Paul D. Buck wrote:
> 
> On Sep 25, 2009, at 8:34 AM, Lynn W. Taylor wrote:
> 
>> Even I can figure out that returning work late would have a negative 
>> effect on credit.
> 
> And I never suggested that it would not.  Variety in the queue and in 
> the work selected to be run on the processor has nothing to do with 
> deadlines.  It has to do with Resource Share a bad allocation system 
> that has no meaning to a person that is single project oriented and we 
> are now throwing under the bus in a variety of ways and pretending we 
> aren't.  

Variety in the cache at any given instant is simply not workable.

If I have two projects, and a 10000:1 resource share, there is no way 
that my 0.01% project can always have a work unit in the cache.

It only gets worse when you have more than two.

In my opinion, the problem comes from trying to always keep the cache 
full.  If people set their cache to four days, and BOINC has to go to 
the most-overworked project to get work, then we've got what we've got.

You simply cannot have everything.  You can't have a perennially full 
cache, and have all projects represented in that cache all at the same time.

You have to let go of something.

In my opinion, half of the cache should be set aside for the project 
that is owed the most.  That means if you're crunching for LHC (which is 
perenially out of work) and all of the other projects have been 
overworked, your four day cache is a two-day cache.

Maybe it shouldn't be half, maybe it's a formula based on average debt, 
but there has to be room for the "most needed" work, or you run the risk 
of not being able to get "x" when "x" has work.

... and unless I've missed something pretty major, it all works off of 
actual CPU time, and benchmarks aren't a factor in the long-term.

At this point, that's my OPINION.

Until I have time to sit down and code something, to demonstrate that 
I'm right, I'm not going to berate those who are working very hard to 
keep BOINC functional -- and actually doing the work.
_______________________________________________
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.

Reply via email to