https://community.intel.com/t5/Intel-Moderncode-for-Parallel/Scalable-hash-map-algorithm/m-p/887929#M3641 I think this can interesting for this forum too.<BR />I've posted implementation of highly scalable concurrent hash map on TBB forum:<BR /><A href="http://software.intel.com/en-us/forums//topic/60494" target="_blank">http://software.intel.com/en-us/forums//topic/60494</A><BR />(there is a C++ source attached)<BR /><BR />Implementation uses a bunch of advanced synchronization techniques, like timestamp validation, deferred memory reclamation and cache conscious data layout. As a result this implementation beats TBB's concurrent_hash_map by a factor of 50 on read-mostly workload on Intel Core 2 Quad (Q6600), and by a factor of 20 on modest write workload. The cost of read transaction (find operation) is about 30 cycles, which is basically equal to that of single-threaded hash map. Algorithm also have perfect linear scalability on read-mostly workload so on greater number of cores/processors it will have even higher performance difference with traditional lock-based synchronization (on 8 cores I expect &gt;100x).<BR /><BR />Such synchronization techniques, I believe, is the way to deal with future many-core systems.<BR /><BR /> Thu, 11 Sep 2008 20:30:19 GMT Dmitry_Vyukov 2008-09-11T20:30:19Z https://community.intel.com/t5/Intel-Moderncode-for-Parallel/Scalable-hash-map-algorithm/m-p/887929#M3641 I think this can interesting for this forum too.<BR />I've posted implementation of highly scalable concurrent hash map on TBB forum:<BR /><A href="http://software.intel.com/en-us/forums//topic/60494" target="_blank">http://software.intel.com/en-us/forums//topic/60494</A><BR />(there is a C++ source attached)<BR /><BR />Implementation uses a bunch of advanced synchronization techniques, like timestamp validation, deferred memory reclamation and cache conscious data layout. As a result this implementation beats TBB's concurrent_hash_map by a factor of 50 on read-mostly workload on Intel Core 2 Quad (Q6600), and by a factor of 20 on modest write workload. The cost of read transaction (find operation) is about 30 cycles, which is basically equal to that of single-threaded hash map. Algorithm also have perfect linear scalability on read-mostly workload so on greater number of cores/processors it will have even higher performance difference with traditional lock-based synchronization (on 8 cores I expect &gt;100x).<BR /><BR />Such synchronization techniques, I believe, is the way to deal with future many-core systems.<BR /><BR /> Thu, 11 Sep 2008 20:30:19 GMT https://community.intel.com/t5/Intel-Moderncode-for-Parallel/Scalable-hash-map-algorithm/m-p/887929#M3641 Dmitry_Vyukov 2008-09-11T20:30:19Z