Hi!

>But I think you've chosen a poor example with the CPU-GPU speed 
relationship

Agreed, and I hope I didn't open a can of worms here because getting the 
RAM requirement handled better was my main concern :-) , e.g.  we have a 
factor of about 2 between the required RAM for x86 CPUs and ARM CPUs for 
one application. 
While the "speed up factor" for tweaking the runtime estimates especially 
for GPUs might not be ideal, it seems to work kind of reasonably well at 
least for our project, E@H. But that's a different issue.

Cheers
HB


-----------------------------------------------------------------
Heinz-Bernd Eggenstein
Max Planck Institute for Gravitational Physics
Callinstrasse 38
D-30167 Hannover,  Germany
Tel.: +49-511-762-19466 (Room 037)



From:   Richard Haselgrove <[email protected]>
To:     Heinz-Bernd Eggenstein <[email protected]>, BOINC 
Developers Mailing List <[email protected]>, 
Date:   07/23/2013 04:01 PM
Subject:        Re: [boinc_dev] RAM requirent depending on plan class



I think we need to be careful about the scoping validity of data, and we 
haven't always been in the past.

In this particular case, I think that it's a reasonable assumption that 
the RAM requirements of an application (again in the BOINC terminology) on 
one platform have a deterministic relationship with the RAM requirements 
on other platforms. All of that is knowable by the project team, and 
suitable for implementation on the server.

But I think you've chosen a poor example with the CPU-GPU speed 
relationship. There is no necessary or deterministic relationship between 
the speed of a GPU, and the speed of the CPU of the host it's mounted in 
(and to confuse matters further, there's no necessary relationship between 
the speeds of two different GPUs in the same host).

The current code assumes (as a starting point) that a generic GPU is 10x 
the speed of the associated CPU: I suspect that this assumption - along 
with a sub-optimal choice of plan_class by the server - is the underlying 
cause of a recent report of 'EXIT_TIME_LIMIT_EXCEEDED' errors at SETI 
Beta.

http://setiweb.ssl.berkeley.edu/beta/forum_thread.php?id=2008

From: Heinz-Bernd Eggenstein <[email protected]>
To: BOINC Developers Mailing List <[email protected]> 
Sent: Tuesday, 23 July 2013, 12:54
Subject: [boinc_dev] RAM requirent depending on plan class

Dear all,

I'd like to hear comments on the following problem/idea:

For Android on ARM, we at E@H have optimised our science app to require 
less RAM than our regular CPU app versions for x86 CPUs.

OTOH, we would like to have only one application (in the BOINC 
terminology) for all CPU platforms, so only one suite of work generators, 
validators etc is needed and x86 and ARM app versions get the same tasks 
and can even cross-validate against each other, which we think is good.

But as the memory requirements are associated to the work unit,  we have 
to use the maximum value for all the different app versions/ plan classes. 



IMHO the situation is somewhat similar to the execution speed differences 
for different plan classes that cannot be fully modelled by the BOINC 
benchmark results (e.g. for CPU-GPU comparison) and I understand that the 
solution for this was a scaling factor that can be configured with plan 
classes to let the scheduler know that results are computed faster by a 
certain factor for a particular plan class compared to what BOINC would 
assume otherwise. 

Would it be feasible/a good idea to also have a configurable scaling 
factor for RAM requirements, perhaps ?? I think many projects will find 
that optimising for RAM size is important for Android. 

Cheers
HB


-----------------------------------------------------------------
Heinz-Bernd Eggenstein
Max Planck Institute for Gravitational Physics
Callinstrasse 38
D-30167 Hannover,  Germany
Tel.: +49-511-762-19466 (Room 037)
_______________________________________________
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