Date: Sat, 2 Aug 2014 13:23:09 +0400 From: Solar Designer <solar@...nwall.com> To: john-users@...ts.openwall.com Subject: Re: PKZIP and GPU acceleration Hi, On Fri, Aug 01, 2014 at 10:36:04PM -0500, Dylan Staley wrote: > I've been playing around with John, compiled from the bleeding-edge branch > (without modifying the Makefile), and I've noticed that a lot of the > formats that have CUDA or OpenCL support are listed as such under the > format option. Now, I've read countless articles about using John on > CUDA-enabled EC2 instances, but seeing as how PKZIP *doesn't* have a CUDA > or OpenCL version listed in the format option, does it even matter? For this format, GPUs don't matter. (We do have a GPU-enabled format for the newer WinZip/AES archives, but not for the old PKZIP.) > I wanted to try and figure this out myself, so I ran john --format=PKZIP > --test on two EC2 instances: > > * g2.2xlarge: 26 ECUs, 8 vCPUs, 2.6 GHz, Intel Xeon E5-2670, 15 GiB memory, > NVIDIA GRID K520 GPU > * c3.8xlarge: 108 ECUs, 32 vCPUs, 2.8 GHz, Intel Xeon E5-2680v2, 60 GiB > memory > > The results were as follows: > > g2.2xlarge > $ ./john --format=PKZIP --test > Will run 8 OpenMP threads > Benchmarking: PKZIP [32/64]... (8xOMP) DONE > Many salts: 50288K c/s real, 6301K c/s virtual > Only one salt: 22616K c/s real, 2837K c/s virtual > > c3.8xlarge > $ ./john --format=PKZIP --test > Will run 32 OpenMP threads > Benchmarking: PKZIP [32/64]... (32xOMP) DONE > Many salts: 144437K c/s real, 4519K c/s virtual > Only one salt: 33959K c/s real, 1061K c/s virtual You need to use --fork=8 or --fork=32, respectively, to achieve much better cumulative speeds on these machines. Our OpenMP scaling at speeds this high isn't great. > At this point, I was completely convinced that john is using the CPU for > PKZIP, but I wanted to see just how much its performance could be increased > in the real world. I used zip2john to get the hash for a zip file with a > fairly simple password ("mi12345"). It took the g2 instance (the one with > the GPU) 21 minutes to find the password, and the more powerful c3 instance > 15 minutes to find the password. You'll get much lower running times, and a bigger difference between the two systems (should be 3x or so), with --fork. > If I'm correct in assuming that John doesn't support using a GPU on PKZIP, > what would be needed to enable this? Code would need to be written, and for speeds this high some redesign would also be needed (simply adding a GPU-enabled format would make little sense as we'd bump into the communication bottleneck right away). > Does anyone know of anything that is > able to use the GPU with better performance than a CPU? For the old PKZIP archives, no. Alexander
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.