From: vmakarov@... Date: 2016-11-04T21:23:48+00:00 Subject: [ruby-core:77934] [Ruby trunk Feature#12142] Hash tables with open addressing Issue #12142 has been updated by Vladimir Makarov. File size16.png added File size256.png added File size60000.png added Koichi's investigation reported that small hash tables are majority during work of Ruby on Rails: https://gist.github.com/ko1/dda8243134bc40f7fc05e293abc5f4b2#file-report-md Memory footprint of cloud applications like Ruby on Rails is important topic. Since Koichi's report some changes have been done in my (https://github.com/vnmakarov/ruby/tree/hash_tables_with_open_addressing) and Yura's (https://github.com/funny-falcon/ruby/tree/st_table_with_array3) implementations. Here are memory consumption graphs. The data produced by Koichi's script https://gist.githubusercontent.com/ko1/dda8243134bc40f7fc05e293abc5f4b2/raw/dfccc2cc5e6f4f749c8a9390eae3104da84eb32f/stbench.rb on 4.2 GHz i7-4790K under FC24. !size16.png! !size256.png! !size60000.png! On these graphs * 'vlad current' means my current variant (https://github.com/vnmakarov/ruby.git, branch hash_tables_with open addressing, f4d2bf3ab78e59118e5dd89bcc5c3d8c4371ed35). * 'vlad april' means my April variant without the last two commits to st.c. * trunk means the current trunk (03c9bc2) * yura means latest Yura's variant (https://github.com/funny-falcon/ruby.git, branch st_table_with_array3, 39d9b2c3658c285fbe4b82edc1c7aebaabec6aaf) of tables with open addressing. I used a 64-bit variant, there is no practically difference in performance and memory consumption with 32-bit variant. Here are the tables speed improvements relative to the trunk on MRI hash benchmarks: ``` Yura My April My current bighash 1.716 1.509 1.443 hash_aref_dsym 1.019 1.018 0.994 hash_aref_dsym_l1.394 1.414 1.377 hash_aref_fix 0.984 1.041 1.015 hash_aref_flo 1.713 1.832 1.828 hash_aref_miss 1.117 1.191 1.131 hash_aref_str 1.177 1.225 1.222 hash_aref_sym 0.959 1.015 1.012 hash_aref_sym_l 1.047 1.066 1.042 hash_flatten 1.165 1.095 1.087 hash_ident_flo 1.005 0.949 0.958 hash_ident_num 0.932 0.979 0.954 hash_ident_obj 0.886 0.934 0.919 hash_ident_str 0.919 0.942 0.919 hash_ident_sym 0.983 1.000 0.967 hash_keys 3.024 3.039 3.073 hash_long 0.751 1.483 1.486 hash_shift 1.363 1.371 1.346 hash_shift_u16 1.388 1.392 1.359 hash_shift_u24 1.344 1.345 1.310 hash_shift_u32 1.330 1.326 1.290 hash_small2 0.918 1.016 1.043 hash_small4 0.928 1.074 1.048 hash_small8 1.581 2.268 1.894 hash_to_proc 1.024 1.067 1.051 hash_values 2.801 2.833 2.822 vm2_bighash* 2.669 3.261 3.001 Average 1.33841 1.43278 1.39226 ``` The data are obtained by running the following script on the same machine used exclusively for benchmarking: ``` ruby ../ruby/benchmark/driver.rb -p hash -r 3 -e trunk::trunk/miniruby -e yura::yura/miniruby -e yura::yura64/miniruby -e current::./miniruby 2>/dev/n\ ull|awk 'NF==4 && /hash/ {s1+=$2;s2+=$3;s3+=$4;n++;print} END{print s1/n, s2/n, s3/n}' ``` Conclusions: * My current variant requires less memory than April one. The current variant is close to average memory consumption of Yura's one (see the integral, a square of areas under the curves). * Yura's curve is more smooth because of more frequent and smaller rebuilding but probably it results in the speed decrease. * There is trade-off in my implementation (and probably Yura's ones) in table speed and its size. * More size reduction in my tables by increasing table fill rate (e.g. from 0.5 to 3/4) is not worth it as one bin size is 1/24 - 1/12 of one entry size for tables with less 2^15 elements. * Unfortunately, it is not known what part of the overall footprint belongs to the tables in Rails. Hash tables might be a small part in comparison with other Rails data. Even if the tables are the lion part, their elements and keys might require more memory than table entries. * Therefore it is hard to say what trade-off in performance/table size should be for Rails. I personally still prefer my April variant. ---------------------------------------- Feature #12142: Hash tables with open addressing https://bugs.ruby-lang.org/issues/12142#change-61253 * 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='' 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. ~~~ ---Files-------------------------------- st-march31.patch (114 KB) base.patch (93.8 KB) hash.patch (4.48 KB) strong_hash.patch (8.08 KB) city.patch (19.4 KB) new-hash-table-benchmarks.patch (1.34 KB) size16.png (6.91 KB) size256.png (6.95 KB) size60000.png (7.59 KB) -- https://bugs.ruby-lang.org/ Unsubscribe: