gems -I.ext/common --disable-gem" --executables="built-ruby::./miniruby -I./lib -I. -I.ext/common -r./prelude --disable-gem" -v --repeat-count=10 -r memory $HOME/src/lazy_init_thread_locals.yml
compare-ruby: ruby 2.7.0dev (2019-09-22T01:11:51Z master a0ce0b6297) [x86_64-linux]
built-ruby: ruby 2.7.0dev (2019-09-22T01:21:06Z lazy-init-thread-l.. 24463b7252) [x86_64-linux]
Calculating -------------------------------------
compare-ruby built-ruby
thread_variable_get 11.632M 11.528M bytes - 1.000M times
thread_variables 11.668M 11.472M bytes - 1.000M times
thread_variable_p 11.692M 11.452M bytes - 1.000M times
thread_variable_set 11.652M 11.568M bytes - 1.000M times
Comparison:
thread_variable_get
built-ruby: 11528000.0 bytes
compare-ruby: 11632000.0 bytes - 1.01x larger
thread_variables
built-ruby: 11472000.0 bytes
compare-ruby: 11668000.0 bytes - 1.02x larger
thread_variable_p
built-ruby: 11452000.0 bytes
compare-ruby: 11692000.0 bytes - 1.02x larger
thread_variable_set
built-ruby: 11568000.0 bytes
compare-ruby: 11652000.0 bytes - 1.01x larger