[one-dev] Live migration / recovery

Gareth Bult gareth at linux.co.uk
Mon Jan 13 08:28:56 PST 2014


Ok, sounds good .. could I also draw your attention to the "deploy" script, which is the other 
script I need to modify .. my version looks like this; 


-- 

mkdir -p `dirname $domain` 
cat > $domain 


logger "DEPLOY :: /var/lib/one/remotes/hooks/vdc/deploy.py $domain" 
/var/lib/one/remotes/hooks/vdc/deploy.py $domain 

data=`virsh --connect $LIBVIRT_URI create $domain` 

-- 

Deploy does three relatively critical things you may wish to consider; 

a. adds the "sharable" flag to the libvirt config - without this libvirt won't live migrate (on the latest versions of libvirt) 
b. adds "error_policy=stop" to ensure the VM pauses in the case of an IO error, which lets you fix the IO issue, then unpause it 
c. adds "discard=unmap", which you need in order for KVM to support TRIM requests properly 
(we use TRIM to allow a VM to pass references to unused space to the back-end storage so it can be released back to the OS) 

If it's of any use, you're more than welcome to use my deploy.py script .. for what it does it seems fairly concise .. :) 

#!/usr/bin/python 

import xml.etree.ElementTree as ET 
from sys import argv 

original = ET.parse(argv[1]) 
params = original.findall(".//disk") 
for p in params: 
if p.get('device') == 'disk': p.insert(0,ET.Element("shareable",{})) 

params = original.findall(".//driver") 
for p in params: 
if p.get('type') == 'raw': 
p.set("error_policy","stop") 
p.set("discard","unmap") 

original.write(argv[1]) 

Regards, 
Gareth. 


-- 
	
Gareth Bult 
“The odds of hitting your target go up dramatically when you aim at it.” 
See the status of my current project at http://vdc-store.com 


----- Original Message -----

From: "Ruben S. Montero" <rsmontero at opennebula.org> 
To: "Gareth Bult" <gareth at linux.co.uk> 
Cc: dev at lists.opennebula.org 
Sent: Monday, 13 January, 2014 3:41:50 PM 
Subject: Re: [one-dev] Live migration / recovery 

Totally agree, in fact we've seen this in the past. 

This is fairly easy to add, one_vmm_exec.rb includes a pseudo-dsl to specify the actions, each action includes a fail action. Looking at the code we have also to roll-back the networking configuration on the target host. 

Added a new issue for this. In the meantime we can just use the previous workaround. 

http://dev.opennebula.org/issues/2633 



On Mon, Jan 13, 2014 at 2:59 PM, Gareth Bult < gareth at linux.co.uk > wrote: 



Ok, this almost seems too easy ... :) 

What I was trying to avoid was more tweaking of ON files when installing VDC. 

The issue I see is that each time someone upgraded ON, they will potentially have a chain of 
small patches to apply ... if I supply a few small patches based on "if <vdc installed> - do something extra" 
would you include these in the stock scripts?? 

Gareth. 

-- 
	
Gareth Bult 
“The odds of hitting your target go up dramatically when you aim at it.” 
See the status of my current project at http://vdc-store.com 



From: "Ruben S. Montero" < rsmontero at opennebula.org > 
To: "Gareth Bult" < gareth at linux.co.uk > 
Cc: dev at lists.opennebula.org 
Sent: Sunday, 12 January, 2014 10:36:30 PM 
Subject: Re: [one-dev] Live migration / recovery 

Hi Gareth 

As the migrate script can be easily updated we do not provide any hook for that. I'd go to kvm/migrate, and do a simple if [ $? ... after the virsh command to kill the cache on the target host. 

Cheers 

Ruben 


On Thu, Dec 19, 2013 at 12:54 PM, Gareth Bult < gareth at linux.co.uk > wrote: 

<blockquote>

Hi, 

I implemented live migration for the VDC driver a few weeks back and on the whole it seems to work 
quite well. The "premigrate" script creates a cache instance on the target server and puts the source 
cache into proxy mode, then the "postmigrate" script kills the original cache instance. 

Problem :: if the migration fails, I'm left with a running cache on both the source and target servers, with 
the source cache in proxy mode. I have a 1-line CLI command to revert the issue, but I need to hook into 
the system in order to call it. 

How to do I do this, I guess effectively I need something like "postmigrate_fail" .. ??? 

tia 
Gareth. 

-- 
	Gareth Bult 
“The odds of hitting your target go up dramatically when you aim at it.” 




_______________________________________________ 
Dev mailing list 
Dev at lists.opennebula.org 
http://lists.opennebula.org/listinfo.cgi/dev-opennebula.org 

-- 
-- 
Ruben S. Montero, PhD 
Project co-Lead and Chief Architect 
OpenNebula - Flexible Enterprise Cloud Made Simple 
www.OpenNebula.org | rsmontero at opennebula.org | @OpenNebula 





</blockquote>




-- 
-- 
Ruben S. Montero, PhD 
Project co-Lead and Chief Architect 
OpenNebula - Flexible Enterprise Cloud Made Simple 
www.OpenNebula.org | rsmontero at opennebula.org | @OpenNebula 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.opennebula.org/pipermail/dev-opennebula.org/attachments/20140113/68154d58/attachment-0002.htm>


More information about the Dev mailing list