[ruby-core:72069] [Ruby trunk - Feature #11405] [PATCH] hash.c: minor speedups to int/fixnum keys

Issue #11405 has been updated by Yusuke Endoh.


Eric Wong wrote:
>  Perhaps teach developers to rehash explicitly.  Maybe Hash#freeze
>  can imply rehash, too.  But yes, I worry about memory increase more
>  than speed nowadays.

I'm not sure if explicit/implicit rehasing will solve any issue.

The problem is that the linked lists are too(?) long.  Long linked lists decrease the average access speed, and also makes per-element access speed non-uniform.
Rehasing will not improve the average speed, nor eliminate/relax the inequality.  It will just make us impossible to predict which element is fast to access and which is slow.  Will this make us happy?


Yura Sokolov wrote:
> Why not decrease rehashing threshold?
> 1 is very good choice :)
> Yeah, it will increase memory consumption a bit.
> ok, let it be 2

Yes, it will solve the issue.  But I'm unsure if we need to fix this issue.  Accessing elements takes just some nano seconds.  Is it a bottleneck in a real-life use-case?
And, to find an appropreate threshold, we need to perform exhaustive benchmark with some real-life applications.  Hard work!

-- 
Yusuke Endoh <[e-mail]>

----------------------------------------
Feature #11405: [PATCH] hash.c: minor speedups to int/fixnum keys
https://bugs.ruby-lang.org/issues/11405#change-55478

* Author: Eric Wong
* Status: Closed
* Priority: Normal
* Assignee: Eric Wong
----------------------------------------
Noticed with [ruby-core:70159] [Bug #11396]

The low bits of Ruby object IDs are rarely populated in the current
implementation, so ensure the get used.

Early versions of this patch redundantly shifted static symbols in
`any_hash`, causing regressions with static symbols in `hash_aref_sym`

* hash.c (any_hash): skip rb_objid_hash for static syms
  (rb_num_hash_start): extract from rb_ident_hash
  (rb_objid_hash): call rb_num_hash_start
  (rb_ident_hash): ditto

~~~
target 0: a (ruby 2.3.0dev (2015-07-30 trunk 51437) [x86_64-linux]
target 1: b (ruby 2.3.0dev (2015-07-30 patch 51437) [x86_64-linux]

benchmark results from Xeon E3-1230 v3 @ 3.30GHz (turbo disabled):
minimum results in each 10 measurements.
Execution time (sec)
name                a       b
hash_aref_dsym        0.316   0.300
hash_aref_dsym_long   5.106   5.063
hash_aref_fix         0.304   0.297
hash_aref_flo         0.061   0.060
hash_aref_miss        0.433   0.430
hash_aref_str         0.408   0.396
hash_aref_sym         0.312   0.306
hash_aref_sym_long    0.482   0.469
hash_flatten          0.385   0.273
hash_ident_flo        0.036   0.037
hash_ident_num        0.277   0.276
hash_ident_obj        0.291   0.284
hash_ident_str        0.289   0.286
hash_ident_sym        0.285   0.281
hash_keys             0.269   0.271
hash_shift            0.020   0.016
hash_values           0.264   0.264
loop_whileloop2       0.101   0.099
vm2_bighash*          3.066   2.972

Speedup ratio: compare with the result of `a' (greater is better)
name                b
hash_aref_dsym        1.052
hash_aref_dsym_long   1.008
hash_aref_fix         1.024
hash_aref_flo         1.015
hash_aref_miss        1.007
hash_aref_str         1.031
hash_aref_sym         1.018
hash_aref_sym_long    1.027
hash_flatten          1.410
hash_ident_flo        0.994
hash_ident_num        1.001
hash_ident_obj        1.022
hash_ident_str        1.012
hash_ident_sym        1.016
hash_keys             0.992
hash_shift            1.237
hash_values           1.001
loop_whileloop2       1.013
vm2_bighash*          1.032
~~~

I will commit in a few weeks if everyone is OK with this.


---Files--------------------------------
0001-hash.c-improve-integer-fixnum-hashing.patch (5.17 KB)


-- 
https://bugs.ruby-lang.org/