A few more thoughts.

For the purpose of reliability, a host should be considered to have failed at a 
task if it is aborted, generates a runtime error, reaches the deadline without 
being returned, or fails to validate.  If the project is down at the time of 
the deadline, the deadline should be extended to be the time that the project 
comes back up + 24 hours, or the time between connections specified by the 
server.  This avoids having reliable clients dinged because of a server problem.

Perhaps reliability should be based on a reducing average with a 2 week half 
life.  Yes, there can be some lag in the calculation because tasks can be 
validated much later than the submission.  However, a host that fails to 
validate most of its tasks cannot be considered to be reliable.

-----Original Message-----
From: [email protected] 
[mailto:[email protected]] On Behalf Of McLeod, John
Sent: Thursday, August 30, 2012 10:55 PM
To: David Anderson; [email protected]
Subject: Re: [boinc_dev] resent WUs added to the end of the WU queue

Give every host a rating of 0 to 1 on reliability.  Hosts that fail everything 
get a 0.  Hosts that return everything correctly on time get a 1.

Reliability should have a "recent" component to it.  If I had a bad 
configuration and was failing everything two months ago, and am getting 
everything right now, the problem 2 months ago should have little effect on now.

The older the task is, the higher the probability of handing it out to a lower 
reliability client should be (to avoid starvation).  However, a very untrusted 
client should still have a much lower chance of getting the task.

If the client has the required files for locality, it should get a higher 
chance.

If you have special tasks on hand, and a host connects, give the special tasks 
if a random number is < f(reliability, locality, age of task, time from 
committing a task till deadline).  

f(double reliability, bool localityl, double age, double deadline) might look 
something like:

reliability ^2 * (locality ? 1.0 : 0.5) + age / deadline;

If the task is new, it is all about reliability and locality.  As the task 
ages, it becomes all about age.


-----Original Message-----
From: [email protected] 
[mailto:[email protected]] On Behalf Of David Anderson
Sent: Thursday, August 30, 2012 3:40 PM
To: [email protected]
Subject: Re: [boinc_dev] resent WUs added to the end of the WU queue

In the shared-mem job cache framework, server-side scheduling has 2 parts:
- the order on which jobs are put in the cache
- the policy for selecting jobs from cache to send
   to a particular host.

This part of BOINC has evolved in a piecemeal fashion,
and is due for a remodel.
Plus it needs to support optimized batch completion,
which is something we've talked about but not done anything.

I created a wiki page with some notes on this:
http://boinc.berkeley.edu/trac/wiki/JobPrioritization
Comments/ideas welcome.

-- David

On 29-Aug-2012 2:19 PM, Travis Desell wrote:
> See:
>
> http://volunteer.cs.und.edu/subset_sum/forum_thread.php?id=47
>
> Is there any kind of project setting which will have resent WUs added to the
> beginning of the WU queue (so they are sent out quicker for validation)?
>
> thanks,
>
> --Travis
>
> ---------------------------------------------------------------------------
> Travis Desell,  Assistant Professor University of North Dakota - Dept. of
> Computer Science [email protected] - cell: 518-867-1054 Streibel Hall
> Room 220 - office: 701-777-3477 3950 Campus Road Stop 9015 Grand Forks, North
> Dakota 52802-9015
>
> Homepage ( http://people.cs.und.edu/~tdesell/ ) MilkyWay@Home (
> http://milkyway.cs.rpi.edu/ ) DNA@Home ( http://volunteer.cs.und.edu/dna )
> Worldwide Computing Laboratory ( http://wcl.cs.rpi.edu/ )
> ----------------------------------------------------------------------------
>
> _______________________________________________ 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.

Reply via email to