I'd prefer to figure out why the static estimates are off.
If an app's jobs are of a size proportional to wu.rsc_fpops_est,
the static estimates should be almost exact, even for a host's first jobs.
If there's a bug, let's fix it rather than hide it.
-- D

On 13-Feb-2014 10:41 PM, Rytis Slatkevičius wrote:
This is exactly why Jon and Mike are proposing two different modes of runtime
estimation: for linear projects (probably applications only, as a project can 
have
multiple apps), add a flag to mark the application as linear; the remaining 
projects
would use the same method as now. Since the new developers probably won't know 
about
the flag, the current estimate should be the default.

David, I want to make sure: is it that you oppose the suggestion itself, or is 
the
time constraints that stop you from implementing it? If the latter, we could 
submit
patches.



--
Pagarbiai / Sincerely
Rytis Slatkevičius
+370 670 77777


On Fri, Feb 14, 2014 at 4:54 AM, David Anderson <[email protected]
<mailto:[email protected]>> wrote:

    I want to make sure everyone realizes that some apps inherently can't
    supply an accurate fraction done.
    It's easy if your app has the form

       for i=1,N
         fixed computation

    It not easy if your app has the form

       lengthy preprocessing
       for i=1,n
         fixed computation
       lengthy postprocessing

    or

       while true do
         fixed computation
         if convergence criterion met
           break

    _________________________________________________
    boinc_dev mailing list
    [email protected] <mailto:[email protected]>
    http://lists.ssl.berkeley.edu/__mailman/listinfo/boinc_dev
    <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