diff options
author | Deepak S <deepak.s@linux.intel.com> | 2017-09-04 02:52:15 -0700 |
---|---|---|
committer | Deepak S <deepak.s@linux.intel.com> | 2017-09-05 20:53:02 -0700 |
commit | dc8e018823c18c7ef0afdd84336064c417b05287 (patch) | |
tree | 4f630d103dcfac099ab78e1bdfa0a371ca550f1b /docs | |
parent | 4a31f76b5f3e1caad12b7cc3bfe97ee1f0dacfc7 (diff) |
Updating SampleVNF Highlevel design
Change-Id: I16edf82eec02c671c1c72d07348d1badf58fc42d
Signed-off-by: Deepak S <deepak.s@linux.intel.com>
Diffstat (limited to 'docs')
-rw-r--r-- | docs/testing/developer/design/04-SampleVNF_Desgin.rest | 123 | ||||
-rw-r--r-- | docs/testing/developer/design/04-SampleVNF_Design.rst | 630 | ||||
-rw-r--r-- | docs/testing/developer/design/images/prox-qo-img01.png | bin | 0 -> 20326 bytes | |||
-rw-r--r-- | docs/testing/developer/design/images/prox-qo-img02.png | bin | 0 -> 53045 bytes | |||
-rw-r--r-- | docs/testing/developer/design/images/prox-screen-01.png | bin | 0 -> 45792 bytes |
5 files changed, 630 insertions, 123 deletions
diff --git a/docs/testing/developer/design/04-SampleVNF_Desgin.rest b/docs/testing/developer/design/04-SampleVNF_Desgin.rest deleted file mode 100644 index 6c39da73..00000000 --- a/docs/testing/developer/design/04-SampleVNF_Desgin.rest +++ /dev/null @@ -1,123 +0,0 @@ -.. This work is licensed under a Creative Commons Attribution 4.0 International License. -.. http://creativecommons.org/licenses/by/4.0 -.. (c) OPNFV, Intel Corporation and others. - -.. OPNFV SAMPLEVNF Documentation design file. - -=================================== -SampleVNF Highlevel Desing -=================================== - -vFW - Design -============= - -Requirements ------------------ -Following are the design requierments of the vFW. - -- The firewall will examine packets and verify that they are appropriate for the - current state of the connection. Inappropriate packets will be discarded, and - counter(s) incremented. -- Support both IPv4 and IPv6 traffic type for TCP/UDP/ICMP. -- All packet inspection features like firewall, synproxy, connection tracker - in this component may be turned off or on through CLI commands -- The Static filtering is done thorugh ACL using DPDK libraries. The rules - can be added/modified through CLI commands. -- Multiple instance of the vFW Pipeline running on multipe cores should be - supported for scaling the performance scaling. -- Should follow the DPDK IP pipeline framework -- Sould use the DPDK libraries and functionalities for better performance -- The memory should be allocated in Hugepages using DPDK RTE calls for better - performance. - - -High Level Design -================= - -The Firewall performs basic filtering for malformed packets and dynamic packet -filtering incoming packets using the connection tracker library. -The connection data will be stored using a DPDK hash table. There will be one -entry in the hash table for each connection. The hash key will be based on -source address/port,destination address/port, and protocol of a packet. The -hash key will be processed to allow a single entry to be used, regardless of -which direction the packet is flowing (thus changing source and destination). -The ACL is implemented as libray stattically linked to vFW, which is used for -used for rule based packet filtering. - -TCP connections and UDP pseudo connections will be tracked separately even if -theaddresses and ports are identical. Including the protocol in the hash key -will ensure this. - -The Input FIFO contains all the incoming packets for vFW filtering. The vFW -Filter has no dependency on which component has written to the Input FIFO. -Packets will be dequeued from the FIFO in bulk for processing by the vFW. -Packets will be enqueued to the output FIFO. - -The software or hardware loadbalancing can be used for traffic distribution -across multiple worker threads. The hardware loadbalancing require ethernet -flow director support from hardware (eg. Fortville x710 NIC card). -The Input and Output FIFOs will be implemented using DPDK Ring Buffers. - -Components of vFW -================= - -In vFW, each component is constructed using packet framework pipelines. -It includes Rx and Tx Driver, Master pipeline, load balancer pipeline and -vfw worker pipeline components. A Pipeline framework is a collection of input -ports, table(s),output ports and actions (functions). - ---------------------------- -Receive and Transmit Driver ---------------------------- -Packets will be received in bulk and provided to LoadBalancer(LB) thread. -Transimit takes packets from worker threads in a dedicated ring and sent to -hardware queue. - ---------------- -Master Pipeline ---------------- -The Master component is part of all the IP Pipeline applications. This component -does not process any packets and should configure with Core 0, to allow -other cores for processing of the traffic. This component is responsible for -1. Initializing each component of the Pipeline application in different threads -2. Providing CLI shell for the user control/debug -3. Propagating the commands from user to the corresponding components - ----------------- -ARPICMP Pipeline ----------------- -This pipeline processes the APRICMP packets. - --------------- -TXRX Pipelines --------------- -The TXTX and RXRX pipelines are pass through pipelines to forward both ingress -and egress traffic to Loadbalancer. This is required when the Software -Loadbalancer is used. - ----------------------- -Load Balancer Pipeline ----------------------- -The vFW support both hardware and software balancing for load balancing of -traffic across multiple VNF threads. The Hardware load balancing require support -from hardware like Flow Director for steering of packets to application through -hardware queues. - -The Software Load balancer is also supported if hardware load balancing can't be -used for any reason. The TXRX along with LOADB pipeline provides support for -software load balancing by distributing the flows to Multiple vFW worker -threads. -Loadbalancer (HW or SW) distributes traffic based on the 5 tuple (src addr, src -port, dest addr, dest port and protocol) applying an XOR logic distributing to -active worker threads, thereby maintaining an affinity of flows to worker -threads. - ------------- -vFW Pipeline ------------- -The vFW performs the basic packet filtering and will drop the invalid and -malformed packets.The Dynamic packet filtering done using the connection tracker -library. The packets are processed in bulk and Hash table is used to maintain -the connection details. -Every TCP/UDP packets are passed through connection tracker library for valid -connection. The ACL library integrated to firewall provide rule based filtering. diff --git a/docs/testing/developer/design/04-SampleVNF_Design.rst b/docs/testing/developer/design/04-SampleVNF_Design.rst new file mode 100644 index 00000000..cdef7448 --- /dev/null +++ b/docs/testing/developer/design/04-SampleVNF_Design.rst @@ -0,0 +1,630 @@ +.. This work is licensed under a Creative Commons Attribution 4.0 International License. +.. http://creativecommons.org/licenses/by/4.0 +.. (c) OPNFV, Intel Corporation and others. + +.. OPNFV SAMPLEVNF Documentation design file. + +=================================== +SampleVNF Highlevel Design +=================================== + +Introduction +-------------- +This project provides a placeholder for various sample VNF (Virtual Network Function) +development which includes example reference architecture and optimization methods +related to VNF/Network service for high performance VNFs. This project provides benefits +to other OPNFV projects like Functest, Models, yardstick etc to perform real life +use-case based testing and NFVi characterization for the same. +The sample VNFs are Open Source approximations* of Telco grade VNF’s using optimized +VNF + NFVi Infrastructure libraries, with Performance Characterization of Sample† Traffic Flows. + • * Not a commercial product. Encourage the community to contribute and close the feature gaps. + • † No Vendor/Proprietary Workloads + +About DPDK +^^^^^^^^^^^ +The DPDK IP Pipeline Framework provides set of libraries to build a pipeline +application. In this document, CG-NAT application will be explained with its +own building blocks. + +This document assumes the reader possess the knowledge of DPDK concepts and IP +Pipeline Framework. For more details, read DPDK Getting Started Guide, DPDK +Programmers Guide, DPDK Sample Applications Guide. + +Scope +-------- +These application provides a standalone DPDK based high performance different +Virtual Network Function implementation. + + + +Common Code - L2L3 stack +------------------------- + +Introduction +^^^^^^^^^^^^^^^ +L2L3 stack comprises of a set of libraries which are commonly used by all +other VNF's. The different components of this stack is shown in the picture +below. + +.. image:: l2l3-components.png + +It comprises of following components. + + (i) Interface Manager + (ii) RTM Lock Library + (iii) ARP/ND & L2 adjacency Library + (iv) L3 stack components + + +Interface Manager +^^^^^^^^^^^^^^^^^ +Interface manager is a set of API's which acts as a wrapper for the physical +interfaces initialization & population. This set of api's assists in configuring +an ethernet device, setting up TX & RX queues & starting of the devices. It +provides various types of interfaces like L2 interface, IPv4/IPv6 interface. +It helps in Configuration (set/get) operations and status updates like (UP/DOWN) +from admin or operations. It provides Callback functions registration for other +components who wants to listen to interface status. It Maintains table of all +the interfaces present. It provides API for getting interface statistics. + +It Provides wrapper APIs on top of DPDKs LAG(link Aggregation) APIs, This +includes creating/deleting BOND interfaces, knowing the properties like Bond mode, +xmit policy, link up delay, link monitor frequency. + + +RTM Lock Library +^^^^^^^^^^^^^^^^^ +It provides basic lock & unlock functions which should be used for synchronization +purposes. + +ARP/ND & L2 adjacency Library +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + +The ARP/ND state machine is given in the following diagram. + +.. image:: state-machine.png + +This library provides api's for handling ARP/ICMPv4 & ND/ICMPV6 packets +handling. It provides api's for creating/deleting & populating an entry. +It handles ARP request/response messages, Handles ICMP v4 echo request & +echo response messages. It handles ND solicitation/advertisement messages +for IPV6 packets. It Provide API for L2 Adjacency creation/deletion and +retrieval based on nexthop & port_id. It handles Gratuitous ARP. + +:: + + Basic commands for ARP/ND table + p 1 arpls 0 (for ARP) + p 1 arpls 1 (for ND) + p 1 arpadd 0 <ip> <mac address> (for adding arp entry) + p 1 arpdel 0 <ip> (for deleting an arp entry) + p 1 arpreq 0 <ip> (for sending an arp request) + + +L3 stack Library +^^^^^^^^^^^^^^^^^ + +This library provides API for taking decision of whether pkt belongs to local +system or to forwarding.It Provides API for IPv4/IPv6 local packet out send +function. It Provides API for packet forwarding - LPM lookup function. + + +vFW - Design +============= + +Requirements +------------- + +Following are the design requierments of the vFW. + +- The firewall will examine packets and verify that they are appropriate for the + current state of the connection. Inappropriate packets will be discarded, and + counter(s) incremented. +- Support both IPv4 and IPv6 traffic type for TCP/UDP/ICMP. +- All packet inspection features like firewall, synproxy, connection tracker + in this component may be turned off or on through CLI commands +- The Static filtering is done thorugh ACL using DPDK libraries. The rules + can be added/modified through CLI commands. +- Multiple instance of the vFW Pipeline running on multipe cores should be + supported for scaling the performance scaling. +- Should follow the DPDK IP pipeline framework +- Should use the DPDK libraries and functionalities for better performance +- The memory should be allocated in Hugepages using DPDK RTE calls for better + performance. + +High Level Design +------------------- + +The Firewall performs basic filtering for malformed packets and dynamic packet +filtering incoming packets using the connection tracker library. +The connection data will be stored using a DPDK hash table. There will be one +entry in the hash table for each connection. The hash key will be based on +source address/port,destination address/port, and protocol of a packet. The +hash key will be processed to allow a single entry to be used, regardless of +which direction the packet is flowing (thus changing source and destination). +The ACL is implemented as libray stattically linked to vFW, which is used for +used for rule based packet filtering. + +TCP connections and UDP pseudo connections will be tracked separately even if +theaddresses and ports are identical. Including the protocol in the hash key +will ensure this. + +The Input FIFO contains all the incoming packets for vFW filtering. The vFW +Filter has no dependency on which component has written to the Input FIFO. +Packets will be dequeued from the FIFO in bulk for processing by the vFW. +Packets will be enqueued to the output FIFO. + +The software or hardware loadbalancing can be used for traffic distribution +across multiple worker threads. The hardware loadbalancing require ethernet +flow director support from hardware (eg. Fortville x710 NIC card). +The Input and Output FIFOs will be implemented using DPDK Ring Buffers. + +Components of vFW +----------------- + +In vFW, each component is constructed using packet framework pipelines. +It includes Rx and Tx Driver, Master pipeline, load balancer pipeline and +vfw worker pipeline components. A Pipeline framework is a collection of input +ports, table(s),output ports and actions (functions). + +Receive and Transmit Driver +^^^^^^^^^^^^^^^^^^^^^^^^^^^ +Packets will be received in bulk and provided to LoadBalancer(LB) thread. +Transimit takes packets from worker threads in a dedicated ring and sent to +hardware queue. + +Master Pipeline +^^^^^^^^^^^^^^^ +The Master component is part of all the IP Pipeline applications. This component +does not process any packets and should configure with Core 0, to allow +other cores for processing of the traffic. This component is responsible for +1. Initializing each component of the Pipeline application in different threads +2. Providing CLI shell for the user control/debug +3. Propagating the commands from user to the corresponding components + +ARPICMP Pipeline +^^^^^^^^^^^^^^^^ +This pipeline processes the APRICMP packets. + +TXRX Pipelines +^^^^^^^^^^^^^^ +The TXTX and RXRX pipelines are pass through pipelines to forward both ingress +and egress traffic to Loadbalancer. This is required when the Software +Loadbalancer is used. + +Load Balancer Pipeline +^^^^^^^^^^^^^^^^^^^^^^ +The vFW support both hardware and software balancing for load balancing of +traffic across multiple VNF threads. The Hardware load balancing require support +from hardware like Flow Director for steering of packets to application through +hardware queues. + +The Software Load balancer is also supported if hardware load balancing can't be +used for any reason. The TXRX along with LOADB pipeline provides support for +software load balancing by distributing the flows to Multiple vFW worker +threads. +Loadbalancer (HW or SW) distributes traffic based on the 5 tuple (src addr, src +port, dest addr, dest port and protocol) applying an XOR logic distributing to +active worker threads, thereby maintaining an affinity of flows to worker +threads. + +vFW Pipeline +^^^^^^^^^^^^ +The vFW performs the basic packet filtering and will drop the invalid and +malformed packets.The Dynamic packet filtering done using the connection tracker +library. The packets are processed in bulk and Hash table is used to maintain +the connection details. +Every TCP/UDP packets are passed through connection tracker library for valid +connection. The ACL library integrated to firewall provide rule based filtering. + + +vCGNAPT - Design +================= + +Introduction +^^^^^^^^^^^^^^ +This application implements vCGNAPT. The idea of vCGNAPT is to extend the life of +the service providers IPv4 network infrastructure and mitigate IPv4 address +exhaustion by using address and port translation in large scale. It processes the +traffic in both the directions. + +It also supports the connectivity between the IPv6 access network to IPv4 data network +using the IPv6 to IPv4 address translation and vice versa. + +Scope +^^^^^^ +This application provides a standalone DPDK based high performance vCGNAPT +Virtual Network Function implementation. + +Features +^^^^^^^^^ +The vCGNAPT VNF currently supports the following functionality: + • Static NAT + • Dynamic NAT + • Static NAPT + • Dynamic NAPT + • ARP (request, response, gratuitous) + • ICMP (terminal echo, echo response, passthrough) + • UDP, TCP and ICMP protocol passthrough + • Multithread support + • Multiple physical port support + • Limiting max ports per client + • Limiting max clients per public IP address + • Live Session tracking to NAT flow + • NAT64 + + +High Level Design +^^^^^^^^^^^^^^^^^^^ +The Upstream path defines the traffic from Private to Public and the downstream +path defines the traffic from Public to Private. The vCGNAPT has same set of +components to process Upstream and Downstream traffic. + +In vCGNAPT application, each component is constructed as IP Pipeline framework. +It includes Master pipeline component, load balancer pipeline component and vCGNAPT +pipeline component. + +A Pipeline framework is collection of input ports, table(s), output ports and +actions (functions). In vCGNAPT pipeline, main sub components are the Inport function +handler, Table and Table function handler. vCGNAPT rules will be configured in the +table which translates egress and ingress traffic according to physical port +information from which side packet is arrived. The actions can be forwarding to the +output port (either egress or ingress) or to drop the packet. + +vCGNAPT Background +^^^^^^^^^^^^^^^^^^^ +The idea of vCGNAPT is to extend the life of the service providers IPv4 network infrastructure +and mitigate IPv4 address exhaustion by using address and port translation in large scale. +It processes the traffic in both the directions. :: ++------------------+ +| +-----+ +| Private consumer | CPE ---- +| IPv4 traffic +-----+ | ++------------------+ | + | +-------------------+ +------------------+ + | | +------------+ - + |-> - Private IPv4 - vCGNAPT - Public - + |-> - access network - NAT44 - IPv4 traffic - + | | +------------+ - + | +-------------------+ +------------------+ ++------------------+ | +| +-----+ | +| Private consumer - CPE ---- +| IPv4 traffic +-----+ ++------------------+ + Figure: vCGNAPT deployment in Service provider network + + +Components of vCGNAPT +--------------------- +In vCGNAPT, each component is constructed as a packet framework. It includes Master pipeline +component, driver, load balancer pipeline component and vCGNAPT worker pipeline component. A +pipeline framework is a collection of input ports, table(s), output ports and actions +(functions). + +Receive and transmit driver +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +Packets will be received in bulk and provided to load balancer thread. The transmit takes +packets from worker thread in a dedicated ring and sent to the hardware queue. + +Master pipeline +^^^^^^^^^^^^^^^^ +This component does not process any packets and should configure with Core 0, +to save cores for other components which processes traffic. The component +is responsible for: + 1. Initializing each component of the Pipeline application in different threads + 2. Providing CLI shell for the user + 3. Propagating the commands from user to the corresponding components. + 4. ARP and ICMP are handled here. + +Load Balancer pipeline +^^^^^^^^^^^^^^^^^^^^^^^ +Load balancer is part of the Multi-Threaded CGMAPT release which distributes +the flows to Multiple ACL worker threads. + +Distributes traffic based on the 2 or 5 tuple (source address, source port, +destination address, destination port and protocol) applying an XOR logic +distributing the load to active worker threads, thereby maintaining an +affinity of flows to worker threads. + +Tuple can be modified/configured using configuration file + +vCGNAPT - Static +------------------ +The vCGNAPT component performs translation of private IP & port to public IP & +port at egress side and public IP & port to private IP & port at Ingress side +based on the NAT rules added to the pipeline Hash table. The NAT rules are +added to the Hash table via user commands. The packets that have a matching +egress key or ingress key in the NAT table will be processed to change IP & +port and will be forwarded to the output port. The packets that do not have a +match will be taken a default action. The default action may result in drop of +the packets. + +vCGNAPT- Dynamic +----------------- +The vCGNAPT component performs translation of private IP & port to public IP & port +at egress side and public IP & port to private IP & port at Ingress side based on the +NAT rules added to the pipeline Hash table. Dynamic nature of vCGNAPT refers to the +addition of NAT entries in the Hash table dynamically when new packet arrives. The NAT +rules will be added to the Hash table automatically when there is no matching entry in +the table and the packet is circulated through software queue. The packets that have a +matching egress key or ingress key in the NAT table will be processed to change IP & +port and will be forwarded to the output port defined in the entry. + +Dynamic vCGNAPT acts as static one too, we can do NAT entries statically. Static NAT +entries port range must not conflict to dynamic NAT port range. + +vCGNAPT Static Topology: +-------------------------- +IXIA(Port 0)-->(Port 0)VNF(Port 1)-->(Port 1) IXIA +operation: + Egress --> The packets sent out from ixia(port 0) will be CGNAPTed to ixia(port 1). + Igress --> The packets sent out from ixia(port 1) will be CGNAPTed to ixia(port 0). + +vCGNAPT Dynamic Topology (UDP_REPLAY): +-------------------------------------- +IXIA(Port 0)-->(Port 0)VNF(Port 1)-->(Port 0)UDP_REPLAY +operation: + Egress --> The packets sent out from ixia will be CGNAPTed to L3FWD/L4REPLAY. + Ingress --> The L4REPLAY upon reception of packets (Private to Public Network), + will immediately replay back the traffic to IXIA interface. (Pub -->Priv). + +How to run L4Replay: +-------------------- + 1. After the installation of ISB on L4Replay server + go to /opt/isb_bin + 2. ./UDP_Replay -c core_mask -n no_of_channels(let it be as 2) -- -p PORT_MASK --config="(port,queue,lcore)" + eg: ./UDP_Replay -c 0xf -n 4 -- -p 0x3 --config="(0,0,1)" + + +vACL - Design +================= + +Introduction +-------------- +This application implements Access Control List (ACL). ACL is typically used for rule +based policy enforcement. It restricts access to a destination IP address/port based +on various header fields, such as source IP address/port, destination IP address/port +and protocol. It is built on top of DPDK and uses the packet framework infrastructure. + +Scope +------ +This application provides a standalone DPDK based high performance ACL Virtual Network +Function implementation. + +High Level Design +------------------ +The ACL Filter performs bulk filtering of incoming packets based on rules in current ruleset, +discarding any packets not permitted by the rules. The mechanisms needed for building the +rule database and performing lookups are provided by the DPDK API. +http://dpdk.org/doc/api/rte__acl_8h.html + +The Input FIFO contains all the incoming packets for ACL filtering. Packets will be dequeued +from the FIFO in bulk for processing by the ACL. Packets will be enqueued to the output FIFO. +The Input and Output FIFOs will be implemented using DPDK Ring Buffers. + +The DPDK ACL example: http://dpdk.org/doc/guides/sample_app_ug/l3_forward_access_ctrl.html +#figure-ipv4-acl-rule contains a suitable syntax and parser for ACL rules. + +Components of ACL +------------------ +In ACL, each component is constructed as a packet framework. It includes Master pipeline +component, driver, load balancer pipeline component and ACL worker pipeline component. A +pipeline framework is a collection of input ports, table(s), output ports and actions +(functions). + +Receive and transmit driver +--------------------------- +Packets will be received in bulk and provided to load balancer thread. The transmit takes +packets from worker thread in a dedicated ring and it is sent to the hardware queue. + +Master +------- +This component does not process any packets and should configure with Core 0, +to save cores for other components which processes traffic. The component +is responsible for: + 1. Initializing each component of the Pipeline application in different threads + 2. Providing CLI shell for the user + 3. Propagating the commands from user to the corresponding components. + 4. ARP and ICMP are handled here. + +Load Balancer +-------------- +Load balancer is part of the Multi-Threaded ACL release which distributes +the flows to Multiple ACL worker threads. + +Distributes traffic based on the 5 tuple (source address, source port, destination +address, destination port and protocol) applying an XOR logic distributing the +load to active worker threads, thereby maintaining an affinity of flows to +worker threads. + +ACL +--- +Visit the following link for DPDK ACL library implementation. +http://dpdk.org/doc/api/rte__acl_8h.html +http://dpdk.org/doc/guides/prog_guide/packet_classif_access_ctrl.html + +Provides shadow copy for runtime rule configuration support + +Implements policy based packet forwarding + +vPE - Design +============= + +Introduction +--------------- +An Edge Router typically sits between two networks such as the provider core +network and the provider access network. In the below diagram, Customer Edge +(CE) Router sits in the provider access network and MPLS cloud network +represents the provide core network. +The edge router processes the traffic in both the directions. The functionality +of the Edge Router varies while processing each direction of traffic. The +packets to the core network will be filtered, classified and metered with QoS +parameters. The packets to the access network will be shaped according to the +subscription policy. +The idea of Edge Router application is to provide the benchmark for the +functionality of Provider Edge routers in each direction. + +The DPDK IP Pipeline Framework provides set of libraries to build a pipeline +application. The Provider Edge Router functionality delivered as a virtual +network function (VNF) is integrated with DPDK, optimized for Intel hardware +architecture. +This document assumes the reader possess the knowledge of DPDK concepts and +IP Pipeline Framework. For more details, read DPDK Getting Started Guide, DPDK +Programmers Guide, DPDK Sample Applications Guide. + +Scope +------ +This application provides a standalone DPDK based high performance Provide +Edge Router Network Function implementation. + +High Level Design +------------------- +The Edge Router application processes the traffic between Customer and the core +network. +The Upstream path defines the traffic from Customer to Core and the downstream +path defines the traffic from Core to Customer. The Edge Router has different +set of components to process Upstream and Downstream traffic. + +In Edge Router application, each component is constructed as building blocks in +IP Pipeline framework. As in Pipeline framework, each component has its own +input ports, table and output ports. The rules of the component will be +configured in the table which decides the path of the traffic and any action to +be performed on the traffic. The actions can be forwarding to the output port, +forwarding to the next table or drop. For more details, please refer Section 24 +of DPDK Programmers Guide (3). + +The Core-to-Customer traffic is mentioned as downstream. For downstream +processing, Edge Router has the following functionalities in Downstream + + ---> Packet Rx --> Routing --> Traffic Manager --> Packet Tx --> + + Routing + To identify the route based on the destination IP. + To provide QinQ label based on the destination IP. + Encapsulation + Updates the MAC address based on the route entry. + Appends the QinQ label based on the route entry. + Traffic Manager + To perform QoS traffic management (5-level hierarchical scheduling) based on + the predefined set of Service Level Agreements (SLAs) + SVLAN, CVLAN, DSCP fields are used to determine transmission priority. + Traffic Manager Profile which contains the SLA parameters are provided as + part of the application. + +The Customer-to-Core traffic is mentioned as upstream. For upstream processing, +Edge Router has the following functionalities in Upstream. + + ---> Packet Rx --> ACL filters --> Flow control --> Metering Policing & + Marking --> Routing --> Queueing & Packet Tx --> + + Firewall + To filter the unwanted packets based on the defined ACL rules. + Source IP, Destination IP, Protocol, Source Port and Destination Port are + used to derive the ACL rules. + Flow Classification + To classify the packet based on the QinQ label + To assign a specific flow id based on the classification. + Metering + Two stages of QoS traffic metering and policing is applied. + 1st stage is performed per flow ID using trTCM algorithm + 2nd stage is performed per flow ID traffic class using trTCM algorithm + Packets will be either dropped or marked Green, Yellow, Red based on the + metering rules. + Routing + To identify the route based on the destination IP + To provide MPLS label to the packets based on destination IP. + Encapsulation + Updates the MAC address based on the route entry. + Appends the MPLS label based on the route entry. + Update the packet color in MPLS EXP field in each MPLS header. + +Components of vPE +------------------- +The vPE has downstream and upstream pipelines controlled by Master component. +Edge router processes two different types of traffic through pipelines +I. Downstream (Core-to-Customer) + 1. Receives TCP traffic from core + 2. Routes the packet based on the routing rules + 3. Performs traffic scheduling based on the traffic profile + a. Qos scheduling is performed using token bucket algorithm + SVLAN, CVLAN, DSCP fields are used to determine transmission priority. + 4. Appends QinQ label in each outgoing packet. +II. Upstream (Customer-to-Core) + 1. Receives QinQ labelled TCP packets from Customer + 2. Removes the QinQ label + 3. Classifies the flow using QinQ label and apply Qos metering + a. 1st stage Qos metering is performed with flow ID using trTCM algorithm + b. 2nd stage Qos metering is performed with flow ID and traffic class using + trTCM algorithm + c. traffic class maps to DSCP field in the packet. + 4. Routes the packet based on the routing rules + 5. Appends two MPLS labels in each outgoing packet. + +Master Component +----------------- +The Master component is part of all the IP Pipeline applications. This +component does not process any packets and should configure with Core0, +to save cores for other components which processes traffic. The component +is responsible for + 1. Initializing each component of the Pipeline application in different threads + 2. Providing CLI shell for the user + 3. Propagating the commands from user to the corresponding components. + +Upstream and Downstream Pipelines +---------------------------------- +The downstream will have Firewall, Pass-through, Metering and Routing pipelines. +The upstream will have Pass-through and Routing pipelines. + +To run the VNF, execute the following: +isb_root/VNFs/vPE$ ./build/ip_pipeline -p 0x3 \ + -f config/auto_combo_1_instances_1_queues_2_ports_v2.cfg \ + -s config/auto_combo_1_instances_1_queues_2_ports_v2.txt + +Prox - Packet pROcessing eXecution engine +========================================== + +Overview: +---------- +Packet pROcessing eXecution Engine (PROX) which is a DPDK application. +PROX can do operations on packets in a highly configurable manner. +The PROX application is also displaying performance statistics that can +be used for performance investigations. +Intel® DPPD - PROX is an application built on top of DPDK which allows creating +software architectures, such as the one depicted below, through small and readable +configuration files. + +.. image:: images/prox-qo-img01.png + +The figure shows that each core is executing a set of tasks. Currently, +a task can be any one of the following: +1. Classify +2. Drop +3. Basic Forwarding (no touch) +4. L2 Forwarding (change MAC) +5. GRE encap/decap +6. Load balance based on packet fields +7. Symmetric load balancing +8. QinQ encap/decap IPv4/IPv6 +9. ARP +10. QoS +11. Routing +12. Unmpls +13. Policing +14. ACL ... + +One of the example configurations that is distributed with the source code is a +Proof of Concept (PoC) implementation of a Broadband Network Gateway (BNG) with Quality of Service (QoS). +The software architecture for this PoC is presented below. + +.. image:: images/prox-qo-img02.png + +The display shows per task statistics through an ncurses interface. +Statistics include: estimated idleness; per second statistics for packets received, +transmitted or dropped; per core cache occupancy; cycles per packet. +These statistics can help pinpoint bottlenecks in the system. +This information can then be used to optimize the configuration. +Other features include debugging support, scripting, +Open vSwitch support... A screenshot of the display is provided below. + +.. image:: images/prox-screen-01.png diff --git a/docs/testing/developer/design/images/prox-qo-img01.png b/docs/testing/developer/design/images/prox-qo-img01.png Binary files differnew file mode 100644 index 00000000..b3e32d89 --- /dev/null +++ b/docs/testing/developer/design/images/prox-qo-img01.png diff --git a/docs/testing/developer/design/images/prox-qo-img02.png b/docs/testing/developer/design/images/prox-qo-img02.png Binary files differnew file mode 100644 index 00000000..0b40c03a --- /dev/null +++ b/docs/testing/developer/design/images/prox-qo-img02.png diff --git a/docs/testing/developer/design/images/prox-screen-01.png b/docs/testing/developer/design/images/prox-screen-01.png Binary files differnew file mode 100644 index 00000000..d97645d7 --- /dev/null +++ b/docs/testing/developer/design/images/prox-screen-01.png |