[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 18 Aug 2015 23:31:52 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Long duration formats' many-salts figure incorrect in short benchmarks
(was: 7z's KDF is unsalted)
Solar,
On 2015-08-18 22:59, magnum wrote:
> $ ../run/john -test -form:7z
> Will run 8 OpenMP threads
> Benchmarking: 7z, 7-Zip (512K iterations) [SHA256 AES 32/64]... (8xOMP) DONE
> Speed for cost 1 (iteration count) of 524288
> Many salts: 7876 c/s real, 1085 c/s virtual
> Only one salt: 42.1 c/s real, 5.5 c/s virtual
I did the same change for the OpenCL version:
$ ../run/john -test -form:7z-opencl -dev=2
Will run 8 OpenMP threads
Device 2: GeForce GT 650M
Benchmarking: 7z-opencl, 7-Zip (512K iterations) [SHA256 OPENCL AES]...
(8xOMP) DONE
Speed for cost 1 (iteration count) of 524288
Many salts: 108 c/s real, 12800 c/s virtual
Only one salt: 108 c/s real, 12800 c/s virtual
So why don't we see *any* boost here? After some digging around I
realized the benchmark timer ends before we even get to try the second
salt even once! This example is a toy GPU but it happens even on
flagship ones.
So running a much longer test does show the boost:
$ ../run/john -test=10 -form:7z-opencl -dev=2
Will run 8 OpenMP threads
Device 2: GeForce GT 650M
Benchmarking: 7z-opencl, 7-Zip (512K iterations) [SHA256 OPENCL AES]...
(8xOMP) DONE
Speed for cost 1 (iteration count) of 524288
Many salts: 13818 c/s real, 308404 c/s virtual
Only one salt: 108 c/s real, 12800 c/s virtual
Longer runs show better figures (within limits). This problem is with
"your code", but it never surfaces in non-Jumbo. I'm not sure whether to
fix this, or how. Should we simply ensure the while loop runs long
enough for both salts to be tested at least once? Maybe that would end
up showing a mere 2x boost in this case? Should we ensure the loop does
all BENCHMARK_MANY virtual salts? But that would likely end up in some
GPU formats benchmarking for hours...
BTW this also boosts the CPU format figure:
$ ../run/john -test=4 -form:7z
Will run 8 OpenMP threads
Benchmarking: 7z, 7-Zip (512K iterations) [SHA256 AES 32/64]... (8xOMP) DONE
Speed for cost 1 (iteration count) of 524288
Many salts: 9606 c/s real, 1356 c/s virtual
Only one salt: 41.7 c/s real, 5.4 c/s virtual
A four second test yields a 219x boost instead of the original 187x.
(BTW I see now I recalled incorrectly; BENCHMARK_MANY isn't 500 but 256,
so the theoretical max boost is < 256x in this situation).
magnum
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
