Re: SSL trouble

[hallam@w3.org]

There are more effective solutions than simple random search, these have been 
known in the distributed processing arena for years.

What you effectivelly have is a farmed solution to a problem with a high degree 
of trivial parallelism. Farms always suffer from the server bottleneck problem. 
The alternative is to use a multifarm, its a bit complicated to explain bu the 
essence is that you distribute the farmming mechanism. The most extreeme example 
of this is to have every slave also act as a master for some part of the problem. 
Since the bandwidth/processing ratio is unfavourable it would be better to have a 
small but non trivial (5-10) number of master controllers.


The basic principles are to leverage pipelined parallelism, a slave does not 
simply ask for a chunk of keyspace, process it, return results and ask for the 
next chunk. Instead overlap work packages, give them more than one to work at at 
once so that the system does not suspend waiting on the server.

Size the chunks adaptively, the more keyspace a processor works through the more 
packets it is given at once.

Use integrity checks to ensure that the slaves are acting properly. One method of 
doing this is to keep secret part of the known plaintext (say 16 bits). A slave 
is required to report _all_ matches in the range to the master. Slaves who report 
a statistically low number of matches may be considered suspicious. It is a 
simple matter to allocate part of that keyspace to another processor for a 
double-check.

	[Its so obvious I'll apply for a patent on that technique]

Another usefull technique is to require the slave to checksum some collateral 
result from the calculation mix. Then if its simply braindead software it can be 
detected.

When running a multi-master farm it is important to realise that the slaves serve 
all the masters, not just a single one. Masters can distribute work chunks 
amongst themselves in larger chunks, as chunks are completed this is communicated 
to the other workers.

If we used the Web as a substrate for this work the control software could then 
be used for other related tasks requiring large scale parallel processing on 
networked workstations.  This was one of the original applications I looked at 
back in 1992 when I was doing an awful lot of this type of work.


		Phill Hallam-Baker