Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > comp.programming > #4980
| From | Ramine <ramine@1.1> |
|---|---|
| Newsgroups | comp.programming |
| Subject | About my previous post |
| Date | 2014-12-06 15:49 -0800 |
| Organization | A noiseless patient Spider |
| Message-ID | <m5vq2q$csn$1@dont-email.me> (permalink) |
Hello, I have wrote in my previous post this: "I have benchmarked the distribute reader-write lock of Dmitry Vyukov and i have found that it is so expensive that it takes over 16 or 32 cores to make it scale as Seqlock with 4 cores" You will say that in computer science we need proof of what i am saying above.. Here is my proof by using the Amdahl's law: The distributed reader-writer mutex using a rwlock of Dmitry Vyukov don't eliminate the full memory barrier in the reader side of the RWLock used by the distributed reader-writer mutex of Dmitry Vyukov, this full memory barrier does take 350 CPU cycles on my quadcore Q6600 and i think it takes around the same cycles on Intel i7, i have just done a parallel program benchmark between Seqlock and the distributed reader-writer mutex of Dmitry Vyukov using a rwlock on x86 and it gives on read-mostly scenario 25 milliseconds with Seqlock and it takes around 165 milliseonds with the Dmitry Vykov distributed reader-writer mutex using a rwlock, it takes 165 milliseconds because it is using an expensive call to GetCurrentProcessorNumber() of Windows and it is using full memory fences on the reader side and it is using more code on the reader side , but this GetCurrentProcessorNumber() and full memory fences on the Dmitry Vyukov algorithm are part of the parallel part of the Amdahl's law, so they will scale, but they are so expensive compared to Seqlock that it takes 16 to 32 cores on small to medium reader section size to make it scale as a Seqlock on 4 cores, cause look at the 165 milliseconds compared to the 25 seconds, there is a big difference, this is my proof , so the distributed reader-writer mutex using a rwlock is scalable but it is too much slow than Seqlock or RCU on reader-mostly scenarios, this is why i have invented my scalable distributed sequential lock that beats Seqlock and that is as fast and as scalable as both Seqlock and RCU on read-mostly scenarios. You can download my scalable distributed sequential lock version 1.1 from: https://sites.google.com/site/aminer68/scalable-distributed-sequential-lock Thank you, Amien Moulay Ramdane.
Back to comp.programming | Previous | Next | Find similar | Unroll thread
About my previous post Ramine <ramine@1.1> - 2014-12-06 15:49 -0800
csiph-web