[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <e3a707a14cb562270796840b6bc6ac69@smtp.hushmail.com>
Date: Thu, 28 Mar 2013 12:24:51 +0100
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: MPI-openCL build; fast hashes are faster!!
On 28 Mar, 2013, at 9:59 , Solar Designer <solar@...nwall.com> wrote:
> On Thu, Mar 28, 2013 at 09:45:13AM +0100, magnum wrote:
>> Please note that the benchmark output from MPI runs is unreliable in this case - for legacy reasons it's just the fastest individual figure multipled by number of nodes. Normally this is OK but in this case you'll likely get an inflated value.
>>
>> We could change this to a true sum of individual speeds
>
> We (you) should do that, please.
It was not by far as bad as I thought. What it did was take the max. values of the real and virtual times, and the sum of crypts. Normally this would produce correct averages even for heterogenous clusters, at least for "real" speed.
I fixed it anyway so now we do a proper average of times. More importantly, I fixed the reduction of crypts (composite int64 struct) for 32-bit archs. It would only sum the low 32 bits until now (and worse on BE - only the high ones) and possibly leave the other half uninitialized.
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.
