Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2018-05-17
12:33:03 mriedem belmoreira: yeah https://docs.openstack.org/osc-placement/latest/index.html
12:33:11 mriedem openstack resource class create - to create your custom resource class
12:33:17 gibi mriedem: and it seems that the fix https://review.openstack.org/#/c/569048 shows a race condition as well
12:33:22 mriedem openstack resource provider list - to find your compute node resource provider
12:33:58 mriedem belmoreira: and then you want to set that custom resource class inventory on the compute node resource provider? use https://docs.openstack.org/osc-placement/latest/cli/index.html#resource-provider-inventory-set
12:34:11 mriedem or actually https://docs.openstack.org/osc-placement/latest/cli/index.html#resource-provider-inventory-class-set is probably easier
12:34:27 mriedem openstack resource provider inventory set - makes you specify the full inventory with the changes, since it's a full overwrite
12:34:40 mriedem openstack resource provider class set - just modifies the inventory for the specified class
12:34:58 mriedem gibi: yeah i was wondering why the osc-placement functional test was still failng
12:36:27 mriedem gibi: are you sure that https://review.openstack.org/#/c/569048/ ran against the merged nova change?
12:36:55 belmoreira mriedem: thanks. This will be always defined with placement API. There isn't the option for the compute_node advertise the custom resource classes that it supports?
12:37:42 gibi mriedem: the nova patch merged yesterday, I pushed the fix for osc_placement today
12:37:46 gibi mriedem: so I assume
12:38:29 mriedem belmoreira: the compute node doesn't make up custom resource classes on the fly randomly :)
12:38:39 mriedem the compute node will report standard resource classes
12:38:39 gibi mriedem: and the osc_placement test fails differently with the fix than before
12:38:58 mriedem gibi: it looks like assertCommandFailed might just be doing a prefix vs assertIn check?
12:39:01 mriedem because you got HTTP 404 back
12:39:06 mriedem http://logs.openstack.org/48/569048/1/check/osc-placement-dsvm-functional/2800b3e/job-output.txt.gz#_2018-05-17_08_20_04_900926
12:39:33 mriedem belmoreira: we plan on reporting compute driver capabilities as standard traits on the compute node provider, see https://review.openstack.org/#/c/538498/
12:39:57 mriedem belmoreira: so we can report things like, "does this compute node support device tags, does this compute note support multiattach volumes, etc"
12:40:29 mriedem those are being defined here https://github.com/openstack/os-traits/tree/master/os_traits/compute
12:40:35 gibi mriedem: ahh. I've just realized that it fails in cleanup not in the test itself http://logs.openstack.org/48/569048/1/check/osc-placement-dsvm-functional/2800b3e/job-output.txt.gz#_2018-05-17_08_20_04_880092
12:40:44 mriedem i just need to cleanup that patch and get on the update_provider_tree() train
12:40:58 mriedem gibi: aha
12:41:17 gibi mriedem: so the test deletes the trait, then fails to clean up as that want to delete the trait again
12:41:32 mriedem gibi: ah yes https://review.openstack.org/#/c/569048/1/osc_placement/tests/functional/base.py@212
12:41:43 mriedem if 'not found' not in err_message:
12:41:46 mriedem just need to fix that
12:41:52 gibi mriedem: good point
12:41:59 mriedem if if 'not found' not in err_message or 'HTTP 404' in err_message:
12:42:00 gibi mriedem: I missed that
12:42:02 mriedem to support old and new placement
12:42:18 mriedem er or 'HTTP 404' not in err_message:
12:42:20 gibi mriedem: even better point to support old and new placement
12:42:31 mriedem gibi: want to fix that quick and then i'll rebase avolkov's series
12:42:39 mriedem belmoreira: does that help you out?
12:42:50 gibi mriedem: I'm doing it as we speak
12:44:16 belmoreira mriedem traits would do the trick as well. What I was thinking is an option to specify "custom" traits/resource classes per host.
12:45:02 belmoreira report compute_driver capabilities is good but we may want to use different properties as well
12:48:44 gibi mriedem: both the old and the new error message contains HTTP 404, so we dont need the ORed conditional
12:48:52 gibi stack@ubuntu:~/osc-placement$ openstack trait delete ALMA --os-placement-api-version 1.6
12:48:55 gibi No such trait(s): ALMA. (HTTP 404)
12:49:06 mriedem belmoreira: you can do that yourself
12:49:22 mriedem belmoreira: you mean you want to set a config option for nova to proxy all of that information for you?
12:50:04 gibi stack@ubuntu:~/nova$ openstack trait delete ALMA --os-placement-api-version 1.6
12:50:07 gibi Not found (HTTP 404)
12:50:14 mriedem belmoreira: the beauty of custom resource classes and traits and placement is that it's external to nova, so you don't need to proxy stuff through nova and add the complexity / bottleneck there
12:50:27 mriedem gibi: cool
12:51:00 openstackgerrit Balazs Gibizer proposed openstack/osc-placement master: Fix error message asserts in functional test https://review.openstack.org/569048
12:51:04 gibi mriedem: ^^ here it is
12:51:20 mriedem yup, watching zuul
12:54:42 mriedem tssurya: i guess belmiro didn't like that answer?
12:55:02 tssurya mriedem: I am sure that's not the case :D
12:55:09 mriedem he hates me now doesn't he
12:55:14 tssurya haha
12:55:19 mriedem I BLEW IT
12:55:48 mriedem tssurya: btw, got any cells v2 upgrade news?
12:56:32 jroll alex_xu: thanks
12:57:03 tssurya mriedem: its actually okay..
12:57:24 tssurya we are just trying to understand what is taking time in the scheduler
12:57:39 tssurya I mean what takes most of the time in the path of the scheduling
12:58:13 mriedem my guess would be that it's pulling more compute node records from the db than it used to do with the per-cell scheduler
12:58:48 tssurya and what's more weird is, how do I test exactly this scatter-gather versus normal sequential querying ?
12:58:52 gibi mriedem: the lates ps of the osc_placement fix still failing the functional test locally :/
12:58:56 tssurya like the time taken
12:59:02 mriedem so if you have 50 cells with 200 computes, before the top level scheduler would pick a cell and that would pull at most 200 compute nodes, right? but now it could be pulling at most 10K compute nodes
12:59:44 tssurya mriedem: we use dan's pre-placement filter to ensure we don't pull that much
12:59:54 mriedem tssurya: hmmm, we could add some debug logging to the scatter gather routine to log how long an operation took per cell...
13:00:26 tssurya mriedem: that is what I did, but it gives weird times,
13:00:32 tssurya not convincing ones
13:01:20 mriedem is the time being spent in pulling the compute nodes from the cells or from the filters?
13:02:21 tssurya mriedem: filters don't seem to take much time, pulling computes (the query directly done on the DB is very fast, however logging the scatter-gather routine gives weird longer times)
13:02:34 tssurya what is taking time is in the select destinations
13:02:44 tssurya I am working on it now to narrow it down further
13:02:47 mriedem also, do you know if you're hitting the same compute nodes for concurrent requests (packing)? just wondering if you're getting 409s from placement when trying to allocate resources against a target node
13:02:49 gibi mriedem: ohh, there is an upper() call on the message somewhere as {{{err_message:no such trait(s): custom_fake_hw_gpu_class_264c6d7df889415ea5d3b428d7a4a2fc. (http 404)}}}
13:03:32 mriedem tssurya: hmm, might be the alternate selection code?
13:03:51 tssurya mriedem: yea I am right at that point now
13:04:06 tssurya will see in prod if that's the one
13:04:52 openstackgerrit Balazs Gibizer proposed openstack/osc-placement master: Fix error message asserts in functional test https://review.openstack.org/569048
13:05:11 mriedem that will add more time compared to before queens because we are trying to now find up to max_attempts alternate hosts in the same cell, and that means running those alternates through the filters again
13:05:38 tssurya mriedem: oh right!
13:05:42 mriedem and that's per instance
13:06:08 mriedem so if you do server multicreate in a single request with like 10 instances and max_attempts is 3, we are trying to find i think 30 hosts
13:07:37 tssurya mriedem: hmm yea, will know if this is the one taking time, will check the logs
13:07:45 tssurya meanwhile mriedem I have another issue : http://logs.openstack.org/71/568271/2/check/tempest-full-py3/7142c54/testr_results.html.gz , not sure on how to fix this
13:08:18 gibi mriedem: ps 3 passes the functional locally
13:08:19 tssurya it totally depends on if the host runs both cell and super conductors or runs one of them
13:09:40 openstackgerrit Nguyen Hai proposed openstack/nova-specs master: Follow the new PTI for document build https://review.openstack.org/551802
13:11:53 mriedem tssurya: hmm, the cell conductor should be reporting into the cell1 database
13:12:22 mriedem this is n-super-cond: http://logs.openstack.org/71/568271/2/check/tempest-full-py3/7142c54/controller/logs/screen-n-super-cond.txt.gz#_May_17_08_20_09_133873
13:12:35 mriedem http://logs.openstack.org/71/568271/2/check/tempest-full-py3/7142c54/controller/logs/etc/nova/nova_conf.txt.gz
13:12:43 mriedem which is in cell0
13:12:44 mriedem [database] connection = mysql+pymysql://root:secretdatabase@127.0.0.1/nova_cell0?charset=utf8
13:13:03 mriedem the cell conductor is using the nova cell1 config
13:13:03 mriedem http://logs.openstack.org/71/568271/2/check/tempest-full-py3/7142c54/controller/logs/screen-n-cond-cell1.txt.gz#_May_17_08_20_10_145541
13:13:19 mriedem http://logs.openstack.org/71/568271/2/check/tempest-full-py3/7142c54/controller/logs/etc/nova/nova_cell1_conf.txt.gz
13:13:23 mriedem [database] connection = mysql+pymysql://root:secretdatabase@127.0.0.1/nova_cell1?charset=utf8
13:14:43 mriedem tssurya: it's a buggy test http://git.openstack.org/cgit/openstack/tempest/tree/tempest/api/compute/admin/test_services.py#n61

Earlier   Later