On Sat, Jun 15, 2013 at 09:30:27AM +0200, Loic Dachary wrote: > Hi Martin, > > Just listened to > > https://2459d6dc103cb5933875-c0245c5c937c5dedcca3f1764ecc9b2f.ssl.cf2.rackcdn.com/fast13/plank_screaming.mp4 > from > https://www.usenix.org/conference/fast13/screaming-fast-galois-field-arithmetic-using-sse2-extensions > > and it's very entertaining ;-) > If video is what you like the Microsoft presentation of the previously mentioned PDF is here: https://www.usenix.org/conference/usenixfederatedconferencesweek/high-performance-vehicular-connectivity-opportunistic They also talk about why they made their choices. I guess that it's also in the PDF. > Cheers > > > On 06/14/2013 10:13 PM, Martin Flyvbjerg wrote: > > Dear Community > > I am a young engineer (not software or math, please bare with me) with some suggestions regarding erasure codes. I never used Ceph before or any other distributed file system. > > > > I stumped upon the suggestion for adding erasure codes to Ceph, as > > described in this article > > http://wiki.Ceph.com/01Planning/02Blueprints/Dumpling/Erasure_encoding_as_a_storage_backend > > > > first I would like to say great initiative to add erasure codes to Ceph. > > Ceph needs its own implementation and it have to be done right, I cannot stress this enough, suggested software mentioned in that article would result in very low performance. > > > > Why? > > Reed-Solomon is normally something regarded as being very slow compared to other erasure codes, because the underlying Galois-Field multiplication is slow. Please see video at usenix.org forexplanation. > > > > The implementations of Zfec library and other suggested software the others rely on the Vandermonde matrix, a matrix used in in Reed-Solomon erasure codes, a faster approach would be to use the Cauchy-Reed-Solomon implementation. Please see [1,2,3] > > Although there is something even better, by using the Intel SSE2/3 SIMD instructions it is possible to do the as fast as any other XOR based erasure codes (RaptorQ LT-codes, LDPC etc.). > > > > The suggested FECpp lib uses the optimisation but with a relative small Galois-field only 2^8, since Ceph aimes at unlimited scalability increasing the size of the Galois-Field would improve performance [4]. Of course the configured Ceph Object Size and/or Stripe width have to be taken into account. > > Please see > > https://www.usenix.org/conference/fast13/screaming-fast-galois-field-arithmetic-using-sse2-extensions > > > > > > The solution > > Using the GF-Complete open source library [4] to implement Reed-Solomon in Ceph in order to allow Ceph to scale to infinity. > > James S. Plank the author of GF-complete have developed a library implementing various Reed-Solomon codes called Jerasure. http://web.eecs.utk.edu/~plank/plank/www/software.html > > Jerasure 2.0 using the GF-complete artimetric based in Intel SSE SIMD instructions, is current in development expected release august 2013. Will be released under the new BSD license. Jerasure 2.0 also supports arbitrary Galois-field sizes 8,16,32,64 or 128 bit. > > > > The limit of this implementation would be the processors L2/L3 cache not the underlying arithmetic. > > > > Best Regards > > Martin Flyvbjerg > > > > [1] http://web.eecs.utk.edu/~plank/plank/papers/CS-05-569.pdf > > [2] http://web.eecs.utk.edu/~plank/plank/papers/CS-08-625.pdf > > [3] http://web.eecs.utk.edu/~plank/plank/papers/FAST-2009.pdf > > [4] http://web.eecs.utk.edu/~plank/plank/papers/FAST-2013-GF.pdf > > -- > > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > > the body of a message to majordomo@xxxxxxxxxxxxxxx > > More majordomo info at http://vger.kernel.org/majordomo-info.html > > -- > Loïc Dachary, Artisan Logiciel Libre > All that is necessary for the triumph of evil is that good people do nothing. > -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
