2.11. Network Services Benchmarking (NSB)¶
2.11.1. Abstract¶
This chapter provides an overview of the NSB, a contribution to OPNFV Yardstick from Intel.
2.11.2. Overview¶
Network Services Benchmarking (NSB) uses the Yardstick framework for performing VNF and NFVI characterisation in an NFV environment.
For VNF characterisation, NSB will onboard a VNF, source and sink traffic to it via traffic generators, and collect a variety of key performance indicators (KPI) during VNF execution. The stream of KPI data is stored in a database, and it is visualized in a performance-visualization dashboard.
For NFVI characterisation, a fixed test VNF, called PROX is used. PROX implements a suite of test cases and visualizes the output data of the test suite. The PROX test cases implement various execution kernels found in real-world VNFs, and the output of the test cases provides an indication of the fitness of the infrastructure for running NFV services, in addition to indicating potential performance optimizations for the NFVI.
NSB extends the Yardstick framework to do VNF characterization in three different execution environments - bare metal i.e. native Linux environment, standalone virtual environment and managed virtualized environment (e.g. OpenStack). It also brings in the capability to interact with external traffic generators, both hardware and software based, for triggering and validating the traffic according to user defined profiles.
NSB extension includes:
Generic data models of Network Services, based on ETSI spec ETSI GS NFV-TST001
Standalone context for VNF testing SRIOV, OVS-DPDK, etc
Generic VNF configuration models and metrics implemented with Python classes
Traffic generator features and traffic profiles
L1-L3 stateless traffic profiles
L4-L7 state-full traffic profiles
Tunneling protocol/network overlay support
Scenarios that handle NSB test cases execution
NSPerf - scenario that handles generic NSB test case execution (setup and init tg/vnf, trigger traffic on tg, collect kpi)
NSPerf-RFC2544 - scenario that allows repeatable triggering of traffic on traffic generators until test case acceptance criteria is met (for example RFC2544 binary search)
Test case samples
Ping
Trex
vPE, vCGNAT, vFirewall etc - ipv4 throughput, latency etc
Traffic generators i.e. Trex, ab/nginx, ixia, iperf, etc
KPIs for a given use case:
System agent support for collecting NFVi KPI. This includes:
CPU statistic
Memory BW
OVS-DPDK Stats
Network KPIs e.g. inpackets, outpackets, thoughput, latency
VNF KPIs e.g. packet_in, packet_drop, packet_fwd
2.11.3. Architecture¶
The Network Service (NS) defines a set of Virtual Network Functions (VNF) connected together using NFV infrastructure.
The Yardstick NSB extension can support multiple VNFs created by different vendors including traffic generators. Every VNF being tested has its own data model. The Network service defines a VNF modelling on base of performed network functionality. The part of the data model is a set of the configuration parameters, number of connection points used and flavor including core and memory amount.
ETSI defines a Network Service as a set of configurable VNFs working in some NFV Infrastructure connecting each other using Virtual Links available through Connection Points. The ETSI MANO specification defines a set of management entities called Network Service Descriptors (NSD) and VNF Descriptors (VNFD) that define real Network Service. The picture below makes an example how the real Network Operator use-case can map into ETSI Network service definition.
Network Service framework performs the necessary test steps. It may involve:
Interacting with traffic generator and providing the inputs on traffic type / packet structure to generate the required traffic as per the test case. Traffic profiles will be used for this.
Executing the commands required for the test procedure and analyses the command output for confirming whether the command got executed correctly or not e.g. as per the test case, run the traffic for the given time period and wait for the necessary time delay.
Verify the test result.
Validate the traffic flow from SUT.
Fetch the data from SUT and verify the value as per the test case.
Upload the logs from SUT onto the Test Harness server
Retrieve the KPI’s provided by particular VNF
2.11.3.1. Components of Network Service¶
Models for Network Service benchmarking: The Network Service benchmarking requires the proper modelling approach. The NSB provides models using Python files and defining of NSDs and VNFDs.
The benchmark control application being a part of OPNFV Yardstick can call that Python models to instantiate and configure the VNFs. Depending on infrastructure type (bare-metal or fully virtualized) that calls could be made directly or using MANO system.
Traffic generators in NSB: Any benchmark application requires a set of traffic generator and traffic profiles defining the method in which traffic is generated.
The Network Service benchmarking model extends the Network Service definition with a set of Traffic Generators (TG) that are treated same way as other VNFs being a part of benchmarked network service. Same as other VNFs the traffic generator are instantiated and terminated.
Every traffic generator has own configuration defined as a traffic profile and a set of KPIs supported. The python models for TG is extended by specific calls to listen and generate traffic.
The stateless TREX traffic generator: The main traffic generator used as Network Service stimulus is open source TREX tool.
The TREX tool can generate any kind of stateless traffic.
+--------+ +-------+ +--------+
| | | | | |
| Trex | ---> | VNF | ---> | Trex |
| | | | | |
+--------+ +-------+ +--------+
Supported testcases scenarios:
Correlated UDP traffic using TREX traffic generator and replay VNF.
Using different IMIX configuration like pure voice, pure video traffic etc
Using different number IP flows e.g. 1, 1K, 16K, 64K, 256K, 1M flows
Using different number of rules configured e.g. 1, 1K, 10K rules
For UDP correlated traffic following Key Performance Indicators are collected for every combination of test case parameters:
RFC2544 throughput for various loss rate defined (1% is a default)
2.11.3.2. KPI Collection¶
KPI collection is the process of sampling KPIs at multiple intervals to allow for investigation into anomalies during runtime. Some KPI intervals are adjustable. KPIs are collected from traffic generators and NFVI for the SUT. There is already some reporting in NSB available, but NSB collects all KPIs for analytics to process.
Below is an example list of basic KPIs:
Throughput
Latency
Packet delay variation
Maximum establishment rate
Maximum tear-down rate
Maximum simultaneous number of sessions
Of course, there can be many other KPIs that will be relevant for a specific NFVI, but in most cases these KPIs are enough to give you a basic picture of the SUT. NSB also uses collectd in order to collect the KPIs. Currently the following collectd plug-ins are enabled for NSB testcases:
Libvirt
Interface stats
OvS events
vSwitch stats
Huge Pages
RAM
CPU usage
Intel® PMU
Intel® RDT
2.11.4. Graphical Overview¶
NSB Testing with Yardstick framework facilitate performance testing of various VNFs provided.
+-----------+
| | +-------------+
| vPE | -->| TGen Port 0 |
| TestCase | | +-------------+
| | |
+-----------+ +---------------+ +-------+ |
| | ---> | VNF | <--->
+-----------+ | Yardstick | +-------+ |
| Test Case | --> | NSB Testing | |
+-----------+ | | |
| | | |
| +---------------+ |
+-----------+ | +-------------+
| Traffic | -->| TGen Port 1 |
| patterns | +-------------+
+-----------+
Figure 1: Network Service - 2 server configuration
2.11.4.1. VNFs supported for chracterization¶
CGNAPT - Carrier Grade Network Address and port Translation
vFW - Virtual Firewall
vACL - Access Control List
- PROX - Packet pROcessing eXecution engine:
VNF can act as Drop, Basic Forwarding (no touch), L2 Forwarding (change MAC), GRE encap/decap, Load balance based on packet fields, Symmetric load balancing
QinQ encap/decap IPv4/IPv6, ARP, QoS, Routing, Unmpls, Policing, ACL
UDP_Replay