Path: csiph.com!usenet.pasdenom.info!weretis.net!feeder4.news.weretis.net!eternal-september.org!feeder.eternal-september.org!mx04.eternal-september.org!.POSTED!not-for-mail From: "aminer" Newsgroups: comp.lang.java.programmer Subject: Distributed Reader-Writer Mutex version 1.03 Date: Thu, 13 Sep 2012 16:28:22 -0500 Organization: A noiseless patient Spider Lines: 130 Message-ID: Injection-Date: Thu, 13 Sep 2012 20:28:40 +0000 (UTC) Injection-Info: mx04.eternal-september.org; posting-host="c43ca82f9e8d62a602307fe9d2e9b807"; logging-data="12003"; mail-complaints-to="abuse@eternal-september.org"; posting-account="U2FsdGVkX1/C/0mDSt+K2oDUVECXxaxB" X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2900.5512 X-RFC2646: Format=Flowed; Original X-Newsreader: Microsoft Outlook Express 6.00.2900.5512 Cancel-Lock: sha1:4qhKno4CWQom/oHq2SJnnJzczTI= X-Priority: 3 X-MSMail-Priority: Normal Xref: csiph.com comp.lang.java.programmer:18745 Hello, Distributed Reader-Writer Mutex 1.03 Description: Distributed Reader-Writer Mutex, based on the Dmitry Vyukov C++ Distributed Reader-Writer Mutex , I have included the following Reader-Writer Mutexes inside this Distributed Reader-Writer mutex: TOmniMREW a light weight MREW that is very fast and TMultiReadExclusiveWrite from JCL and now both of them can scale better, and i have modified the Dmitry Vyukov Distributed Reader-Writer Mutex, in the first version i have not used GetCurrentProcessor() but i have used GetCurrentThreadID(), and i have also provided you with a second version that scales better, to be able to use the second version please use the version2 in defines.inc, i have given you a test.pas example for the first version and test1.pas for the second version, but don't forget to use version2 inside defines.inc, to use the second version just uncomment the version2 inside defines.inc and comment version1. I have also done a cache line alignement in TOmniMREW, this has allowed Drwlock to scale better. I have provided you with the source code, please take a look at the source code to understand better.The Object Pascal Distributed Reader-Writer Mutex is based on the following C++ Distributed Reader-Writer Mutex by Dmitry Vyukov, read more here: http://www.1024cores.net/home/lock-free-algorithms/reader-writer-problem/distributed-reader-writer-mutex I have also modified the Dmitry Vyukov's Distributed Reader-Writer Mutex to use a variable number of MREWs, you can pass the number of MREWs to the constructor like this: drw:=TDRWLOCK.create(100); You have four methods: procedure wlock; // like the EnterWriteLock of TOmniMREW procedure wunlock; // like the ExitWriteLock procedure rlock; // like EnterReadLock procedure runlock; // like the ExitReadLock and you have to pass the number of MREWs(multiple-readers-exclusive-writer) to the constructor like this: drw:=TDRWLOCK.create(200); // here we are creating 200 MEWs Here is some scalability numbers: I have used TOmniMREW of the Omnithread library and used only EnterReadLock() and ExitReadLock() with four threads on a quad cores and TOmniMREW gave a negative scalability of -5.51x And when i have used the second version of Distributed Reader-Writer Mutex using only rlock() and runlock() , it gave me +3.94x scalability with four threads on four cores. So now it's scaling. And about the second version , don't forget to initialize the number that you pass to rlock() and runlock() to 0 before calling rlock() and runlock() . In the previous versions i have aligned the array elements on cache line bounderies like have done Dmitry Vyukov, and it didn't work correctly when i have tested the second version, so i have thought about that and after that i have decided to not align the array elements on cache line bounderied but just add a cache line padding to TOmniMREW for example and this time it has worked perfectly and now the second version is scaling perfectly.. And if you have noticed , there is still a weakness with the Dmitry Vyukov C++ Distributed Reader-Writer Mutex, cause since he is using GetCurrentProcessorNumber() he is limiting the array of rwlocks to the number of avaiblable cores and this is not good i think , cause if you have many more threads than the avaiblable cores and there is high contention this will cause the performance to degrade, so i have decided to change that in my implementation and i have used a variable number of rwlocks/MREWs so that you can lower more the contention and this is better for scalability. Language: FPC Pascal v2.2.0+ / Delphi 7+: http://www.freepascal.org/ Operating Systems: Win and Linux (x86). Required FPC switches: -O3 -Sd -dFPC -dWin32 -dFreePascal -Sd for delphi mode.... Required Delphi switches: -DMSWINDOWS -$H+ For Delphi -DDelphi And inside defines.inc you can use the following defines: {$DEFINE CPU32} for 32 bits systems {$DEFINE CPU64} for 64 bits systems {$DEFINE TOmniMREW} to use Omnithread MREW {$DEFINE TMultiReadExclusiveWrite} to use the jcl TMultiReadExclusiveWrite You can download Distributed Reader-Writer Mutex version 1.03 from: http://pages.videotron.com/aminer/ Thank you, Amine Moulay Ramdane.