Hi guys,
I'm contemplating implementing a new fast hash algorithm in Illumos' ZFS
implementation to supplant the currently utilized sha256. On modern
64-bit CPUs SHA-256 is actually much slower than SHA-512 and indeed much
slower than many of the SHA-3 candidates, so I went out and did some
testing (details attached) on a possible new hash algorithm that might
improve on this situation.
However, before I start out on a pointless endeavor, I wanted to probe
the field of ZFS users, especially those using dedup, on whether their
workloads would benefit from a faster hash algorithm (and hence, lower
CPU utilization). Developments of late have suggested to me three
possible candidates:
* SHA-512: simplest to implement (since the code is already in the
kernel) and provides a modest performance boost of around 60%.
* Skein-512: overall fastest of the SHA-3 finalists and much faster
than SHA-512 (around 120-150% faster than the current sha256).
* Edon-R-512: probably the fastest general purpose hash algorithm I've
ever seen (upward of 300% speedup over sha256) , but might have
potential security problems (though I don't think this is of any
relevance to ZFS, as it doesn't use the hash for any kind of security
purposes, but only for data integrity & dedup).
My testing procedure: nothing sophisticated, I took the implementation
of sha256 from the Illumos kernel and simply ran it on a dedicated
psrset (where possible with a whole CPU dedicated, even if only to a
single thread) - I tested both the generic C implementation and the
Intel assembly implementation. The Skein and Edon-R implementations are
in C optimized for 64-bit architectures from the respective authors (the
most up to date versions I could find). All code has been compiled using
GCC 3.4.3 from the repos (the same that can be used for building
Illumos). Sadly, I don't have access to Sun Studio.
Cheers,
--
Saso
Hash preformances on 10 GB of data
gcc (GCC) 3.4.3 (csl-sol210-3_4-20050802)
CFLAGS: -O3 -fomit-frame-pointer -m64
MACHINE #1
CPU: dual AMD Opteron 4234
Options: single thread on no-intr whole-CPU psrset
Algorithm Result Improvement
sha256 (ASM) 21.19 cycles/byte (baseline)
sha256 (C) 27.66 cycles/byte -23.34%
sha512 (ASM) 13.48 cycles/byte 57.20%
sha512 (C) 17.35 cycles/byte 22.13%
Skein-512 (C) 8.95 cycles/byte 136.76%
Edon-R-512 (C) 4.94 cycles/byte 328.94%
MACHINE #2
CPU: single AMD Athlon II Neo N36L
Options: single thread on no-intr 1-core psrset
Algorithm Result Improvement
sha256 (ASM) 15.68 cycles/byte (baseline)
sha256 (C) 18.81 cycles/byte -16.64%
sha512 (ASM) 9.95 cycles/byte 57.59%
sha512 (C) 11.84 cycles/byte 32.43%
Skein-512 (C) 6.25 cycles/byte 150.88%
Edon-R-512 (C) 3.66 cycles/byte 328.42%
MACHINE #3
CPU: dual Intel Xeon E5645
Options: single thread on no-intr whole-CPU psrset
Algorithm Result Improvement
sha256 (ASM) 15.49 cycles/byte (baseline)
sha256 (C) 17.90 cycles/byte -13.46%
sha512 (ASM) 9.88 cycles/byte 56.78%
sha512 (C) 11.44 cycles/byte 35.40%
Skein-512 (C) 6.88 cycles/byte 125.15%
Edon-R-512 (C) 3.35 cycles/byte 362.39%
MACHINE #4
CPU: single Intel Xeon E5405
Options: single thread on no-intr 1-core psrset
Algorithm Result Improvement
sha256 (ASM) 17.45 cycles/byte (baseline)
sha256 (C) 18.34 cycles/byte -4.85%
sha512 (ASM) 10.24 cycles/byte 70.41%
sha512 (C) 11.72 cycles/byte 48.90%
Skein-512 (C) 7.32 cycles/byte 138.39%
Edon-R-512 (C) 3.86 cycles/byte 352.07%
MACHINE #5
CPU: dual Intel Xeon E5450
Options: single thread on no-intr whole-CPU psrset
Algorithm Result Improvement
sha256 (ASM) 16.43 cycles/byte (baseline)
sha256 (C) 18.50 cycles/byte -11.19%
sha512 (ASM) 10.37 cycles/byte 58.44%
sha512 (C) 11.85 cycles/byte 38.65%
Skein-512 (C) 7.38 cycles/byte 122.63%
Edon-R-512 (C) 3.88 cycles/byte 323.45%
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
