Port creation times out for some VMs in large group
fsbiz at yahoo.com
fsbiz at yahoo.com
Mon Oct 7 20:33:20 UTC 2019
Thanks. Yes, it helps breathe some CPU cycles.
This was traced to afixed bug.
which was applied to Queens in April 2019.
Unfortunately, the patch simply makes the code more elegant by removing the semaphores.But it does not really fix the real issue that is dhcp-client serializes all the port update messages and eachmessage is processed too slowly resulting in PXE boot timeouts.
The issue still remains open.
On Wednesday, October 2, 2019, 11:34:39 AM PDT, Chris Apsey <bitskrieg at bitskrieg.net> wrote:
Is that still spitting out a vif plug failure or are your instances spawning but not getting addresses? I've found that adding in the no-ping option to dnsmasq lowers load significantly, but can be dangerous if you've got potentially conflicting sources of address allocation. While it doesn't address the below bug report specifically, it may breathe some more CPU cycles into dnsmasq so it can handle other items better.
-------- Original Message --------
On Oct 2, 2019, 12:41, fsbiz at yahoo.com < fsbiz at yahoo.com> wrote:
Thanks. This definitely helps.
I am running a stable release of Queens.Even after this change I still see 10-15 failures when I create 100 VMs in our cluster.
I have tracked this down (to a reasonable degree of certainty) to the SIGHUPs caused by DNSMASQ reloadsevery time a new MAC entry is added, deleted or updated.
It seems to be related tohttps://bugs.launchpad.net/neutron/+bug/1598078
The fix for the above bug was abandoned. Gerrit Code Review
Gerrit Code Review
Any further fine tuning that can be done?
On Friday, September 27, 2019, 09:37:51 AM PDT, Chris Apsey <bitskrieg at bitskrieg.net> wrote:
Do this: https://cloudblog.switch.ch/2017/08/28/starting-1000-instances-on-switchengines/
The problem will go away. I'm of the opinion that daemon mode for rootwrap should be the default since the performance improvement is an order of magnitude, but privsep may obviate that concern once its fully implemented.
Either way, that should solve your problem.
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
On Friday, September 27, 2019 12:17 PM, Albert Braden <Albert.Braden at synopsys.com> wrote:
When I create 100 VMs in our prod cluster:
openstack server create --flavor s1.tiny --network it-network --image cirros-0.4.0-x86_64 --min 100 --max 100 alberttest
Most of them build successfully in about a minute. 5 or 10 will stay in BUILD status for 5 minutes and then fail with “ BuildAbortException: Build of instance <UUID> aborted: Failed to allocate the network(s), not rescheduling.”
If I build smaller numbers, I see less failures, and no failures if I build one at a time. This does not happen in dev or QA; it appears that we are exhausting a resource in prod. I tried reducing various config values in dev but am not able to duplicate the issue. The neutron servers don’t appear to be overloaded during the failure.
What config variables should I be looking at?
Here are the relevant log entries from the HV:
2019-09-26 10:10:43.001 57008 INFO os_vif [req-dea54d9a-3f3e-4d47-b901-a4f41b1947a8 d28c3871f61e4c8c8f8c7600417f7b14 e9621e3b105245ba8660f434ab21016c - default 4fb72165eee4468e8033cdc7d506ddf0] Successfully plugged vif VIFBridge(active=False,address=fa:16:3e:8b:45:07,bridge_name='brq49cbe55d-51',has_traffic_filtering=True,id=18f4e419-b19c-4b62-b6e4-152ec78e72bc,network=Network(49cbe55d-5188-4183-b5ad-e65f9b46f8f2),plugin='linux_bridge',port_profile=<?>,preserve_on_delete=False,vif_name='tap18f4e419-b1')
2019-09-26 10:15:44.029 57008 WARNING nova.virt.libvirt.driver [req-dea54d9a-3f3e-4d47-b901-a4f41b1947a8 d28c3871f61e4c8c8f8c7600417f7b14 e9621e3b105245ba8660f434ab21016c - default 4fb72165eee4468e8033cdc7d506ddf0] [instance: dc58f154-00f9-4c45-8986-94b10821cbc9] Timeout waiting for [('network-vif-plugged', u'18f4e419-b19c-4b62-b6e4-152ec78e72bc')] for instance with vm_state building and task_state spawning.: Timeout: 300 seconds
More logs and data:
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the openstack-discuss