My apologies, I must have confused your setup with someone that of someone
else.
There are two reasonable fixes for the DCF that I can see.
1) Have DCF per application - with new applications getting an average of
the DCFs for existing applications for the project as a starting point.
2) Have DCF be the mean correction for work fetch and the mean + 3
standard deviations for CPU scheduling. (this is a very different
calculation than it is now).
The current DCF is a good safety for preventing late work, but it does tend
to under fetch sometimes.
jm7
Mikus Grinbergs
<[email protected]>
Sent by: To
<boinc_dev-bounce [email protected]
[email protected] cc
u> [email protected]
Subject
Re: [boinc_dev] BOINC 6.10 release
10/27/2009 01:25
PM
> If I recall correctly, you have refused to set the "Connect Every X" and
> "Extra Work" settings to what we have advised to allow your computers to
> work. My memory is that "Connect Every X" is set to 0, and "Extra Work"
is
> set to something like 10. If you would change "Connect Every X" to a
> value between 1 and 2, I expect that your woes about not enough work to
> make it through the day would be reduced.
I believe you have confused me with someone else. I do not __ever__
recall you advising me how to set the "Connect every X" and "Extra
Work" settings.
My "big" system (8 core) has <work_buf_min_days>1.1 and
<work_buf_additional_days>1.4.
These values have been "tuned" to result in the ready queue
typically containing 28-40 hours of work after a download.
The system with the 32-bit client has <work_buf_min_days>1.0 and
<work_buf_additional_days>3.0.
These values have not been "tuned" - since here workunits have
wildly varying run times (within each project). Even with the total
ready queue size being four days, at least once a month this system
goes idle sooner than 24 hours after a download. [That situation
appears to be caused by the DCF ramping-up quickly when a single
"grossly-underestimated" WU occurs, but ramping-down slowly despite
that WU being followed by many "normally-estimated" WUs.]
My other systems have total work queue sizes between those given
above. All have <work_buf_min_days> (in global_prefs_override.xml)
around 1.0. And I have *never* set <work_buf_additional_days>
particularly large - you are mistaken about me there.
mikus
_______________________________________________
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.
