Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Wed, 15 Apr 2015 04:55:19 +0200
From: magnum <>
Subject: Re: bechmark versus self-test

On 2015-04-08 14:56, Solar Designer wrote:
> On Sun, Apr 05, 2015 at 07:48:19PM -0300, Claudio Andr? wrote:
>> JtR should allow the format to set which ciphers in test[] should be 
>> used to benchmark. Reason:
>> - work only with passwords of the same size/cost/whatever.
> IIRC, during benchmarking we currently use only the first two test
> vectors' salt strings (we alternate them for the "many salts" benchmark),
> but potentially with all of the test vectors' plaintexts.
> So the requirement to "work only with passwords of the same [...]
> cost/whatever" is currently addressed by making those the first two in
> tests[].  As to "size", if it means plaintext password length, then yes,
> it appears we currently lack the ability to lock it to a fixed value.
> We do have the ability to benchmark separately for passwords shorter
> than some length and longer than it, but only when we're not also
> benchmarking for one vs. many salts - see the logic around
> params.benchmark_length in bench_set_keys().  I initially introduced
> this for the AFS format, which differs greatly for length up to 8 vs.
> longer (with longer being much faster).

For formats like RAR3 it's really about longer passwords being
exponentially slower and we have wanted to settle for an exact length
for comparison with eg, cRARk. For others it's more about *different*
lengths in a batch being slower regardless of absolute length.

If we do use mask mode, problem is solved right there. Perhaps we should
just add the mask fields for now and see where they take us.

> While we could add a length lock feature, I think it doesn't fully solve
> the problem with benchmarking some of the recent formats, especially the
> PHC finalists where caching potentially plays a great role.  On a GPU,
> we should be benchmarking with tens of thousands of _different_ test
> vectors, not just repeat a few same suitable-length test vectors over
> and over.
>> A POC can be seen here:

(the above includes:)

struct fmt_tests {
        char *ciphertext, *plaintext;
        char *fields[10];
        int skip_on_bench;
        char *mask_plain;
        char *mask;

> I have mixed feelings about this.  If we introduce an extra field to
> each test vector, it better be a flags field where we can add more flags
> later if we need to.

Good point. If we add that field at all, the "int skip_on_bench;" would
be "int flags;" and bit 0 of it could be "skip in benchmark".

> Also, I am not sure if the proposed mask* fields address the caching
> issue I mentioned above or not. 

Instead of re-using a few different plaintexts over and over again, we
let mask mode produce unique ones. Sounds good to me. What problem would

> We need to address this issue even for slow hashes where it doesn't
> make sense for the format to bother providing its own mask mode
> support.  Will this mask be used with our generic mask support code
> in such cases?

Yes, we'd use mask mode to produce plaintexts for any format that
defines those fields.


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.