https://community.intel.com/t5/Intel-Moderncode-for-Parallel/Optimistic-shychronization/m-p/967588#M5459 <P>Hello,<BR /><BR />I have come to an interresting subject...<BR /><BR />I was thinking more about my ParallelVarFiler and about database engines, first i have told you that since database engines are disk <BR />bound and/or memory bound that means that they are not scalable <BR />in multicores systems, but more than that that i was thinking more<BR />and asking my self why to use optimistic synchronization mechanism like in transactional memory ? i think that an optimistic synchronization mechanism like in transactional memory can give better throughput and better speed, prof of that ? take a look at my ParallelVarFiler for exemple, <BR />since it can be disk bound when you are saving automaticly the <BR />data to a disk file that means that there is no need to use <BR />an RWLock since the hardisk is not trully parallel.. so a Lock will have <BR />the same performance as an RWLock in a disk bound scenario, it's why i have used a FIFO fair Lock inside my ParallelVarFiler, but if the hardisk or memory can be truly parallel an RWLock will be better , <BR />it's the same with an optimistic sychronization like in transactional memory, first you have to know that a Lock will serialize the read and write transactions, an RWLock will parallelize the read transaction, but&nbsp; with optimistic synchronization mechanism like in transactional memory you can have in theory many writes transactions in parallel <BR />and many read transactions in parallel , but since the hardisk is not truly parallel, this <BR />optimistic synchronization like in transactional memory will not bring <BR />better performance than a RWLock , so i think that transactional memory and <BR />optimistic synchronization can not give better performance if the hardisk and memory are not trully parallel, but since the memory system <BR />and hardisk system is not truly parallel today, so this <BR />optmistic schychronisation that is using transactional memory or other systemes, will not bring you better performance than a RWLock.<BR /><BR /><BR />Thank you,<BR />Amine Moulay Ramdane.<BR /><BR /><BR /><BR />&nbsp;<BR /><BR /><BR /></P> Mon, 28 Oct 2013 03:00:44 GMT aminer10 2013-10-28T03:00:44Z https://community.intel.com/t5/Intel-Moderncode-for-Parallel/Optimistic-shychronization/m-p/967588#M5459 <P>Hello,<BR /><BR />I have come to an interresting subject...<BR /><BR />I was thinking more about my ParallelVarFiler and about database engines, first i have told you that since database engines are disk <BR />bound and/or memory bound that means that they are not scalable <BR />in multicores systems, but more than that that i was thinking more<BR />and asking my self why to use optimistic synchronization mechanism like in transactional memory ? i think that an optimistic synchronization mechanism like in transactional memory can give better throughput and better speed, prof of that ? take a look at my ParallelVarFiler for exemple, <BR />since it can be disk bound when you are saving automaticly the <BR />data to a disk file that means that there is no need to use <BR />an RWLock since the hardisk is not trully parallel.. so a Lock will have <BR />the same performance as an RWLock in a disk bound scenario, it's why i have used a FIFO fair Lock inside my ParallelVarFiler, but if the hardisk or memory can be truly parallel an RWLock will be better , <BR />it's the same with an optimistic sychronization like in transactional memory, first you have to know that a Lock will serialize the read and write transactions, an RWLock will parallelize the read transaction, but&nbsp; with optimistic synchronization mechanism like in transactional memory you can have in theory many writes transactions in parallel <BR />and many read transactions in parallel , but since the hardisk is not truly parallel, this <BR />optimistic synchronization like in transactional memory will not bring <BR />better performance than a RWLock , so i think that transactional memory and <BR />optimistic synchronization can not give better performance if the hardisk and memory are not trully parallel, but since the memory system <BR />and hardisk system is not truly parallel today, so this <BR />optmistic schychronisation that is using transactional memory or other systemes, will not bring you better performance than a RWLock.<BR /><BR /><BR />Thank you,<BR />Amine Moulay Ramdane.<BR /><BR /><BR /><BR />&nbsp;<BR /><BR /><BR /></P> Mon, 28 Oct 2013 03:00:44 GMT https://community.intel.com/t5/Intel-Moderncode-for-Parallel/Optimistic-shychronization/m-p/967588#M5459 aminer10 2013-10-28T03:00:44Z