Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2021-04-28
11:01:13 gibi fyi we are expected failures and therefore work here ^^
11:01:13 gibi fyi we are expected failures and therefore work here ^^
11:01:21 gibi s/expected/expecting/
11:01:22 gibi s/expected/expecting/
11:11:22 openstackgerrit Brin Zhang proposed openstack/nova master: Replaces tenant_id with project_id from List/Show usage APIs https://review.opendev.org/c/openstack/nova/+/768509
11:11:33 gibi reported https://bugs.launchpad.net/nova/+bug/1926426 to track the work
11:11:35 openstack Launchpad bug 1926426 in OpenStack Compute (nova) "Nova is not compatible with sqlalchemy 1.4" [High,Triaged]
11:11:35 gibi reported https://bugs.launchpad.net/nova/+bug/1926426 to track the work
11:11:35 openstack Launchpad bug 1926426 in OpenStack Compute (nova) "Nova is not compatible with sqlalchemy 1.4" [High,Triaged]
11:13:17 sean-k-mooney is the issue the test or the production code?
11:13:17 sean-k-mooney is the issue the test or the production code?
11:13:56 sean-k-mooney looks like maybe both
11:13:56 sean-k-mooney looks like maybe both
11:14:02 sean-k-mooney have we capped it in UC
11:14:02 sean-k-mooney have we capped it in UC
11:15:24 sean-k-mooney ah we had to 1.3.23 previously
11:15:24 sean-k-mooney ah we had to 1.3.23 previously
11:15:50 gibi it is a test coming from a proposed UC bump
11:15:50 gibi it is a test coming from a proposed UC bump
11:16:07 gibi so far we see one oslo.db issue but we also seem to have some nova specific failure modes as well
11:16:07 gibi so far we see one oslo.db issue but we also seem to have some nova specific failure modes as well
11:56:26 gibi sean-k-mooney: Did I understood the comment from the PTG well about when we need direction aware and when we need direction less pps resource? https://review.opendev.org/c/openstack/neutron-specs/+/785236/3/specs/xena/qos-minimum-guaranteed-packet-rate.rst#266
11:56:26 gibi sean-k-mooney: Did I understood the comment from the PTG well about when we need direction aware and when we need direction less pps resource? https://review.opendev.org/c/openstack/neutron-specs/+/785236/3/specs/xena/qos-minimum-guaranteed-packet-rate.rst#266
11:58:37 sean-k-mooney sorry have nto got to your spec yet today but ill take a look at that comment now one sec
11:58:37 sean-k-mooney sorry have nto got to your spec yet today but ill take a look at that comment now one sec
11:59:41 sean-k-mooney gibi: so ya
11:59:41 sean-k-mooney gibi: so ya
12:00:01 sean-k-mooney we can either put the sum of the direct based values in the request
12:00:01 sean-k-mooney we can either put the sum of the direct based values in the request
12:00:22 sean-k-mooney or we can have 3 resources class and just make the user choose the correct policy
12:00:22 sean-k-mooney or we can have 3 resources class and just make the user choose the correct policy
12:01:46 sean-k-mooney i kind of like having 3 reosuce classes PPS_TOTAL, PPS_INGRESS and PPS_EGRESS
12:01:47 sean-k-mooney i kind of like having 3 reosuce classes PPS_TOTAL, PPS_INGRESS and PPS_EGRESS
12:01:52 gibi so then in case of hw offloaded OVS the packet processing bottleneck is not some kind of shared resource (like the CPU in case of normal OVS) but some direction specific hardware resoruce?
12:01:52 gibi so then in case of hw offloaded OVS the packet processing bottleneck is not some kind of shared resource (like the CPU in case of normal OVS) but some direction specific hardware resoruce?
12:02:37 gibi hence we need to track resource inventory per direction separately
12:02:37 gibi hence we need to track resource inventory per direction separately
12:02:41 sean-k-mooney well yes and know. the pcie bandwith is a shared resouce but hardware offloaded ovs use sriov VFs for the dataplane
12:02:41 sean-k-mooney well yes and know. the pcie bandwith is a shared resouce but hardware offloaded ovs use sriov VFs for the dataplane
12:03:03 sean-k-mooney so we can pretend it just sriov in this case
12:03:03 sean-k-mooney so we can pretend it just sriov in this case
12:03:22 sean-k-mooney vdpa is more or less the same in that the parent of the vdpa device atleat today is a VF
12:03:23 sean-k-mooney vdpa is more or less the same in that the parent of the vdpa device atleat today is a VF
12:03:54 gibi so in this case the logic that is implemented in OVS runs on the specific hardware?
12:03:55 gibi so in this case the logic that is implemented in OVS runs on the specific hardware?
12:04:59 gibi anyhow I think I get it, the HW offload OVS is like the SRIOV VF case
12:04:59 gibi anyhow I think I get it, the HW offload OVS is like the SRIOV VF case
12:05:00 sean-k-mooney how it works is the ovs vswitchd process caluates openflow rules as normal and then use the tcflower protocal to install those flows into the hardware dataplane if it supprot them
12:05:01 sean-k-mooney how it works is the ovs vswitchd process caluates openflow rules as normal and then use the tcflower protocal to install those flows into the hardware dataplane if it supprot them
12:05:24 gibi sean-k-mooney: thanks that helps visualizing the situation for me
12:05:24 gibi sean-k-mooney: thanks that helps visualizing the situation for me
12:06:19 sean-k-mooney basically you can thinkg of the programming model like a cache if the hardware can cache the rule in its hardware swtich it will process them in hardware. the ovs proces just caculates them as normal and they are then copied to the hardware
12:06:19 sean-k-mooney basically you can thinkg of the programming model like a cache if the hardware can cache the rule in its hardware swtich it will process them in hardware. the ovs proces just caculates them as normal and they are then copied to the hardware
12:06:50 sean-k-mooney what that means in partcie is the first packet(s) of each connection is processed in software and then hardware takes over
12:06:50 sean-k-mooney what that means in partcie is the first packet(s) of each connection is processed in software and then hardware takes over
12:08:07 gibi I see
12:09:18 gibi and in this hardwer the incoming and outgoing packet processing is implemented independently so we can run out of one direction while the other direction still has capacity
12:09:18 gibi and in this hardwer the incoming and outgoing packet processing is implemented independently so we can run out of one direction while the other direction still has capacity
12:09:40 gibi and therefore we need separated resource tracking
12:09:40 gibi and therefore we need separated resource tracking
12:10:11 sean-k-mooney i think that depens on the nic but in principal yes the internal capasity of the switch is higher then the phsyical connectors
12:10:11 sean-k-mooney i think that depens on the nic but in principal yes the internal capasity of the switch is higher then the phsyical connectors
12:10:45 sean-k-mooney so it can run out of capasity to transmit to the datacenter network before its ablity to swtich between vms on the same host
12:10:45 sean-k-mooney so it can run out of capasity to transmit to the datacenter network before its ablity to swtich between vms on the same host
12:11:17 sean-k-mooney although intel tend and mellonox until recently had a habbit of oversubscibing ther port bandwith
12:11:17 sean-k-mooney although intel tend and mellonox until recently had a habbit of oversubscibing ther port bandwith
12:11:59 sean-k-mooney i.e. intel 2*40G fortviles only had 56Gb/s of pcie bandwith
12:12:00 sean-k-mooney i.e. intel 2*40G fortviles only had 56Gb/s of pcie bandwith
12:12:53 sean-k-mooney which is fine for the 2*25G card but ya pcie bandwith and phycial port bandwith can often be less the the chip bandwith in the data center space
12:12:53 sean-k-mooney which is fine for the 2*25G card but ya pcie bandwith and phycial port bandwith can often be less the the chip bandwith in the data center space
12:13:40 sean-k-mooney the assumtion often is that you are using 2 port cards for HA not more bandwitr/pps
12:13:40 sean-k-mooney the assumtion often is that you are using 2 port cards for HA not more bandwitr/pps
12:14:16 gibi when you say "run out of capasity to transmit to the datacenter" do you mean it runs out of physical nework bandwidth?
12:14:16 gibi when you say "run out of capasity to transmit to the datacenter" do you mean it runs out of physical nework bandwidth?
12:14:29 sean-k-mooney yep
12:14:43 gibi so while it can still swtich packets, it cannot transmit them
12:14:43 gibi so while it can still swtich packets, it cannot transmit them
12:15:09 gibi but then this is showing just that the packet processing capacity is independent from the bandwidth
12:15:09 gibi but then this is showing just that the packet processing capacity is independent from the bandwidth
12:15:43 gibi it does not show that switching incoming packets are independend from switching outgoing packets
12:15:43 gibi it does not show that switching incoming packets are independend from switching outgoing packets
12:15:47 sean-k-mooney ya the 2*25G intel fortvales actully had 2 40G chips one per uplink port but had 25G phys
12:15:47 sean-k-mooney ya the 2*25G intel fortvales actully had 2 40G chips one per uplink port but had 25G phys
12:17:27 sean-k-mooney gibi: you are right it does not show they are independeint on the recive side if you have an intel cpu with vt-c
12:17:27 sean-k-mooney gibi: you are right it does not show they are independeint on the recive side if you have an intel cpu with vt-c
12:17:56 sean-k-mooney that allows supported nic to place recived packet directly into the cpus l3 cache
12:17:56 sean-k-mooney that allows supported nic to place recived packet directly into the cpus l3 cache
12:18:40 sean-k-mooney on transmit i know the cpu can do a dma transfre to the nic queue or indead to a zerocopy transmit in some cases
12:18:40 sean-k-mooney on transmit i know the cpu can do a dma transfre to the nic queue or indead to a zerocopy transmit in some cases
12:20:36 gibi OK, I rest my case. I will do allow the admin to configure either two direction aware pps resource inventory (in case of offloaded OVS) or a single direction less resource inventory (for the normal OVS)
12:20:36 gibi OK, I rest my case. I will do allow the admin to configure either two direction aware pps resource inventory (in case of offloaded OVS) or a single direction less resource inventory (for the normal OVS)
12:20:53 sean-k-mooney well no
12:20:53 sean-k-mooney well no
12:21:16 sean-k-mooney if you think that scoping this down to just normal ovs for now is the best way forward that would be ok
12:21:16 sean-k-mooney if you think that scoping this down to just normal ovs for now is the best way forward that would be ok
12:21:58 sean-k-mooney it would proably be nice to support both modesl as you said but we could start with the one you want
12:21:58 sean-k-mooney it would proably be nice to support both modesl as you said but we could start with the one you want
12:22:18 gibi I can do both, this is not the hard part (I guess :D)_
12:22:18 gibi I can do both, this is not the hard part (I guess :D)_
12:22:21 sean-k-mooney i just would not like to see logic to traslate driction based request into driectionless

Earlier   Later