[RC5] Tree search algorithms and the next d.net

Adam Zilinskas AZilinskas at SolutionsIQ.com
Tue Apr 20 15:55:22 EDT 1999


From: David Cantrell:
>This suffers from the same problem as we found when talking about 
>distributed graphics rendering.  The customers DEMAND a particular
> turn-around time; they DEMAND reliability. 

I guess that I am thinking of a co-op type system. 
There are times you want to crunch a big problem
 but can't afford to own a Cray-equivalent. You can 
 afford a $10,000 system (or pick a reasonable price).
 You can get Cray-like performance by teaming up with
 a thousand other machines like yours. The pay back to
 the group is that when you are finished, your machine 
 is idle so that others can make use of it in exchange
 of using their CPU cycles.

Now I guess its like using city streets. Its a communal
 property and works well when each person needs it only 
 for a period of time. There will be occasional traffic 
 jams when everyone wants to use it at the same time, but
 if planned right we generally get good access to the
 resources.

Now if there is a hog, someone that will need many CPU-years
 from many machines to work their problem. They will have to 
 either do their own system or compensate the others for 
 the resources the hog uses and does ot replace. The street
 equivalent is now putting a mining company on your street 
 and the street is always full of dump trucks entering and
 leaving the mine. No one else can make use of the street.
 The best thing would be for the mining company to make
 their own private road and stay off the communal one, or
 pay to expand this communal street to make more room.

Back to the $10,000 example, doing a problem on it 
 alone may take thousands of CPU minutes to solve.
 We could buy a $10,000,000 system and solve the
 problem in a few minutes. Using to co-op network,
 we spread the thousands of CPU minutes compute time
 over thousands of machines (given the problem can
  be done that way). We get the effect of the $10,000,000
  machine while only spending $10,000. 
Now if we become a hog, we have such a big problem 
 or submit so many problems a day that we take alot
 of CPU cycles but ever sit idle long enough to give
 any back. We should compensate the others for all the 
 sucking we are doing (pay them like a utility) or 
 break down and buy enough machines to do our network
 dedicated for our own uses and our demands.

If everyone understood they are getting something for 
 a little, and like rush hour traffic, accept that some
 times they will have to wait fro traffic to clear. 
 They might accept being part of the co-op.

Now about distributed work units. (Greg Hewgill)
This is where alot of design work would be needed.
I am thinking on an idea like the Java sandbox where
 work unit cores will, by designed, protect the data
 within the unit as well prevent the core code from 
 violating the host.

Somehow work units are in some sort of consistent
 blocks that are checked out of hubs. The core code
 for that work unit will also be downloaded if it 
 is not already cached. The core code works on the
 data and checks the block back in. Security will 
 happen by some sort of mild encryption of the blocks
 as well as each block cotnains only tiny parts of
 the problem. Now for redundancy, the network
 will have a scheme that if a checked out block 
 does not come back within a reasonable time, it
 is assumed to have been lost and the block is 
 unlocked to be tried by another client. This might 
 rule out offline clients that check out blocks 
 and posts them back in 20 days later. 

I can also see a development kit for core code.
It helps people write their tasks to fit the 
 co-op network scheme. The key piece of this
 development kit will be a analyzer that reviews
 that the code to insure that it stays within the 
 work block sandbox (that is, an automated review
 of the code that it is allowable to run on other 
 members of the network).
 

                  Adam Zilinskas
                  Solutions IQ
                  azilinskas at solutionsiq.com



--
To unsubscribe, send 'unsubscribe rc5' to majordomo at lists.distributed.net
rc5-digest subscribers replace rc5 with rc5-digest



More information about the rc5 mailing list