Earlier  
Later
New Search
Posted Nick Remark
#openstack-cyborg - 2019-10-24
03:36:25 shaohe_feng These components may run on the same host node with VM guest.
03:36:39 shaohe_feng Seems this is out of cyborg control. (should we add a reservation mechanism for them?)
03:36:44 Sundar Could you elaborate or provide links?
03:37:11 shaohe_feng I have attend a meeting about accelerators.
03:38:12 shaohe_feng Get the information, not only VM need accelerator, but maybe some other applications run on the host node also need accelerators
03:39:02 Sundar It is not common in an OpenStack cluster to have VMs and non-VM processes running on the same compute node.
03:39:51 shaohe_feng I meam the non-VM processes are openstack required components
03:40:03 shaohe_feng Not sure you know RDT.
03:40:24 Sundar Yes, I know RDT. Are you referring to RMD?
03:40:34 shaohe_feng No
03:40:37 shaohe_feng RDT
03:40:53 shaohe_feng the use case is:
03:41:04 shaohe_feng Divide cache into 2 parts.
03:41:29 Sundar Resource Director technology (RDT) is a set of Intel processor features for cache/memory management. They can be used in OpenStack in different ways.
03:41:42 shaohe_feng 1 parts are reserved for system IAAS sotrware run on the node
03:42:04 shaohe_feng another parts are reserved for VM.
03:42:49 Sundar That part is outside Cyborg though -- it relates to cache/memory and should be configured by the operator. How does it relate to accelerators?
03:43:17 shaohe_feng We defined, cyborg can only allocate accelerators to nova, right?
03:43:48 shaohe_feng I means like RDT.
03:43:58 shaohe_feng we have 5 accelerators.
03:44:11 shaohe_feng we divided into 2 parts.
03:45:20 shaohe_feng 1 parts have 1 acc, it is out of cyborg control. It is reserved for system IAAS sotrware. Admin allocate them
03:45:56 shaohe_feng another parts have the remain 4 acc, cyborg allocate them to nova.
03:46:09 shaohe_feng similar like RDT use case.
03:46:38 shaohe_feng on we can simple define cyborg.
03:47:20 shaohe_feng if cyborg run on a host node. we not allow to allocate accelerators to system IAAS sotrware any more.
03:47:38 shaohe_feng we can note this on cyborg docs.
03:47:44 Sundar If the admin wants to keep aside some accelerators, he/she can update Placement to mark that resource provider as in-use.
03:48:22 Sundar From Cyborg's POV it created the RP but Placement will never pick that RP because it sees it as in-use
03:49:26 xinranwang As we know, cinder has enable qat driver, I guess shaohe want to know if Cyborg can help cinder manage this kind of accelerator, is that what you want say, shaohe_feng?
03:49:44 shaohe_feng not only cinder
03:49:53 shaohe_feng also for some VPN on the host.
03:50:02 chenke what does POV means?
03:50:02 shaohe_feng and storage
03:50:08 xinranwang yes, cinder is just an example
03:50:12 Sundar chenke: point of view
03:50:17 shaohe_feng DPDK, SPDK also need .
03:50:25 shaohe_feng yes, it is an example
03:50:49 shaohe_feng so does that means, let cyborg run first before other system IAAS sotrware, it report resources to placement first
03:51:34 shaohe_feng and amdin reserve some accelerators, and start the remain system IAAS sotrwares
03:52:11 shaohe_feng and amdin reserve some accelerators to the remain system IAAS sotrwares
03:52:57 Sundar Not sure where this discussion is headed. If Cinder is using a device, operator should not configure Cyborg driver for that device. If the admin wants to dedicate some networking/storage devices for DPDK/SPDK and they were claimed by Cyborg, the admin could still mark them in placement, as I said.
03:53:01 shaohe_feng We can add this usage as a guideline in cyborg doc.
03:54:04 shaohe_feng Sundar: can we add this to the cyborg doc?
03:54:17 Sundar shaohe_feng: It may be better to wait for concrete usage by operators and get their feedback, instead of adding hypothetical use cases in the docs -- they could be confusing.
03:54:19 shaohe_feng maybe DPDK as example.
03:55:10 s_shogo I also interested in disabling to allocate some accelerators, as maintenance.
03:55:43 Sundar s_shogo: Only when no VM is using that device, right?
03:55:56 s_shogo Sundar: yes
03:56:00 shaohe_feng Maybe we had better not allow DPDK to use accelerators any more. This can make things more simple.
03:56:09 xinranwang Currently, operator plays this role.
03:57:18 shaohe_feng For example, had better to avoid DPDK to use some network accelerator.
03:57:50 shaohe_feng Then things is simple.
03:57:53 Sundar s_shogo: Would the admin-Placement method work for now? We had plans to add health reporting and enabling/disabling devices. Once we have the /v2/devices API, we could add this functionality. Seems ok?
03:58:46 Sundar shaohe_feng: Cyborg does not support network resources today.
03:59:02 shaohe_feng yes, just a example.
03:59:35 shaohe_feng QAT can be used for some network accelerations
03:59:47 shaohe_feng and it can be used by DPDK.
04:00:00 shaohe_feng not sure cyborg will support QAT
04:00:23 s_shogo Sundar: Yes, the approach is good. I want to discuss the methods continuously with example usecase for maintenance and that's requirements.
04:00:26 shaohe_feng Or cyborg just plan to manager only VM used accelerators.
04:01:37 Sundar shaohe_feng: Do you have specific knowledge that somebody wants to have several QATs in one host, some for VMs, and some for DPDK on the host?
04:02:45 Sundar s_shogo: Sure, let's make this a discussion topic at the PTG. Your I could create an etherpad with the problem statement and some ideas.
04:03:15 Sundar *You or I
04:04:25 shaohe_feng usually，they use kolla to deploy the system IAAS sotrwares include (placement), How do kolla know what's accelerators in placement？
04:04:30 s_shogo Sundar : Thanks, I'll update the etherpad (part of the topic is already written)
04:05:31 Sundar My suggestion: let's get Nova integ done and make something functional. Then we can start talking about more advanced use cases.
04:05:52 Sundar We are over the time. Is there anything else?
04:06:49 shaohe_feng Yes, the meeting tells use DPDK can use accelerators. But they does not say DPDK can not live with VM on same host.
04:07:10 shaohe_feng Ok.
04:07:36 shaohe_feng seems, we did not support much advanced use cases at present.
04:07:40 shaohe_feng step by step
04:08:10 Sundar shaohe_feng: Let's take this offline.
04:08:16 Sundar Good. Thanks a lot, everybody. Have a good day!
04:08:20 Sundar #endmeeting
04:08:22 openstack Meeting ended Thu Oct 24 04:08:20 2019 UTC. Information about MeetBot at http://wiki.debian.org/MeetBot . (v 0.1.4)
04:08:23 openstack Minutes: http://eavesdrop.openstack.org/meetings/openstack_cyborg/2019/openstack_cyborg.2019-10-24-03.02.html
04:08:24 openstack Minutes (text): http://eavesdrop.openstack.org/meetings/openstack_cyborg/2019/openstack_cyborg.2019-10-24-03.02.txt
04:08:25 openstack Log: http://eavesdrop.openstack.org/meetings/openstack_cyborg/2019/openstack_cyborg.2019-10-24-03.02.log.html
04:10:07 s_shogo If you have time, a specific topic related to intel PAC > mainly intel folksI'm trying to build devstack env with intel-PAC board, but the latest devstack config seems to be lack of main body of OPAE driver, is out of scope in devstack?
04:10:34 s_shogo Not in a hurry topic.
04:10:52 Sundar s_shogo: Yes, the OPAE kernel driver is not installed by devstack. Mainly because that requires the right kernel headers etc. and can go wrong in many ways
04:11:09 Sundar Do you have the right documentation to install OPAE drivers?
04:11:09 s_shogo Sundar: OK, thanks. I got it.
04:11:52 s_shogo Sundar: I always refer the official manual and works well.
04:12:12 Sundar Ok, great
04:14:35 s_shogo Thanks all, good evening and lunch, bye
04:14:45 chenke bye~
07:57:19 openstackgerrit YumengBao proposed openstack/cyborg master: Update api-ref: deprecate v1 add v2 device_profile section https://review.opendev.org/690539
12:40:34 openstackgerrit chenker proposed openstack/cyborg master: Enable openstack-cover-jobs in zuul check for cyborg https://review.opendev.org/690950
14:28:33 openstackgerrit chenker proposed openstack/cyborg master: Bugfix: _needs_programming lack one situation process https://review.opendev.org/690983
14:35:55 openstackgerrit chenker proposed openstack/cyborg master: Bugfix: _needs_programming lack one situation process https://review.opendev.org/690983
18:09:15 openstackgerrit Merged openstack/cyborg master: Update docstring for apply_patch https://review.opendev.org/690517
#openstack-cyborg - 2019-10-25
03:07:46 openstackgerrit chenker proposed openstack/cyborg master: Bugfix: modify the log location for successful async bind ARQs https://review.opendev.org/691170
13:11:40 openstackgerrit YumengBao proposed openstack/cyborg master: Update api-ref: Add accelerator requests(ARQs) section https://review.opendev.org/691306
13:13:55 openstackgerrit YumengBao proposed openstack/cyborg master: Update api-ref: Add accelerator requests(ARQs) section https://review.opendev.org/691306
#openstack-cyborg - 2019-10-26
08:09:32 openstackgerrit YumengBao proposed openstack/cyborg master: Update api-ref: Add accelerator requests(ARQs) section https://review.opendev.org/691306
#openstack-cyborg - 2019-10-28
02:09:54 openstackgerrit chenker proposed openstack/python-cyborgclient master: Stop testing python2.7 https://review.opendev.org/691549
06:40:58 openstackgerrit chenker proposed openstack/python-cyborgclient master: (WIP)Stop testing python2.7 https://review.opendev.org/691549
08:23:16 openstackgerrit YumengBao proposed openstack/cyborg master: Update api-ref: Add accelerator requests(ARQs) section https://review.opendev.org/691306

Earlier   Later