|
Message-ID: <20111219121125.GB4696@openwall.com> Date: Mon, 19 Dec 2011 16:11:25 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: Re: DES BS + OMP improvements, MPI direction On Sun, Dec 18, 2011 at 01:29:02AM -0700, RB wrote: > On Sat, Dec 17, 2011 at 15:46, Solar Designer <solar@...nwall.com> wrote: > > Are you suggesting that we'd apply MPI at the same low level where we > > currently apply OpenMP? No, that doesn't sound like a good idea to me, > > except maybe for the slowest hash/cipher types only (BF, MSCash2, RAR, > > crypt when run against SHA-crypt), where it might be acceptable. > > Yes, it is what I'm suggesting. Obviously anything that completes in > less than the RTT of the MPI environment would need to be handled > differently (if at all), but the problem the MPI implementation has > historically had is that it splits work at too high a level. Last I > checked it was splitting work based on individual work steps - each > thread took a given length x count and worked it by itself on that. > For very fast hashes (like LM) I'm sure that's more acceptable, but > for slower ones and with larger MPI implementations (hundreds to > thousands of threads), the network very quickly gets to the point that > every thread is working on an effectively impossible-to-solve step, > which really isn't the end-user's intent of parallelization. I think you don't literally mean "every" above (as that would imply that all easier "steps" were already completed, so everything is as desired), but I get what you're referring to. Yes, there's this drawback of the way the current MPI implementation splits the work for incremental mode. We need to address it. However, I think we should do it without incurring the drawbacks of OpenMP-like low-level parallelization. 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.