Hi David,
> I wish that GPUgrid (and Rosetta, Einstein etc.) would upgrade their server
> code.
> It's easier than they think it is, and it would make their lives easier.
As I have explained before, we at Einstein@Home would like this too. We have
tried and tried on our test project, but we can't get the credit system to work
correctly for us.
Cheers,
Bruce
> On Feb 14, 2015, at 7:54 AM, David Anderson <[email protected]> wrote:
>
>
> On 13-Feb-2015 4:14 PM, Jacob Klein wrote:
>> Hallelujah! This is the right call. Now, when a GPU task needs a full CPU
>> core to
>> keep itself fed, it won't have to compete against the CPU-only tasks.
>> Hurray! ::does
>> funky nerd dance:: Now POEM can properly set their GPU apps to use 1.0 CPU,
>> and
>> it'll work correctly!
>>
>> Now, just so you know, I believe that GPUGrid is breaking new ground with
>> old server
>> code. They have a new beta app they're testing, where the app itself is set
>> up as an
>> MT app, but it also requires an AMD GPU to run it. It's the first ever "MT +
>> GPU"
>> hybrid app that I've ever seen or heard of, and I think they're using the
>> MT-style
>> "app plan class shtuff etc." to pull it off. And they're not going to
>> upgrade server
>> code any time soon, from what they've said.
>>
>> I don't have an AMD GPU, and so I haven't tested it, but I wanted you to be
>> aware of
>> its existence. It's an oddball. Perhaps one day I'll get an AMD GPU in this
>> rig, so
>> I can test it. I love testing oddballs.
>>
>> Thanks for making the code change here.
>>
>> Sincerely,
>> Jacob
>>
>>
>>
>>
>> > Date: Fri, 13 Feb 2015 16:07:00 -0800
>> > From: [email protected]
>> > To: [email protected]; [email protected]
>> > CC: [email protected]
>> > Subject: Re: [boinc_dev] set app priority
>> >
>> > I checked in a change that will run coprocessor jobs at BELOW_NORMAL,
>> > even if they use >= 1 CPU.
>> > -- David
>> >
>> > On 29-Jan-2015 1:28 PM, Jacob Klein wrote:
>> > > David,
>> > >
>> > > I expect coprocessor jobs (including gpu jobs), to be set at a higher
>> > > priority than
>> > > cpu jobs, in order to keep those "other devices" fed. It has a
>> > > meaningful impact on
>> > > the task completion times, for GPU tasks as well as ASIC tasks.
>> > >
>> > > I still maintain that it should be set as:
>> > > If (coprocessor job)
>> > > ... set process priority to BELOW_NORMAL_PRIORITY_CLASS (which I think
>> > > uses default
>> > > thread prio of 6 on Windows)
>> > > Else (cpu job)
>> > > ... set process priority to IDLE_PRIORITY_CLASS (which I think uses
>> > > default thread
>> > > prio of 4 on Windows)
>> > >
>> > > We already do this today, except as Richard mentioned, we are
>> > > erroneously making
>> > > "full-cpu-core GPU tasks" get bumped to the lower priority, affecting my
>> > > ability to
>> > > keep my GPUs loaded, and thus affected task completion time.
>> > >
>> > > Could we please change it to "detect that it's a coprocessor job"
>> > > somehow, when
>> > > determining priorities, per my recommendation above?
>> > > I understand your concern about this, but keep in mind these priorities
>> > > are all
>> > > lower than NORMAL_PRIORITY_CLASS.
>> > >
>> > > So, can we please fix it?
>> > >
>> > >
>> > >
>> > >
>> ------------------------------------------------------------------------------------
>> > > Date: Thu, 29 Jan 2015 20:59:06 +0000
>> > > From: [email protected]
>> > > Subject: Re: [boinc_dev] set app priority
>> > > To: [email protected]; [email protected]
>> > > CC: [email protected]
>> > >
>> > > Maybe we're simply using the wrong discriminant to choose which priority
>> > > to set.
>> > >
>> > > You note that some GPU apps perform poorly at idle priority.
>> > >
>> > > But at present we're determining their GPU-ness, not directly, but
>> > > indirectly by
>> > > their fractional CPU assignment.
>> > >
>> > > There are other GPU apps, apart from the POEM example, where a full 1.00
>> > > CPU
>> > > assignment makes a significant difference in processing time - the
>> > > Einstein@Home
>> > > OpenCL application for imtel_gpu is a case in point.
>> > >
>> > > Would it not be possible to determine the required priority directly
>> > > from the 'use
>> > > GPU' definition instead, and ignore the CPU fraction?
>> > >
>> > >
>> ------------------------------------------------------------------------------------
>> > > *From:* David Anderson <[email protected]>
>> > > *To:* Jacob Klein <[email protected]>
>> > > *Cc:* BOINC Development <[email protected]>
>> > > *Sent:* Thursday, January 29, 2015 8:42 PM
>> > > *Subject:* Re: [boinc_dev] set app priority
>> > >
>> > > Jacob:
>> > > BOINC runs CPU jobs at idle priority to avoid impacting
>> > > the performance of non-BOINC programs.
>> > > The rationale for running jobs that use < 1 CPU at higher priority is
>> > > that
>> > > - because they use little CPU time, running them at higher priority
>> > > won't have much impact on other programs
>> > > - some apps (like QCN, and GPU apps) perform poorly if run at idle
>> > > priority
>> > >
>> > > We could change the policy to run coprocessor apps at higher priority,
>> > > regardless of how much CPU they use.
>> > > But that would impact system performance in some cases.
>> > > I'd hesitate to make this the default.
>> > > Maybe it could be a cc_config.xml option.
>> > >
>> > > -- David
>> > >
>> > > On 26-Jan-2015 2:23 PM, Jacob Klein wrote:
>> > > > Hi David,
>> > > >
>> > > > I've been running POEM@Home GPU tasks, with an app_config.xml file,
>> > > > with
>> > > <cpu_usage>
>> > > > set to 1.000, since each of their GPU tasks uses a full core, and I
>> > > > want to
>> > > > appropriately budget resources on my system.
>> > > >
>> > > > However, when I do that, the task's process runs at Idle priority,
>> > > > instead of
>> > > Below
>> > > > Normal.
>> > > >
>> > > > Martin (a dev at Poem@Home) is researching the issue, here:
>> > > > http://boinc.fzk.de/poem/forum_thread.php?id=1105&postid=10200
>> > > > And I replied here:
>> > > > http://boinc.fzk.de/poem/forum_thread.php?id=1105&postid=10201
>> > > >
>> > > > ... but I think it's being caused by your changes below.
>> > > > Is your logic below, possibly incorrect? I mean, to my knowledge,
>> > > > there are
>> > > some GPU
>> > > > apps (especially OpenCL) which require a full CPU core, and we'd still
>> > > > want their
>> > > > priorities to be kept at Below Normal, to keep the GPU fed with
>> > > > kernels, as a
>> > > > priority over other CPU jobs. I could even imagine non-GPU coprocessor
>> > > > jobs that
>> > > > would also require a full CPU core.
>> > > >
>> > > > Thoughts?
>> > > >
>> > > > Thanks,
>> > > > Jacob Klein
>> > > >
>> > > >
>> > > >
>> > > >
>> > > > > Date: Thu, 9 Oct 2014 14:49:49 -0700
>> > > > > From: [email protected] <mailto:[email protected]>
>> > > > > To: [email protected] <mailto:[email protected]>;
>> > > [email protected] <mailto:[email protected]>
>> > > > > Subject: Re: [boinc_dev] set app priority
>> > > > >
>> > > > > Currently, when the client runs an app:
>> > > > >
>> > > > > - if the app version uses < 1 CPU or has the is_wrapper flag set,
>> > > > > set the priority to BELOW_NORMAL_PRIORITY_CLASS
>> > > > > - otherwise set the priority to IDLE_PRIORITY_CLASS
>> > > > >
>> > > > > So GPU apps (if they use < 1 CPU), non-CPU-intensive apps, and
>> > > > > wrappers
>> > > > > are run at Below Normal.
>> > > > >
>> > > > > When the BOINC wrapper runs an app:
>> > > > >
>> > > > > - if <no_priority_change/> is set in the job description file,
>> > > > > the process inherits the priority of the wrapper
>> > > > > (normally BELOW_NORMAL_PRIORITY_CLASS)
>> > > > > - otherwise set the priority to IDLE_PRIORITY_CLASS
>> > > > >
>> > > > > There's no provision for anything other than Idle and Below Normal.
>> > > > >
>> > > > > (The above is Windows, of course; it's analogous on Unix).
>> > > > >
>> > > > > -- David
>> > > > >
>> > > > > On 09-Oct-2014 1:26 PM, Jacob Klein wrote:
>> > > > > > Does this support running the task's process at a priority that is
>> > > neither Normal
>> > > > > > (8) nor Idle (4)?
>> > > > > > GPU apps, for instance, are best run at Below Normal (6) (which
>> > > > > > GPUGrid
>> > > somehow
>> > > > > > does)... and I feel that ASIC apps also should run at Below Normal
>> > > > > > (6).
>> > > > > >
>> > > > > >
>> > > > > >
>> > > > > > > Date: Thu, 9 Oct 2014 12:38:16 -0700
>> > > > > > > From: [email protected] <mailto:[email protected]>
>> > > > > > > To: [email protected]
>> > > > > > > <mailto:[email protected]>
>> > > > > > > Subject: Re: [boinc_dev] set app priority
>> > > > > > >
>> > > > > > > To address this problem, I added a <no_priority_change/> option
>> > > > > > > in the wrapper's job description file:
>> > > > > > > http://boinc.berkeley.edu/trac/wiki/WrapperApp
>> > > > > > >
>> > > > > > > This tells the wrapper to not lower the priority of the tasks'
>> > > > > > > process.
>> > > > > > >
>> > > > > > > -- David
>> > > > > > >
>> > > > > > > On 09-Oct-2014 1:45 AM, Rebirther wrote:
>> > > > > > > > The <is_wrapper/> option doesnt work for me, the wrapper is
>> > > > > > > > still in
>> > > normal
>> > > > > > > > priority as before and the app in lowest priority running BOINC
>> > > client 7.4.22
>> > > > > > > >
>> > > > > > > > -Reb _______________________________________________ boinc_dev
>> > > mailing list
>> > > > > > > > [email protected] <mailto:[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] <mailto:[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] <mailto:[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.
-----------------------------------------------------------------
Prof. Dr. Bruce Allen, Director
Max Planck Institute for Gravitational Physics
Callinstrasse 38
D-30167 Hannover, Germany
Tel +49-511-762-17145
Fax +49-511-762-17182
Email: [email protected]
_______________________________________________
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.
