Вторник, 16 апреля 2013, 10:41 -07:00 от David Anderson
<[email protected]>:
>The plan class mechanism lets you specify whatever policy you want.
>The predefined (sample) classes don't cover all possible policies.
>-- David
Could you do SETI@home project some favor and devise some example how to
achieve such config, i.e. (being more generic):
1) if accelerator supports 2 APIs and there are apps for both APIs send work
only for one API (customizable) on such device.
2) if accelerator supports only 1 API and there is app that uses that API -
send work for that app.
This would solve particular problem we have with CAL/OpenCL now + this will be
good example how to achieve such thing for CUDA/OpenCL mixes too and can be
actual for OSX -based configs too.
To have example is very good way to learn how existing functional can be
adopted for particular needs.
>
>On 16-Apr-2013 10:31 AM, Eric J Korpela wrote:
>> Unless things have changed since January (the last time I looked) there is no
>> existing plan class that will send CAL apps to CAL capable cards only if they
>> don't also have OpenCL. Maybe they're not supposed to, but they do. If the
>> scheduler believes (erroneously or not) that the CAL application is faster,
>> it
>> will send it. That, and the lack of easy customizability is why I've stopped
>> using the standard plan classes in favor of plan_class_spec.xml .
>>
>> I was planning on adding that capability to the plan_class_spec style plan
>> classes, but have not yet done so. I've also been planning on trying to
>> fully
>> understand the app selection logic and why it sometimes uses processing rate
>> (good) and other times uses elapsed time (bad) to choose one app over
>> another.
>> The app selection is sufficiently bad that I need to put a lot of time into
>> trying to get the right apps to the right GPUs with the right drivers rather
>> than letting processing rate be the decider.
>>
>>
>>
>> On Tue, Apr 16, 2013 at 10:14 AM, David Anderson < [email protected]
>> <mailto: [email protected] >> wrote:
>>
>>
>>
>> On 16-Apr-2013 2:12 AM, Raistmer the Sorcerer wrote:
>>
>> How to properly configure project server to send Brook+ app ONLY to
>> GPUs
>> with
>> CAL support but w/o OpenCL support and OpenCL app to all OpenCL GPUs
>> ? Is
>> this possible in current BOINC implementation ?
>>
>>
>> Yes; see
>> http://boinc.berkeley.edu/__trac/wiki/AppPlan
>> < http://boinc.berkeley.edu/trac/wiki/AppPlan >
>>
>> -- David
>>
>> _________________________________________________
>> 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.
_______________________________________________
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.
