1 Commits

Author	SHA1	Message	Date
Ben Pfaff	0e66616040	cmap: New module for cuckoo hash table. ... This implements an "optimistic concurrent cuckoo hash", a single-writer, multiple-reader hash table data structure. The point of this data structure is performance, so this commit message focuses on performance. I tested the performance of cmap with the test-cmap utility included in this commit. It takes three parameters for benchmarking: - n, the number of elements to insert. - n_threads, the number of threads to use for searching and mutating the hash table. - mutations, the percentage of operations that should modify the hash table, from 0% to 100%. e.g. "test-cmap 1000000 16 1" inserts one million elements, uses 16 threads, and 1% of the operations modify the hash table. Any given run does the following for both hmap and cmap implementations: - Inserts n elements into a hash table. - Iterates over all of the elements. - Spawns n_threads threads, each of which searches for each of the elements in the hash table, once, and removes the specified percentage of them. - Removes each of the (remaining) elements and destroys the hash table. and reports the time taken by each step, The tables below report results for various parameters with a draft version of this library. The tests were not formally rerun for the final version, but the intermediate changes should only have improved performance, and this seemed to be the case in some informal testing. n_threads=16 was used each time, on a 16-core x86-64 machine. The compiler used was Clang 3.5. (GCC yields different numbers but similar relative results.) The results show: - Insertion is generally 3x to 5x faster in an hmap. - Iteration is generally about 3x faster in a cmap. - Search and mutation is 4x faster with .1% mutations and the advantage grows as the fraction of mutations grows. This is because a cmap does not require locking for read operations, even in the presence of a writer. With no mutations, however, no locking is required in the hmap case, and the hmap is somewhat faster. This is because raw hmap search is somewhat simpler and faster than raw cmap search. - Destruction is faster, usually by less than 2x, in an hmap. n=10,000,000: .1% mutations 1% mutations 10% mutations no mutations cmap hmap cmap hmap cmap hmap cmap hmap insert: 6132 2182 6136 2178 6111 2174 6124 2180 iterate: 370 1203 378 1201 370 1200 371 1202 search: 1375 8692 2393 28197 18402 80379 1281 1097 destroy: 1382 1187 1197 1034 324 241 1405 1205 n=1,000,000: .1% mutations 1% mutations 10% mutations no mutations cmap hmap cmap hmap cmap hmap cmap hmap insert: 311 25 310 60 311 59 310 60 iterate: 25 62 25 64 25 57 25 60 search: 115 637 197 2266 1803 7284 101 67 destroy: 103 64 90 59 25 13 104 66 n=100,000: .1% mutations 1% mutations 10% mutations no mutations cmap hmap cmap hmap cmap hmap cmap hmap insert: 25 6 26 5 25 5 25 5 iterate: 1 3 1 3 1 3 2 3 search: 12 57 27 219 164 735 10 5 destroy: 5 3 6 3 2 1 6 4 Signed-off-by: Ben Pfaff <blp@nicira.com> Acked-by: Jarno Rajahalme <jrajahalme@nicira.com>	2014-05-20 16:55:20 -07:00