[ptls][URGENT] Bobcat Release: DPUs, GPUs, and FGPAs, oh my!
Hello folks!
So, as the OpenInfra Foundation has begun putting together the release marking materials for the Bobcat release, we noticed a trend focused on enabling hardware and increasing utilization. Ironic in particular had an excellent cycle highlight about DPUs[1]. We wanted to quickly reach out and see if there were any cycle highlights that projects maybe haven't published in a review, but are related to this topic. Please respond ASAP as we are moving fast on finalizing the press release.
Also, if there is anything related coming in Caracal, we might be able to call that out as well, though I know it's quite early to know what might be landing in 2024.1.
If there are other cycle highlights that should get called out as well, definitely push the patches and add me as a reviewer!
-Kendall Nelson
[1] https://releases.openstack.org/bobcat/highlights.html#ironic
Hello Kendall:
I can confirm that hardware offload (datapath offload in ML2/OVS and ML2/OVN) is currently used in several deployments and it's being tested internally in my company. However there are no upstream specific highlights about this topic right now. In the next cycle (or cycles) our plan is to create an external CI testing both mechanism drivers with HW NICs.
Regards.
On Tue, Sep 19, 2023 at 11:46 PM Kendall Nelson kennelson11@gmail.com wrote:
Hello folks!
So, as the OpenInfra Foundation has begun putting together the release marking materials for the Bobcat release, we noticed a trend focused on enabling hardware and increasing utilization. Ironic in particular had an excellent cycle highlight about DPUs[1]. We wanted to quickly reach out and see if there were any cycle highlights that projects maybe haven't published in a review, but are related to this topic. Please respond ASAP as we are moving fast on finalizing the press release.
Also, if there is anything related coming in Caracal, we might be able to call that out as well, though I know it's quite early to know what might be landing in 2024.1.
If there are other cycle highlights that should get called out as well, definitely push the patches and add me as a reviewer!
-Kendall Nelson
[1] https://releases.openstack.org/bobcat/highlights.html#ironic
On Wed, 2023-09-20 at 11:29 +0200, Rodolfo Alonso Hernandez wrote:
Hello Kendall:
I can confirm that hardware offload (datapath offload in ML2/OVS and ML2/OVN) is currently used in several deployments and it's being tested internally in my company. However there are no upstream specific highlights about this topic right now. In the next cycle (or cycles) our plan is to create an external CI testing both mechanism drivers with HW NICs.
nova also added suport for DPU in yoga https://specs.openstack.org/openstack/nova-specs/specs/yoga/implemented/inte... but there has been significant changes in this area recently.
there may be some new enabling of intel/napatech dpus/smartnice for dpdk based hardware offloads in both neutorn/nova next cycle. In bobcat https://specs.openstack.org/openstack/nova-specs/specs/2023.2/approved/suppo... was prospoed but not completed due to technial reasons. it made some progress but realistically there is more work in ovs to be done before that is sutable to merge upstream. as it stands the current approach is rather vendor specific but they are working to make a more vendor neutral approach work so i expect that to continue in caracal or D while this is relevent to this mail thread i dont think it should be included in any marketing matiral.
you could note that nova/neutron booth already support off-path dpus but i woudl likely omit the napatech spec since that is still in flux. "off-path" means that the contol plane of the dpu is runing on the dpu not the linux host wehere nova/neutron runs, in the case of the previous intergration ovn is the contolplane and it runs on the dpu and is integrated with ml2/ovn in neutron and nova's pci tracker. in the context of the ironic enhancement ironic woudl be used to provision both the host server OS and the linux/frimware image (with ovn) onto the dpu. that would then be configured by external tooling (i.e. ansible) to be integrated with neutron ml2/ovn backend and nova.
one low light is a lot of the work around nova <=> cycborg interaction has more or less stopped since the pandemic and i don't really see that changing anytime soon. The ironic enhancement kind of puts ironic and cyborg in competition for the management of the dpu lifecycle since that was inteneded to be managed via cycborg not ironic. It is the reason that cyborg has a rest api to manage the programing of a dpu os image and firmware. form a nova/neutron point of veiw it does not impact us as we just assume the lifecycle of the dpu is externally managed and wethere that si ironic or cyborg does not directly impact our usage of the dpu.
Regards.
On Tue, Sep 19, 2023 at 11:46 PM Kendall Nelson kennelson11@gmail.com wrote:
Hello folks!
So, as the OpenInfra Foundation has begun putting together the release marking materials for the Bobcat release, we noticed a trend focused on enabling hardware and increasing utilization. Ironic in particular had an excellent cycle highlight about DPUs[1]. We wanted to quickly reach out and see if there were any cycle highlights that projects maybe haven't published in a review, but are related to this topic. Please respond ASAP as we are moving fast on finalizing the press release.
Also, if there is anything related coming in Caracal, we might be able to call that out as well, though I know it's quite early to know what might be landing in 2024.1.
If there are other cycle highlights that should get called out as well, definitely push the patches and add me as a reviewer!
-Kendall Nelson
[1] https://releases.openstack.org/bobcat/highlights.html#ironic
Realistically the work Ironic has completed this cycle is to enable an operator to execute steps against a node and a DPU as part of the same deployment operation. This work is unrelated to past work in ironic to support "smart nics" (the original marketing name used for DPUs), which was more about the deploy time networking interactions. Unfortunately, we're not entirely there as most of this was the underlying substrate and workflow capabilities. We expect this to begin to be extended to aspects such as firmware management of the DPU, and we already have reports of operators using ironic separately to deploy the OS on some DPU models. The issue is before now, they were entirely disjointed. We expect this to improve over the next few years, but is also going to take standardization across the DPU vendor market to really enable. The OPI[0] community is working on trying to drive standardization, but many of the features and requirements being discussed will likely force revision in the silicon, in other words it might be a few years before we see the standardization we're trying to drive there reach a usable state.
That being said, we anticipate continued work in Caracal.
On Tue, Sep 19, 2023 at 2:49 PM Kendall Nelson kennelson11@gmail.com wrote:
Hello folks!
So, as the OpenInfra Foundation has begun putting together the release marking materials for the Bobcat release, we noticed a trend focused on enabling hardware and increasing utilization. Ironic in particular had an excellent cycle highlight about DPUs[1]. We wanted to quickly reach out and see if there were any cycle highlights that projects maybe haven't published in a review, but are related to this topic. Please respond ASAP as we are moving fast on finalizing the press release.
Also, if there is anything related coming in Caracal, we might be able to call that out as well, though I know it's quite early to know what might be landing in 2024.1.
If there are other cycle highlights that should get called out as well, definitely push the patches and add me as a reviewer!
-Kendall Nelson
[1] https://releases.openstack.org/bobcat/highlights.html#ironic
participants (4)
-
Julia Kreger
-
Kendall Nelson
-
Rodolfo Alonso Hernandez
-
smooney＠redhat.com