Has anyone else run the benchmark and reproduced the performance difference?
On Tue, Aug 14, 2018 at 8:16 AM, Dmitriy Pavlov <dpavlov....@gmail.com> wrote: > It depends. > > CRC is a CPU-intensive operation, while WAL logging and page store write > are mostly about IO speed. > > In the same time, it can make the huge impact on machines with fast IO and > slow CPU. So if we can apply change proposed by Evgeniy and Alexey it could > benefit performance because we save CPU. Later we can use it's power in a > more efficient manner (e.g. with compression). > > вт, 14 авг. 2018 г. в 14:03, Yakov Zhdanov <yzhda...@apache.org>: > > > Guys, what time in % does crc calculation take in WAL logging process? > > > > --Yakov > > > > 2018-08-14 13:37 GMT+03:00 Dmitriy Pavlov <dpavlov....@gmail.com>: > > > > > Hi Alex, thank you for this idea. > > > > > > Evgeniy, Alex, would you like to submit the patch with bypassing > > > implementation differences to keep compatibility? > > > > > > Sincerely, > > > Dmitriy Pavlov > > > > > > вт, 14 авг. 2018 г. в 12:06, Alex Plehanov <plehanov.a...@gmail.com>: > > > > > > > Hello, Igniters! > > > > > > > > In java8 java.lang.zip.CRC32 methods become intrinsic, moreover new > > > > "update" method, which use ByteBuffer was introduced. Since we moved > to > > > > java8, perhaps we really can get performance boost by using standard > > > > java.lang.zip.CRC32 instead of PureJavaCrc32. > > > > > > > > About compatibility: looks like PureJavaCrc32 implements the same > > > algorithm > > > > as java.lang.zip.CRC32. These two implementations uses the same > > > polynomial > > > > and the same initial value. The only difference is final xor mask > > > > (0xFFFFFFFF for java.lang.zip.CRC32). So, we can easily convert from > > > > PureJavaCrc32 > > > > to standard CRC32 and vice versa, using this expression: crc32 ^= > > > > 0xFFFFFFFF > > > > > > > > > > > > 2018-08-14 0:19 GMT+03:00 Eduard Shangareev < > > eduard.shangar...@gmail.com > > > >: > > > > > > > > > Evgeniy, > > > > > > > > > > Could you share benchmark code? And please share what version of > JVM > > > > > you have used. > > > > > > > > > > On Mon, Aug 13, 2018 at 10:44 PM Zhenya <arzamas...@mail.ru.invalid > > > > > > > wrote: > > > > > > > > > > > I think it would break backward compatibility, as Nikolay > mentioned > > > > above > > > > > > we would take exception here: > > > > > > > > > > > > [1] > > > > > > > > > > > > https://github.com/apache/ignite/blob/master/modules/ > > > > > core/src/main/java/org/apache/ignite/internal/processors/ > > > > > cache/persistence/file/FilePageStore.java#L372 > > > > > > > > > > > > thats why i question for community thoughts here. > > > > > > > > > > > > > Hi Evgeniy, > > > > > > > > > > > > > > would you like to submit a patch with CRC32 implementation > > change? > > > > > > > > > > > > > > Sincerely, > > > > > > > Dmitriy Pavlov > > > > > > > > > > > > > > пн, 13 авг. 2018 г. в 22:08, Евгений Станиловский > > > > > > > <arzamas...@mail.ru.invalid>: > > > > > > > > > > > > > >> Hi, igniters, i wrote a simple bench, looks like PureJavaCrc32 > > has > > > > > > >> performance problems in compatible with zip.CRC32. > > > > > > >> > > > > > > >> Benchmark Mode Cnt Score Error Units > > > > > > >> BenchmarkCRC.Crc32 avgt 5 1088914.540 ± 368851.822 ns/op > > > > > > >> BenchmarkCRC.pureJavaCrc32 avgt 5 6619408.049 ± 3746712.210 > > ns/op > > > > > > >> > > > > > > >> thoughts? > > > > > > > > > > > > > > > > > > > > >
