There is a problem with credit inflation. If we change the long term resource share to be based on granted credit, then the credit granted will tend to normalize. There are a few open questions on exactly how to do this, but it should be possible.
-----Original Message----- From: boinc_dev [mailto:[email protected]] On Behalf Of David Anderson Sent: Tuesday, August 16, 2016 3:02 PM To: Christian Beer <[email protected]>; Jason Groothuis <[email protected]>; [email protected] Subject: Re: [boinc_dev] 4th Gen BOINC credit system The credit calculation is encapsulated in a single function, assign_credit_set(). You can change policy by changing or replacing this function. Maintaining runtime statistics (the host_app_version table) is used for job runtime estimation as well as the credit system. It needs to stay even if you change credit. -- David On 8/16/2016 7:40 AM, Christian Beer wrote: > I've tried to modularize the runtime estimation and credit calculation > system (RECCS) in a local branch so we could more easily switch between > implementations of different credit systems (on a per app level). But > because CreditNew is very deeply involved in several key components I > find it hard to establish an interface that is generic enough to > decouple the RECCS system from the rest of BOINC. > > Such a modularization would also help in building a simulator that can > use different strategies (credit/runtime systems) just as they are used > in the scheduler. > > I've halted this for now because of lack of time but I would be willing > to share what I have and write a Design document on how to modularize it > further. > > Regards > Christian > > On 15.08.2016 08:18, Jason Groothuis wrote: >> I concur with these statements after having studied the mechanism from an >> engineering perspective over the last couple of years. Steady state and new >> host/app were observed, on seti/seti-beta, along with fresh app launch on >> single appversion regimen with the kind help of Albert@home staff (before me >> running out of resources to take the study further). >> >> >> The fundamental design concepts weren't found to be problematic, though >> some of the implementation assumptions and choices were. Outcomes were that >> the dominant instabilities were (engineering) control systems related, and >> that characterisation of the existing system (via matching such a simulation >> to real-world) would aid communication of the problems and potential >> solutions. >> >> >> What is a relatively simple engineering task of estimate localisation, is >> quickly conflated with passions when Credit is mentioned. That quickly >> mires the genuinely solvable issues in Inertia. >> >> >> So In my view, progress could be made by: >> >> - simulate the existing mechanism, under original CPU-FPUonly state, >> CPU-SIMD enabled state, GPU only state, and para-metrically mixed state >> >> - Use that (refined) simulation as reference comparison to real-world, and >> to formalise potential practical solutions. >> >> - Change the name from CreditNew to 'estimator'; or somesuch >> >> >> >> >> >> >> >> ------------------------------------------------------------------------------------------------------ >> Jason Richard Groothuis >> ------------------------------------------------------------------------------------------------------ >> >> >> ________________________________ >> From: boinc_dev <[email protected]> on behalf of David >> Anderson <[email protected]> >> Sent: 15 August 2016 08:14 >> To: [email protected] >> Subject: Re: [boinc_dev] 4th Gen BOINC credit system >> >> My thoughts: >> >> 1) The basic ideas behind the current credit system are sound, >> but there seem to be some problems with the details, >> in particular what happens when new app versions are added. >> >> 2) We need a simulator that models credit calculations for various scenarios >> (synthetic, or based on trace data from a project). >> Such a simulator would help identify the problems with the current system, >> or would demonstrate that a new scheme works better. >> I don't think we should consider a new system without simulator-based >> validation of it. >> >> -- David >> >> >> On 8/14/2016 9:06 AM, CM wrote: >>> Hey, >>> >>> A '4th gen BOINC credit system' thread was created over at the official >>> BOINC >>> forums: https://boinc.berkeley.edu/dev/forum_thread.php?id=10953 >> [Discussion] 4th Generation BOINC credit >> system<https://boinc.berkeley.edu/dev/forum_thread.php?id=10953> >> boinc.berkeley.edu >> I originally started a discussion over at cryptocointalk about this: >> https://cryptocointalk.com/topic/46130-discussion-3rd-generation-boinc-credit-system/ >> >> >> >>> There's been about 77 posts to the thread, the general consensus was that a >>> new >>> BOINC credit system is a good idea but thus far the definition/specs of >>> such a >>> next gen system has not been agreed upon. >>> >>> What are the BOINC dev's thoughts on this topic? I'd love to continue this >>> discussion & work towards a next-gen BOINC system. >>> >>> Cheers guys :) >>> >>> _______________________________________________ >>> boinc_dev mailing list >>> [email protected] >>> http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev >>> To unsubscribe, visit the above URL and >>> (near bottom of page) enter your email address. >> _______________________________________________ >> boinc_dev mailing list >> [email protected] >> http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev >> To unsubscribe, visit the above URL and >> (near bottom of page) enter your email address. >> _______________________________________________ >> boinc_dev mailing list >> [email protected] >> http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev >> To unsubscribe, visit the above URL and >> (near bottom of page) enter your email address. > _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address. _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
