Earlier  
Later
New Search
Posted Nick Remark
#openstack-cyborg - 2017-11-01
15:20:18 zhipeng #link https://github.com/jaypipes/articles/blob/master/openstack/placement-queens-update.md
15:20:46 zhipeng #info Jay Pipe's writeup on Placement
15:20:52 zhipeng this is as the best as you could get
15:21:12 zhipeng and there are two specs that will be very useful to us, from Ironic
15:21:23 zhipeng #link http://specs.openstack.org/openstack/nova-specs/specs/queens/approved/ironic-driver-traits.html
15:21:40 zhipeng #link https://review.openstack.org/#/c/504531/
15:22:17 zhipeng #info first one is about Ironic report/accept trait with Nova, second one is about trait config in Ironic
15:22:45 zhipeng ok there is another spec that will help our modeling, is the vGPU spec from jianghua
15:23:00 zhipeng #link http://specs.openstack.org/openstack/nova-specs/specs/queens/approved/virt-add-support-for-vgpu.html
15:23:21 zhipeng and even more interestingly
15:23:32 zhipeng we could also learn something from the kubernetes discussion
15:23:35 zhipeng #link https://github.com/kubernetes/community/pull/782
15:23:48 zhipeng #info k8s discussion on resource class for accelerators
15:24:36 zhipeng well based upon all the readings i did via the above mentioned links
15:25:06 zhipeng I think there several points we need to consider when we modeling cyborg maneged resources
15:25:43 zhipeng #info 1. We need to define CUSTOM_ traits for sure, but also some additions to os-trait lib on FPGA if there are any
15:26:41 zhipeng #info 2. other than the reporting from cyborg-agent to the Placement API, we need (a) the driver also support trait api and (b) cyborg-api support trait
15:27:09 crushil zhipeng, Ya, I have been following the traits discussion on the Ironic end too. So, I might have some inputs once there is a spec up on our end
15:27:44 zhipeng #info 3. There is a caveat for Cyborg on resource provider modeling for accelerators which is we need to find a way to model different types of accelerators in a unified way
15:29:17 zhipeng #info 3a. Traits and resource class are used to define qualitive and quantative side of resource, however we might need a third kind to descirbe accelerators, because only using traits and ResourceClass means we will have a lot them for different accelerators
15:30:35 zhipeng #info 3b. We might need to consider using something like capability to describe the accelerator resource. For example FPGA and GPU both could be used for ML, a user might just want a VM to run a ML workload with the help of an attached accelerator
15:31:29 zhipeng #info 3c. This means we need to enable user to be able to select one capable resource instead of identifing very specifically which resource they want
15:31:43 zhipeng ok that is the end of my monologue lol
15:31:49 zhipeng any thoughts ?
15:37:32 shaohe_feng_ need time to reading the link, too many info.
15:38:11 zhipeng ya , any thoughts on my comments above ? besides the links :P
15:43:05 crushil I will chime in once we have specs in our backlog
15:43:36 crushil I have some thoughts about the way we should be implementing traits on our end and handing them over to Nova
15:44:22 zhipeng crushil that brings mine proposals actually
15:44:23 crushil And that will alter the implementation of the generic driver as well which I am working on currently
15:44:36 zhipeng i think besides justin's general spec
15:44:46 zhipeng we will need one on the driver side and one on the api side
15:45:06 zhipeng would you like to lead on the driver trait spec ?
15:45:11 crushil Sure
15:45:37 crushil When is the stable Queens release date?
15:45:40 zhipeng zhuli and I will work on the api side
15:45:54 zhipeng crushil I need to double check on that
15:46:21 crushil I think we need to be reasonable with the amount of work we promise
15:46:30 crushil I would hate to overpromise and underdeliver
15:46:37 zhipeng #action crushil to work on a new spec on cyborg driver trait support, and zhipeng will work on a new spec on api trait extention
15:46:43 zhipeng crushil that is for sure
15:48:15 zhipeng and on that note, let's also start simple
15:48:32 zhipeng we could cover very basic trait descriptions for Queens
15:48:57 zhipeng meaning for example not necessarily model FPGA resources in a Tree structure in full
15:49:11 zhipeng let's deliver something basic in Queens
15:49:42 crushil zhipeng, Exactly with at least a shim that vendors can build upon
15:49:56 zhipeng crushil +1
15:50:13 zhipeng okey any more comments on this topic ?
15:50:38 zhipeng folks you are more than welcomed to provide further feedback on mailing list later on
15:50:53 crushil zhipeng, We need to have something like a midcycle to go over things that need work on
15:51:10 zhipeng that is what I think
15:51:20 zhipeng maybe a video conf meeting for like 2-3 hours ?
15:51:24 crushil Ya
15:51:27 zhipeng maybe the week of Sydney Summit ?
15:51:34 zhipeng since all of us will not be there lol
15:51:38 crushil lol
15:51:47 zhipeng let's do something useful instead
15:51:48 crushil Or we can do that monthly
15:51:58 zhipeng of cource that'd be great
15:52:10 crushil Justin and I can be in the same room
15:52:59 zhipeng lol don't eat ice cream in front of me
15:53:41 crushil We'll join from Coldstone
15:53:50 zhipeng lol
15:54:00 zhipeng okey moving onto next topic
15:54:14 zhipeng #topic Generic Driver progress update
15:54:32 zhipeng crushil any thoughts on what or how much work needed there ?
15:54:45 zhipeng i think shaohe could also lend a hand here
15:55:29 crushil So, I am working on creating interfaces between Cyborg APIs and generic driver. After that is done, I will work on creating a shim for the vendors
15:55:44 crushil That can be helpful if shaohe can help
15:55:54 crushil Does shaohe work with you?
15:56:42 zhipeng no but we are in the same country, so it'll be easy for us to communicate
15:56:53 zhipeng shaohe_feng_ interested ?
15:56:53 shaohe_feng_ I'm planning to support inte
15:57:01 shaohe_feng_ yes.
15:57:07 zhipeng gr8t !
15:58:02 shaohe_feng_ we are planning to support intel's accelerator in cyborg
15:58:09 zhipeng #action shaohe_feng_ and crushil collaborate on the driver development for backend support
15:58:23 shaohe_feng_ such as intel's GPU and FPGA
15:58:32 crushil shaohe_feng_, That's awesome
15:59:04 crushil I have reached out to other vendors. Intel was one of the vendors that I hadn't talked ot
15:59:05 crushil *to
15:59:15 zhipeng nice :)
15:59:40 zhipeng okey moving on to next topic
15:59:43 crushil shaohe_feng_, Are you specifically supporting Intel products or Altera products?
16:00:42 crushil sorry zhipeng We can talk about this offline. I don't mean to start another tangent
16:00:56 shaohe_feng_ crushil: I can help on some general work
16:01:01 zhipeng no worries
16:01:13 zhipeng no important items left :P
16:02:45 shaohe_feng_ we know better Intel's and Altera specifically products
16:02:58 shaohe_feng_ ^ crushil
16:03:01 crushil zhipeng, Please send a recurring calendar invite for every month where we analyze where we are
16:03:15 crushil shaohe_feng_, Cool
16:03:39 zhipeng #action zhipeng to send meeting invites for monthly deep dive
16:03:46 zhipeng #topic patch discussion
16:04:04 zhipeng #link https://review.openstack.org/#/q/project:openstack/cyborg+status:open
16:04:12 zhipeng any outstanding issue there ?
16:04:48 crushil We need more patches
16:04:51 crushil :)
16:05:08 zhipeng haha yep
16:05:34 zhipeng okey then

Earlier   Later