pinggit
commented
4 years ago From: Laurent Antoine Durand ldurand@juniper.net
In my opinion, the first thing to achieve is to split the work between us in order to work in //.
How to work altogether ? I’m proposing that Ping keeps the lead of the whole document consistency. Then : First, under his control, lets first definitely split the document in chapters. Second, let’s work and clean respectively on our allocated chapter. Third, reassemble the whole (ping) = cut and paste the content of each producer in a single doc. Then we can iterate in order to reach our target (we could work in a cycle of one week)
About current TOC
In my opinion current TOC is perfectible. See ny remarks. Anyway, once a TOC has been agreed, let’s split the work. I propose we are assigning each chapter to a main contributor (chapter’s lead), who will be responsible to produce most of the content and in which others can contribute to help him to enrich or cover a part in which he needs help. If agreed, it will allow us to work separately on a subpart of doc, and to reassemble (ping ?) at each end of week.
At the end of this mail, some remarks on TOC changes proposals and also on who could take a first lead for the coming weeks on some chapters in order to get quickly a first result.
About chapter 2 (enclosed draft document)
For the chapter 2, I can take the lead, or Ping can keep it if he prefers. Here is my first (draft) contribution for this chapter. PS: Maybe this chapter 2 could be split in two parts (one links to DPDK and Linux – NIC drivers, PMD, huge pages / and one bound to Linux in General: cpu isolation, hardware discovery, …). My contribution for this chapter anyway will be more focused on DPDK specific, and not Linux Generic).
Here a proposal on how to split the task: . |[underline]#chapter 1: SDN overview# I propose that Ping take the lead on this chapter.
. |[underline]#chapter 2: SDN dataplane fundamentals# I can take the lead on this chapter. With some help on: . . . |smartnic . . . |vDPA . . . |eBPF
= . |[underline]#chapter 3: Contrail vrouter and DPDK# To be discussed. But in my opinion we have almost everything in perf guide. I would prefer following header: #chapter 3: Contrail DPDK vrouter architecture# After cleaning, would be fine to define a leader
. |[underline]#chapter 4: DPDK vrouter packet forwarding# To be discussed. But in my opinion we have around 50 to 60% in perf guide. I would prefer following header: #chapter 4: Contrail DPDK vrouter packet processing# After cleaning, would be fine to define a leader
. |[underline]#chapter 5: huge pages#
To be discussed. But in my opinion we have around of content 90% in perf guide. Seems short for a chapter.
. . I would prefer following header: #chapter 5: Contrail DPDK vrouter memory consideration#
|Huge pages configuration
. . |hugepages for vrouter: --socket-mem
. . |Kubernetes: vrouter and Single Hugepage size
. . |IRQ setup
. |[underline]#chapter 6: CPU pinning#
To be discussed. But in my opinion we have around of content 90% in perf guide. Seems short for a chapter.
. . I would prefer following header: #chapter 6: Contrail DPDK vrouter fine tuning#
(or chapter 6: Contrail DPDK vrouter CPU assignment), if we keel Chapter 5, 6, and 7 and we do not merge them.
To be discussed. But in my opinion this part should me move earlier (chapter 2) because this is generic Linux Concern. . . |Physical CPU cores allocation planning . . |Hyper Threading . . |CPU isolation mechanisms . . . |isolcpus mechanism configuration . . . |Tuned CPU partitioning configuration
To be discussed. But in my opinion we should merge content of chapter 5 in this chapter.
. . |CPU assignment for DPDK vRouter . . . |packet processing threads . . . |service and control threads <= 19.12 . . . |service and control threads pinning >= 20.03 . . |Provisioning concerns . . . |control and service threads >= 20.03 . . . |packet processing threads (forwarding cores) . . |CPU pinning and isolation . . . |tuned and isolcpus - RedHat . . . |isolcpus - Ubuntu . . . |Operating system scheduler CPU assignment (CPUAffinity) - RedHat . . . |Nova CPU assignment (vcpu_pin_set) . . . |vRouter CPU assignment (CPU_LIST, SERVICE_CORE_MASK and DPDK_CTRL_THREAD_MASK) . . . |Check CPU pinning on all processes
. |[underline]#chapter 7: DPDK vrouter Performance fine tuning# . . |Performance tuning essentials
To be discussed. But in my opinion this chapter should be restructured. We are mixing lots of different things …
. . |Enable CPU performance mode To be moved in previous chapter which is talking about CPUs.
But as there is a lots of different things, either we merge all chapter 5,6 and 7 together or we call it
chapter 7: Contrail DPDK vrouter underlay configuration
. . |vrouter physical bond setting . . |MPLSoUDP encapsulation Configuration . . . |Contrail Configuration . . . |SDN Gateway configuration: MPLSoUDP and Load Balancing
But as there is a lots of different things, either we merge all chapter 5,6 and 7 together or we call it
chapter 7: Contrail DPDK vrouter packets switching
. . |Manage a high flow environment . . . |Increase flow table size . . . |Flow setup rate optimisation . . . |Configuration of packet based processing
To be moved in chapter 5 which is talking about huge page.
. . |DPDK buffer size adjustment (mbuf) . . . |Intel recommendation . . . |Problem
To be moved in previous chapter (chapter 6) which is talking about CPUs. . . |vRouter CPU numbers and single queue VM
I propose that Premecz take the lead on this chapter. . |[underline]#chapter 8: Deployment considerations# . . |Define number of resources required for vRouter DPDK . . |Compute Node CPU capacity planning . . |vrouter forwarding cores capacity planning . . . |First step: expected network throughput . . . |Second step: VM DPDK support and multi queues capability . . . |Last step: sibling consideration . . |vrouter services cores capacity planning . . |typical CPU assignment . . |Performance tests figures (DPDK vrouter) . . |Jumbo frame support . . |Configure driver and network settings . . |Configure nova scheduler and flavor
I propose to take the lead on this chapter and to merge it with the next one.
. |[underline]#chapter 9: DPDK vrouter troubleshooting# . . |vrouter fine tuning parameters (kernel and DPDK mode) . . . |Generic vrouter dimensioning parameters . . . |Increase vrouter default value of dimensioning parameters (especially flow table) . . . |32 bits vrouter nexthop limit parameter (contrail release 19.11 and later) . . . |vrouter networking parameters (MTU) . . |vrouter DPDK fine tuning parameters . . . |DPDK vrouter parameters . . . |DPDK fine tuning rules (mainly for Intel NIC) . . . |DPDK vouter parameters provisioning . . . |DPDK vouter parameters configuration (5.0 and later versions) . . |Connectivity Troubleshooting . . . |How to check vrouter connectivity . . . |vRouter vif queues setup . . . |vRouter vif queues activation . . |Packet drop troubleshooting . . . |Interface traffic counters . . . |Packet flow from a Compute Physical NIC to a VM NIC - counters placement . . . |Packet flow from a VM NIC to compute node Physical NIC - counters placement . . . |Traffic Load balancing on forwarding cores . . . |vrouter is dropping packets . . . |virtual instance is dropping packets . . . |Packet Drop statistics . . . |vif command: get-drop-stats option . . . |Dropstats and packet drop log utility . . |Faulty vrouter deployment troubleshooting
I propose to take the lead on this chapter and to merge it with the previous one.
. |[underline]#chapter 10: DPDK vrouter log analysis# . |appendix . . |References
Subject: RE: vrouter/DPDK day one book: extending/polishing the perf guide
see discussion: https://github.com/pinggit/dpdk-contrail-book/blob/master/README.md#0529-meeting-minutes
My 2nd read is till ch5 and was trying to raise lots of questions down to the details. But we’ve decided to split jobs so each one can focus on smaller part to get work done better.
10 chapters in total.
TOCv0:
I’ll start to work on chapter 1/2 and try to finalize in 2 weeks. (Later if time permit I can also work on 3 and 4) I think chapter 5 is short/easy and chapter 10 is almost done (by Laurent) Chapter 3/4/6/7/8/9 needs some effort.
Please share which part you pick. We will help each other.
Please share concerns anytime.
NOTE: we need at least 2 weeks to review. If a draft is too raw it will be rejected, or bring big headache to editors (they are not technical on our content).