aboutsummaryrefslogtreecommitdiffstats
path: root/docs/report/introduction
diff options
context:
space:
mode:
Diffstat (limited to 'docs/report/introduction')
-rw-r--r--docs/report/introduction/methodology.rst1
-rw-r--r--docs/report/introduction/methodology_nfv_service_density.rst106
2 files changed, 107 insertions, 0 deletions
diff --git a/docs/report/introduction/methodology.rst b/docs/report/introduction/methodology.rst
index 1f9fcfe7fb..da8f859d8d 100644
--- a/docs/report/introduction/methodology.rst
+++ b/docs/report/introduction/methodology.rst
@@ -18,6 +18,7 @@ Test Methodology
methodology_kvm_vms_vhost_user
methodology_lxc_drc_container_memif
methodology_k8s_container_memif
+ methodology_nfv_service_density
methodology_vpp_device_functional
methodology_ipsec_on_intel_qat
methodology_trex_traffic_generator
diff --git a/docs/report/introduction/methodology_nfv_service_density.rst b/docs/report/introduction/methodology_nfv_service_density.rst
new file mode 100644
index 0000000000..2946ba2777
--- /dev/null
+++ b/docs/report/introduction/methodology_nfv_service_density.rst
@@ -0,0 +1,106 @@
+NFV Service Density
+-------------------
+
+Network Function Virtualization (NFV) service density tests focus on
+measuring total per server throughput at varied NFV service “packing”
+densities with vswitch providing host dataplane. The goal is to compare
+and contrast performance of a shared vswitch for different network
+topologies and virtualization technologies, and their impact on vswitch
+performance and efficiency in a range of NFV service configurations.
+
+Each NFV service instance consists of a set of Network Functions (NFs),
+running in VMs (VNFs) or in Containers (CNFs), that are connected into a
+virtual network topology using VPP vswitch running in Linux user-mode.
+Multiple service instances share the vswitch that in turn provides per
+service chain forwarding context(s). In order to provide a most complete
+picture, each network topology and service configuration is tested in
+different service density setups by varying two parameters:
+
+- Number of service instances (e.g. 1,2,4..10).
+- Number of NFs per service instance (e.g. 1,2,4..10).
+
+The initial implementation of NFV service density tests in
+|csit-release| is using two NF applications:
+
+- VNF: DPDK L3fwd running in KVM VM, configured with /8 IPv4 prefix
+ routing. L3fwd got chosen as a lightweight fast IPv4 VNF application,
+ and follows CSIT approach of using DPDK sample applications in VMs for
+ performance testing.
+- CNF: VPP running in Docker Container, configured with /24 IPv4 prefix
+ routing. VPP got chosen as a fast IPv4 NF application that supports
+ required memif interface (L3fwd does not). This is similar to all
+ other Container tests in CSIT that use VPP.
+
+Tests are designed such that in all tested cases VPP vswitch is the most
+stressed application, as for each flow vswitch is processing each packet
+multiple times, whereas VNFs and CNFs process each packets only once. To
+that end, all VNFs and CNFs are allocated enough resources to not become
+a bottleneck.
+
+Service Configurations
+~~~~~~~~~~~~~~~~~~~~~~
+
+Following NFV network topologies and configurations are tested:
+
+- VNF Service Chains (VSC) with L2 vswitch
+
+ - *Network Topology*: Sets of VNFs dual-homed to VPP vswitch over
+ virtio-vhost links. Each set belongs to separate service instance.
+ - *Network Configuration*: VPP L2 bridge-domain contexts form logical
+ service chains of VNF sets and connect each chain to physical
+ interfaces.
+
+- CNF Service Chains (CSC) with L2 vswitch
+
+ - *Network Topology*: Sets of CNFs dual-homed to VPP vswitch over
+ memif links. Each set belongs to separate service instance.
+ - *Network Configuration*: VPP L2 bridge-domain contexts form logical
+ service chains of CNF sets and connect each chain to physical
+ interfaces.
+
+- CNF Service Pipelines (CSP) with L2 vswitch
+
+ - *Network Topology*: Sets of CNFs connected into pipelines over a
+ series of memif links, with edge CNFs single-homed to VPP vswitch
+ over memif links. Each set belongs to separate service instance.
+ - *Network Configuration*: VPP L2 bridge-domain contexts connect each
+ CNF pipeline to physical interfaces.
+
+Thread-to-Core Mapping
+~~~~~~~~~~~~~~~~~~~~~~
+
+CSIT defines specific ratios for mapping software threads of vswitch and
+VNFs/CNFs to physical cores, with separate ratios defined for main
+control threads and data-plane threads.
+
+In |csit-release| NFV service density tests run on Intel Xeon testbeds
+with Intel Hyper-Threading enabled, so each physical core is associated
+with a pair of sibling logical cores corresponding to the hyper-threads.
+
+|csit-release| executes tests with the following software thread to
+physical core mapping ratios:
+
+- vSwitch
+
+ - Data-plane on single core
+
+ - (data:core) = (1:1) => 2dt1c - 2 Data-plane Threads on 1 Core.
+ - (main:core) = (1:1) => 1mt1c - 1 Main Thread on 1 Core.
+
+ - Data-plane on two cores
+
+ - (data:core) = (1:2) => 4dt2c - 4 Data-plane Threads on 2 Cores.
+ - (main:core) = (1:1) => 1mt1c - 1 Main Thread on 1 Core.
+
+- VNF and CNF
+
+ - Data-plane on single core
+
+ - (data:core) = (1:1) => 2dt1c - 2 Data-plane Threads on 1 Core per
+ NF.
+ - (main:core) = (2:1) => 2mt1c - 2 Main Threads on 1 Core, 1 Thread
+ per NF, core shared between two NFs.
+
+Maximum tested service densities are limited by a number of physical
+cores per NUMA. |csit-release| allocates cores within NUMA0. Support for
+multi NUMA tests is to be added in future release. \ No newline at end of file