sharing DL_DIR?

sharing DL_DIR?

[Rich Pixley]

[-- Attachment #1: Type: text/plain, Size: 1271 bytes --]

Is it reasonable to expect to share DL_DIR between multiple builds?

That is, are downloads properly locked so that multiple concurrent 
downloads of the same file won't collide?  And if so, are they NFS safe 
locks?

Or must each build download it's own copies of every component?

I ask because our, (Palm), branch has been, but it's a bit of a nuisance 
to pick a suitable locking mechanism that is both functional and performant.

  * Flock/lockf aren't reliably supported in heterogeneous environments.
  * The old lockfiles don't work over NFS. Lock directories apparently
    do, but trying to clean these up after interruptions or failures is
    a losing battle so if we do use these, we end up with orphan locks
    periodically.
  * With NFS as a possibility, we can't assume any kernel local IPC
    mechanism, so sysV ipc is out.
  * It seems like a pretty huge overhead to try to create any sort of
    zeroconf overhead.

Our solution so far has been to use NFS lock directories for multiple 
machine builds and flock/lockf for single machine builds.  This mostly 
works, though it's not ideal, and it's significantly faster than 
downloading/copying all of the component source multiple times even over 
local mirrors.

--rich

[-- Attachment #2: Type: text/html, Size: 1660 bytes --]

Re: sharing DL_DIR?

[Tom King]

[-- Attachment #1: Type: text/plain, Size: 1783 bytes --]

I have a single downloads directory with symlinks for every target build I
have to that directory.  As long as I have the source already in there I
don't fetch  it again

Tom

On Wed, Feb 22, 2012 at 1:33 PM, Rich Pixley <rich.pixley@palm.com> wrote:

>  Is it reasonable to expect to share DL_DIR between multiple builds?
>
> That is, are downloads properly locked so that multiple concurrent
> downloads of the same file won't collide?  And if so, are they NFS safe
> locks?
>
> Or must each build download it's own copies of every component?
>
> I ask because our, (Palm), branch has been, but it's a bit of a nuisance
> to pick a suitable locking mechanism that is both functional and performant.
>
>    - Flock/lockf aren't reliably supported in heterogeneous environments.
>    - The old lockfiles don't work over NFS. Lock directories apparently
>    do, but trying to clean these up after interruptions or failures is a
>    losing battle so if we do use these, we end up with orphan locks
>    periodically.
>    - With NFS as a possibility, we can't assume any kernel local IPC
>    mechanism, so sysV ipc is out.
>    - It seems like a pretty huge overhead to try to create any sort of
>    zeroconf overhead.
>
> Our solution so far has been to use NFS lock directories for multiple
> machine builds and flock/lockf for single machine builds.  This mostly
> works, though it's not ideal, and it's significantly faster than
> downloading/copying all of the component source multiple times even over
> local mirrors.
>  --rich
>
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core
>
>

[-- Attachment #2: Type: text/html, Size: 2591 bytes --]

Re: sharing DL_DIR?

[Christopher Larson]

[-- Attachment #1: Type: text/plain, Size: 2380 bytes --]

I share it as well. I don't see locks being a big issue for most things, as the worst case is a re-download, not use of a corrupt or incomplete file, due to the completion stamps.

-- 
Christopher Larson


On Wednesday, February 22, 2012 at 4:39 PM, Tom King wrote:

> I have a single downloads directory with symlinks for every target build I have to that directory.  As long as I have the source already in there I don't fetch  it again
> 
> Tom
> 
> On Wed, Feb 22, 2012 at 1:33 PM, Rich Pixley <rich.pixley@palm.com (mailto:rich.pixley@palm.com)> wrote:
> > Is it reasonable to expect to share DL_DIR between multiple builds?
> > 
> > That is, are downloads properly locked so that multiple concurrent downloads of the same file won't collide?  And if so, are they NFS safe locks?
> > 
> > Or must each build download it's own copies of every component?
> > 
> > I ask because our, (Palm), branch has been, but it's a bit of a nuisance to pick a suitable locking mechanism that is both functional and performant.
> > Flock/lockf aren't reliably supported in heterogeneous environments. 
> > The old lockfiles don't work over NFS. Lock directories apparently do, but trying to clean these up after interruptions or failures is a losing battle so if we do use these, we end up with orphan locks periodically.
> > With NFS as a possibility, we can't assume any kernel local IPC mechanism, so sysV ipc is out.
> > It seems like a pretty huge overhead to try to create any sort of zeroconf overhead.
> > 
> > Our solution so far has been to use NFS lock directories for multiple machine builds and flock/lockf for single machine builds.  This mostly works, though it's not ideal, and it's significantly faster than downloading/copying all of the component source multiple times even over local mirrors.
> > --rich
> > 
> > _______________________________________________
> > Openembedded-core mailing list
> > Openembedded-core@lists.openembedded.org (mailto:Openembedded-core@lists.openembedded.org)
> > http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core
> > 
> 
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org (mailto:Openembedded-core@lists.openembedded.org)
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core
> 
> 



[-- Attachment #2: Type: text/html, Size: 3730 bytes --]

Re: sharing DL_DIR?

[Saul Wold]

On 02/22/2012 03:33 PM, Rich Pixley wrote:
> Is it reasonable to expect to share DL_DIR between multiple builds?
>
Absolutely, We share DL_DIR via NFS across multiple machines and builds.

> That is, are downloads properly locked so that multiple concurrent
> downloads of the same file won't collide? And if so, are they NFS safe
> locks?
>
Yes, as far as I know.

> Or must each build download it's own copies of every component?
>
No need.

> I ask because our, (Palm), branch has been, but it's a bit of a nuisance
> to pick a suitable locking mechanism that is both functional and performant.
>
>   * Flock/lockf aren't reliably supported in heterogeneous environments.
>   * The old lockfiles don't work over NFS. Lock directories apparently
>     do, but trying to clean these up after interruptions or failures is
>     a losing battle so if we do use these, we end up with orphan locks
>     periodically.
>   * With NFS as a possibility, we can't assume any kernel local IPC
>     mechanism, so sysV ipc is out.
>   * It seems like a pretty huge overhead to try to create any sort of
>     zeroconf overhead.
>
> Our solution so far has been to use NFS lock directories for multiple
> machine builds and flock/lockf for single machine builds. This mostly
> works, though it's not ideal, and it's significantly faster than
> downloading/copying all of the component source multiple times even over
> local mirrors.
>
I think we are using a very basic locking mechism that creates the 
lockfile and a donefile when complete.  It seems to work well enough.

Sau!

> --rich
>
>
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core

Re: sharing DL_DIR?

[Rich Pixley]

[-- Attachment #1: Type: text/plain, Size: 2698 bytes --]

What happens if two separate bitbake invocations both simultaneously 
attempt to download the same file?

--rich

On 2/22/12 15:41 , Christopher Larson wrote:
> I share it as well. I don't see locks being a big issue for most 
> things, as the worst case is a re-download, not use of a corrupt or 
> incomplete file, due to the completion stamps.
>
> -- 
> Christopher Larson
>
> On Wednesday, February 22, 2012 at 4:39 PM, Tom King wrote:
>
>> I have a single downloads directory with symlinks for every target 
>> build I have to that directory.  As long as I have the source already 
>> in there I don't fetch  it again
>>
>> Tom
>>
>> On Wed, Feb 22, 2012 at 1:33 PM, Rich Pixley <rich.pixley@palm.com 
>> <mailto:rich.pixley@palm.com>> wrote:
>>> Is it reasonable to expect to share DL_DIR between multiple builds?
>>>
>>> That is, are downloads properly locked so that multiple concurrent 
>>> downloads of the same file won't collide?  And if so, are they NFS 
>>> safe locks?
>>>
>>> Or must each build download it's own copies of every component?
>>>
>>> I ask because our, (Palm), branch has been, but it's a bit of a 
>>> nuisance to pick a suitable locking mechanism that is both 
>>> functional and performant.
>>>
>>>   * Flock/lockf aren't reliably supported in heterogeneous environments.
>>>   * The old lockfiles don't work over NFS. Lock directories
>>>     apparently do, but trying to clean these up after interruptions
>>>     or failures is a losing battle so if we do use these, we end up
>>>     with orphan locks periodically.
>>>   * With NFS as a possibility, we can't assume any kernel local IPC
>>>     mechanism, so sysV ipc is out.
>>>   * It seems like a pretty huge overhead to try to create any sort
>>>     of zeroconf overhead.
>>>
>>> Our solution so far has been to use NFS lock directories for 
>>> multiple machine builds and flock/lockf for single machine builds.  
>>> This mostly works, though it's not ideal, and it's significantly 
>>> faster than downloading/copying all of the component source multiple 
>>> times even over local mirrors.
>>>
>>> --rich
>>>
>>> _______________________________________________
>>> Openembedded-core mailing list
>>> Openembedded-core@lists.openembedded.org 
>>> <mailto:Openembedded-core@lists.openembedded.org>
>>> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core
>>>
>>
>> _______________________________________________
>> Openembedded-core mailing list
>> Openembedded-core@lists.openembedded.org 
>> <mailto:Openembedded-core@lists.openembedded.org>
>> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core
>

[-- Attachment #2: Type: text/html, Size: 5775 bytes --]

Re: sharing DL_DIR?

[Rich Pixley]

Excellent, thanks.  That's the news I was hoping for.

--rich

On 2/22/12 15:41 , Saul Wold wrote:
> On 02/22/2012 03:33 PM, Rich Pixley wrote:
>> Is it reasonable to expect to share DL_DIR between multiple builds?
>>
> Absolutely, We share DL_DIR via NFS across multiple machines and builds.
>
>> That is, are downloads properly locked so that multiple concurrent
>> downloads of the same file won't collide? And if so, are they NFS safe
>> locks?
>>
> Yes, as far as I know.
>
>> Or must each build download it's own copies of every component?
>>
> No need.
>
>> I ask because our, (Palm), branch has been, but it's a bit of a nuisance
>> to pick a suitable locking mechanism that is both functional and performant.
>>
>>    * Flock/lockf aren't reliably supported in heterogeneous environments.
>>    * The old lockfiles don't work over NFS. Lock directories apparently
>>      do, but trying to clean these up after interruptions or failures is
>>      a losing battle so if we do use these, we end up with orphan locks
>>      periodically.
>>    * With NFS as a possibility, we can't assume any kernel local IPC
>>      mechanism, so sysV ipc is out.
>>    * It seems like a pretty huge overhead to try to create any sort of
>>      zeroconf overhead.
>>
>> Our solution so far has been to use NFS lock directories for multiple
>> machine builds and flock/lockf for single machine builds. This mostly
>> works, though it's not ideal, and it's significantly faster than
>> downloading/copying all of the component source multiple times even over
>> local mirrors.
>>
> I think we are using a very basic locking mechism that creates the
> lockfile and a donefile when complete.  It seems to work well enough.
>
> Sau!
>
>> --rich
>>
>>
>> _______________________________________________
>> Openembedded-core mailing list
>> Openembedded-core@lists.openembedded.org
>> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core

Re: sharing DL_DIR?

[Richard Purdie]

On Wed, 2012-02-22 at 16:47 -0800, Rich Pixley wrote:
> What happens if two separate bitbake invocations both simultaneously
> attempt to download the same file?

The first will download it, the second will wait for it to finish, then
both will continue with the downloaded file.

Cheers,

Richard

Re: sharing DL_DIR?

[Rich Pixley]

On 2/23/12 02:16 , Richard Purdie wrote:
> On Wed, 2012-02-22 at 16:47 -0800, Rich Pixley wrote:
>> What happens if two separate bitbake invocations both simultaneously
>> attempt to download the same file?
> The first will download it, the second will wait for it to finish, then
> both will continue with the downloaded file.
How does the second know that the first is in progress?

--rich

Re: sharing DL_DIR?

[Richard Purdie]

On Thu, 2012-02-23 at 11:21 -0800, Rich Pixley wrote:
> On 2/23/12 02:16 , Richard Purdie wrote:
> > On Wed, 2012-02-22 at 16:47 -0800, Rich Pixley wrote:
> >> What happens if two separate bitbake invocations both simultaneously
> >> attempt to download the same file?
> > The first will download it, the second will wait for it to finish, then
> > both will continue with the downloaded file.
> How does the second know that the first is in progress?

It won't be able to take the lock and the done stamp file doesn't exist
yet.

Cheers,

Richard

Re: sharing DL_DIR?

[Rich Pixley]

On 2/23/12 12:11 , Richard Purdie wrote:
> On Thu, 2012-02-23 at 11:21 -0800, Rich Pixley wrote:
>> On 2/23/12 02:16 , Richard Purdie wrote:
>>> On Wed, 2012-02-22 at 16:47 -0800, Rich Pixley wrote:
>>>> What happens if two separate bitbake invocations both simultaneously
>>>> attempt to download the same file?
>>> The first will download it, the second will wait for it to finish, then
>>> both will continue with the downloaded file.
>> How does the second know that the first is in progress?
> It won't be able to take the lock and the done stamp file doesn't exist
> yet.
What lock?

I've just read through lib/bb/fetch2/__init__.py but I'm not seeing a 
lock.  What am I missing?

--rich

Re: sharing DL_DIR?

[Richard Purdie]

On Thu, 2012-02-23 at 12:36 -0800, Rich Pixley wrote:
> On 2/23/12 12:11 , Richard Purdie wrote:
> > On Thu, 2012-02-23 at 11:21 -0800, Rich Pixley wrote:
> >> On 2/23/12 02:16 , Richard Purdie wrote:
> >>> On Wed, 2012-02-22 at 16:47 -0800, Rich Pixley wrote:
> >>>> What happens if two separate bitbake invocations both simultaneously
> >>>> attempt to download the same file?
> >>> The first will download it, the second will wait for it to finish, then
> >>> both will continue with the downloaded file.
> >> How does the second know that the first is in progress?
> > It won't be able to take the lock and the done stamp file doesn't exist
> > yet.
> What lock?
> 
> I've just read through lib/bb/fetch2/__init__.py but I'm not seeing a 
> lock.  What am I missing?

Try searching that file for expressions like "lock" and "lockfile" ;-)

Cheers,

Richard