Can overclocking skew the scientific results?

[cbuchner1]

Hi,

running the CPU or GPU clients on overclocked machines may lead to erroneous computations, sometimes to EUEs and loss of work units. What kind of safeguards are in place to ensure that results delivered by a client are accurate?

For example BOINC uses two machines to compute the same result and only awards the points when the result is mutually verified. How does Folding@Home address this issue?

Christian

[John Naylor]

I don't know how completed units are checked, but units are protected from tampering by checksums - if the program detects that one file has been changed in the work unit data since the client was last shut down (or since the last checkpoint was written) then the work unit is automatically started over (this also happens if files are corrupted by a bad shutdown). That prevents tampering during processing, and to prevent bogus results being fabricated and sent in, the project uses a 2048-bit digital signature for transmission between client and server. This rules out fabricated results or file tampering from the list of possible reasons why returned units (completed or otherwise) cannot be trusted.

If a work unit is returned as an Early Unit End, then it is reissued to a different computer, and if that machine also EUEs at the same point then that is assumed to be the natural ending point of that series of units (so one particular RCG of a project has completed). However if a unit is EUE'd by one machine, then upon reissuing another machine processes the unit and EUEs at a different point, then the unit is reissued again to a third machine, and again and again upon the unit's return until either the unit is completed or two EUEs match (probably they have to have run longer than any other set of EUEs before it is accepted).

Work units are never "lost" so to speak - if a unit has to be deleted then when its preferred deadline passes it will be reissued automatically to a different system.