-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/24598/
-----------------------------------------------------------
(Updated Aug. 12, 2014, 11:21 a.m.)
Review request for cloudstack, Alex Huang, anthony xu, daan Hoogland, edison
su, Kishan Kavala, Min Chen, Sanjay Tripathi, and Hugo Trippaers.
Bugs: CLOUDSTACK-7319
https://issues.apache.org/jira/browse/CLOUDSTACK-7319
Repository: cloudstack-git
Description (updated)
-------
We noticed that the dd process was way to agressive on Dom0 causing all kinds
of problems on a xenserver with medium workloads.
ACS uses the dd command to copy incremental snapshots to secondary storage.
This process is to heavy on Dom0 resources and even impacts DomU performance,
and can even lead to domain freezes (including Dom0) of more then a minute.
We've found that this is because the Dom0 kernel caches the read and write
operations of dd.
Some of the issues we have seen as a consequence of this are:
- DomU performance/freezes
- OVS freeze and not forwarding any traffic
- Including LACPDUs resulting in the bond going down
- keepalived heartbeat packets between RRVMs not being send/received resulting
in flapping RRVM master state
- Braking snapshot copy processes
- the xenserver heartbeat script reaching it's timeout and fencing the server
- poolmaster connection loss
- ACS marking the host as down and fencing the instances even though they are
still running on the origional host resulting in the same instance running on
to hosts in one cluster
- vhd corruption are a result of some of the issues mentioned above
We've developed a patch on the xenserver scripts
/etc/xapi.d/plugins/vmopsSnapshot that added the direct flag of both input and
output files (iflag=direct oflag=direct).
Our test have shown that Dom0 load during snapshot copy is way lower.
We believe Hot-fix 4 for XS62 sp1 contains a similar fix but for the sparse dd
process used for the first copy of a chain.
http://support.citrix.com/article/CTX140417
== begin quote ==
Copying a virtual disk between SRs uses the unbuffered I/O to avoid polluting
the pagecache in the Control Domain (dom0). This reduces the dom0 vCPU overhead
and allows the pagecache to work more effectively for other operations.
== end quote ==
Diffs
-----
scripts/vm/hypervisor/xenserver/vmopsSnapshot 5fd69a6
Diff: https://reviews.apache.org/r/24598/diff/
Testing
-------
We are running this fix in our beta and prod environment (both using ACS 4.3.0)
with great success.
Thanks,
Joris van Lieshout
