Issue #12142 has been updated by Shyouhei Urabe.


Vladimir Makarov wrote:
> I don't think it is a leak.  What you measure is the maximal residential size.  I think the table is rebuilt many times and memory for previous version of tables is freed but it is not freed to OS by MRI (or glibc.  I don't know yet what allocation library is used by MRI).  Still this is *very* bad.  I should definitely to investigate and fix it.  I believe I know how to fix it.  I should reuse the array elements when it is possible.  Thanks for pointing this out.

OK.  Looking forward to the fix.

> Would you mind if I include your test in the final version of the patch as a benchmark?

No problem.  Do so pleae.

> > When you want to use a hash as an in-memory key-value store, it is quite natural for it to experience lots of random additions / deletions.  I think this situation happens in real-world programs.  Is this intentional or just a work-in-progress?
> >
> 
> The proposed hash tables will work with random additions/deletions.  I only did not know what the exact performance will be in comparison with the current tables.  As I became aware of your case now (Yura Sokolov also wrote about it) it will be a work-in-progress for me.
> 
> I am not sure your case is real world case scenario (removing the last element is) but it is definitely the worst case for my implementation. 

(WIP is definitely OK with me.) Let me think of a more realistic use case.

Often, a key-value store comes with expiration of keys.  This is typically done by storing expiration time in each hash elements, then periodically call st_foreach() with a pointer argument to an expiration-checking function.  To simulate this, I wrote following snippet:

````ruby
h = Hash.new
m = Mutex.new

# Thread to purge expired entries
Thread.start{
  loop{
    t = Time.now
    m.synchronize{
      h.reject!{|k, v| v < t }
    }
  }
}

# insertions
1000.times{|i|
  t = Time.now + 0.1
  m.synchronize {
    1000.times{|j|
      h[1000*i+j] = t
    }
  }
  STDERR.printf"%d: %d\r", i, h.size
}
puts
````
(The printf was for me to check progress.)

It ran with your patch and resulted in 37,668kb max resident memory.  Without the patch, it took 8,988kb.

Above might still be a bit illustrative, but I believe this situation happens in wild.

> The worst case is probably the same as for the current tables.   It is theoretically possible to create test data which results in usage the same entry for the current and proposed tables.   But in practice it is impossible for medium size table even if murmur hash is not a cryptography level hash function as, for example, sha2.
> 
> I've specifically chosen a very small hash table load (0.5) to make chance of collisions very small and rebuilding less frequent (such parameter still results in about the same non-small hash table sizes).  I think even maximal load 3/4 would work well to avoid collision impact.  People can make experiments with such parameters and change them later if something works better of course if the proposed tables will be in the trunk.
>   
> But still if there are a lot of collisions the same strategy can be used -- table rebuilding.  I'll think about this. 

Yes.  Please do.  "People are not evil(smart) enough to do this" -kind of assumptions tends to fail.


----------------------------------------
Feature #12142: Hash tables with open addressing
https://bugs.ruby-lang.org/issues/12142#change-57332

* Author: Vladimir Makarov
* Status: Open
* Priority: Normal
* Assignee: 
----------------------------------------
~~~
 Hello, the following patch contains a new implementation of hash
tables (major files st.c and include/ruby/st.h).

  Modern processors have several levels of cache.  Usually,the CPU
reads one or a few lines of the cache from memory (or another level of
cache).  So CPU is much faster at reading data stored close to each
other.  The current implementation of Ruby hash tables does not fit
well to modern processor cache organization, which requires better
data locality for faster program speed.

The new hash table implementation achieves a better data locality
mainly by

  o switching to open addressing hash tables for access by keys.
    Removing hash collision lists lets us avoid *pointer chasing*, a
    common problem that produces bad data locality.  I see a tendency
    to move from chaining hash tables to open addressing hash tables
    due to their better fit to modern CPU memory organizations.
    CPython recently made such switch
    (https://hg.python.org/cpython/file/ff1938d12240/Objects/dictobject.c).
    PHP did this a bit earlier
    https://nikic.github.io/2014/12/22/PHPs-new-hashtable-implementation.html.
    GCC has widely-used such hash tables
    (https://gcc.gnu.org/svn/gcc/trunk/libiberty/hashtab.c) internally
    for more than 15 years.

  o removing doubly linked lists and putting the elements into an array
    for accessing to elements by their inclusion order.  That also
    removes pointer chaising on the doubly linked lists used for
    traversing elements by their inclusion order.

A more detailed description of the proposed implementation can be
found in the top comment of the file st.c.

The new implementation was benchmarked on 21 MRI hash table benchmarks
for two most widely used targets x86-64 (Intel 4.2GHz i7-4790K) and ARM
(Exynos 5410 - 1.6GHz Cortex-A15):

make benchmark-each ITEM=bm_hash OPTS='-r 3 -v' COMPARE_RUBY='<trunk ruby>'

Here the results for x86-64:

hash_aref_dsym       1.094
hash_aref_dsym_long          1.383
hash_aref_fix        1.048
hash_aref_flo        1.860
hash_aref_miss       1.107
hash_aref_str        1.107
hash_aref_sym        1.191
hash_aref_sym_long           1.113
hash_flatten         1.258
hash_ident_flo       1.627
hash_ident_num       1.045
hash_ident_obj       1.143
hash_ident_str       1.127
hash_ident_sym       1.152
hash_keys            2.714
hash_shift           2.209
hash_shift_u16       1.442
hash_shift_u24       1.413
hash_shift_u32       1.396
hash_to_proc         2.831
hash_values          2.701

The average performance improvement is more 50%.  ARM results are
analogous -- no any benchmark performance degradation and about the
same average improvement.

The patch can be seen as

https://github.com/vnmakarov/ruby/compare/trunk...hash_tables_with_open_addressing.patch

or in a less convenient way as pull request changes

https://github.com/ruby/ruby/pull/1264/files


This is my first patch for MRI and may be my proposal and
implementation have pitfalls.  But I am keen to learn and work on
inclusion of this code into MRI.

~~~



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

Unsubscribe: <mailto:ruby-core-request / ruby-lang.org?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>