[Planetlab-devel] vc_ctx_migrate & nodemanager

[Thierry Parmentelat]

On Mar 25, 2008, at 7:40 PM, Daniel Hokka Zakrisson wrote:

> Hello Thierry,
>
> Quoting Thierry Parmentelat <thierry.parmentelat at sophia.inria.fr>:
>> Daniel, and all
>>
>> I've just installed a couple new (build and test) boxes under   
>> centos5.1 & vserver from your yum repo
>> I am now fully up2date and compliant - to the best of my knowledge  
>> -  with the recommended practices.
>> But I still see the ctx_migrate issue, more or less occasionally
>
> Really? I have not been able to reproduce it with util-vserver  
> 0.30.215, and I have no idea what might be racing if it's still  
> happening...
yes, I can reproduce the issue rather simply; I've been playing with a  
small bench that tries to run the vserver .. build command ten times  
in a raw, and have not seen it succeed ten times yet; now fortunately  
I've never seen if fail the ten times either :)
This is not a big concern for me, but again if it's helpful to you, I  
can give you access to one of those our test boxes.

>
>> at first glance I thought I had seen a way to reproduce more  
>> easily  but that was just coincidence
>> for the time being, I have only noticed that waiting for a short  
>> delay  (3 s) but more numerous times (20)  seems as effective as  
>> the 5 x 1mn  policy I had previously
>> (this is in v*-init-vservers.sh - that we use quite frequently  
>> during  the day as well)
>> I've also tried to make it easier to gather statistics on this  
>> problem;
>>
>> I am just wondering, as we are now eating our own food as Marc  
>> likes  to put it, whether we should implement the same kind of  
>> workaround in  the nodemanager
>> or is the logic of NodeManager flexible enough that it will retry   
>> later anyway if vuseradd fails ?
>
> vuseradd is _much_ simpler, and also doesn't use the yum build  
> method, thus avoiding rpm-fake entirely.
OK so that clears it entirely as far as nodes are concerned, which was  
my major issue here

-- Thierry