Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Thu, 10 Sep 2020 14:52:42 +0200
From: magnum <>
Subject: Re: rar-opencl performance

On 2020-09-08 00:22, Axymeus wrote:
> Is this what you're looking for? Sorry it's in french...
> Type: Fichier
> Taille: 271933564
> Taille compressée: 268645888
> Taux: 98%
> mtime: 2014-01-29 23:49:28,964786900
> Attributs: ..A....
> CRC32: CF45F153
> Système d'exploitation hôte: Windows
> Compression: RAR 3.0(v29) -m3 -md=4M
> Drapeaux: chiffré

So it's not a "stored" file but compressed.  This means you should get 
proper early rejection and the problem is simply your file is so large 
(rar2john will pick the smallest possible file from each archive).  Even 
with few false candidates *not* getting rejected early, the CPU 
post-processing take so long it hits overall performance as we need a 
full decompress of all 300 MB (per candidate, picture that) in order to 
see the false CRC and finally reject them.

I did some testing: Created a 3 KB file, a 3 MB file and a 300 MB one, 
all from /dev/urandom data, and rar'ed them with password.  This on a 
4-core, 8-thread Intel CPU and nvidia 2080ti:

John Jumbo 30-sec run at length 8:
   3 KB: 40K p/s, util ~ 77%, end temp 51C
   3 MB: 27K p/s, util ~ 60%, end temp 46C
300 MB: 2.9K p/s, util most often 0% seen, best 41%, end temp 40C

Even the 3 KB one did not show any utilization from JtR status line, I 
repeatedly ran nvidia-smi during run to get figures.

I got (much) worse performance from cRARk 5.5 for all sizes, not quite 
sure I used it properly.  Admittedly I used the OpenCL version and not 
the CUDA one.

There is one thing we could try to get a little better performance: 
Moving the early rejection Huffman checks to GPU.  I'm not sure it will 
fly but it I think it could do some good.  Even so, the worst case false 
candidates will always hit us hard.


> On Tue, Sep 8, 2020 at 12:00 AM magnum <> wrote:
>> On 2020-09-07 17:37, Axymeus wrote:
>>> On Sat, Sep 5, 2020 at 9:17 PM magnum <> wrote:
>>>> Axymeus, please post the result of "unrar vt <archive name>".  My guess
>>>> is this is a very large "stored" file.  I believe cRARk will process
>>>> such files fully on GPU if possible, while we will not - as of now.  We
>>>> could definitely implement that, I just never considered it worthwhile
>>>> because it's so uncommon.
>>> Can't find unrar, not even within cygwin, where should I get it?
>>> The archive does contain a single large file (~300 MB)
>> So use "rar vt <archive name>", it's the same thing.
>> 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.