The Bridge - Issue 3, 2021 - 19

Feature
Analyzing Service Mesh Performance
generator. To highlight the performance impact of a
service mesh and its set of data plane proxies, consider
running Meshery in two different environments outside
the Kubernetes cluster.
* First: load generator running as a process outside of
Kubernetes cluster in master-vm.
Figure 2: Snippet of the Service Mesh Performance (SMP) specification
describing how to capture statistical analysis test results.
V. DEFINING DEPLOYMENTS
Virtualized deployments involve deploying microservice
orchestration and service mesh stack in virtual machines
(VMs). Although bare metal usage has performance
benefits, customers often use VMs to provide hardwarelevel
isolation between various applications. This
deployment involves two VMs across two nodes, with
one acting as a Kubernetes master with the other a
worker node. Customers deploy VMs on a single nonuniform
memory access (NUMA) node to avoid cross
traffic on an ultra path interconnect (UPI). Results in
virtualized testing have shown that depending on pinning
of QEMU threads to a set of isolated cores - either
sequentially or clustering the threads together to all the
cores - tail latencies are heavily impacted.
Microservice deployments could use a wide variety
of deployment scenarios. The following list provides a
sample set of how a service mesh performance could
be analyzed either on a same node or in a multi-node
cluster:
* Pod to pod communication.
* Pod to service communication.
* Ingress controller to pod and vice-versa.
* Load balancer to pod and vice-versa.
* Pod to Egress Gateway.
* Mutual transport layer security (TLS) termination across
any of the above endpoints.
* Different security rules and policies.
* Communication protocol.
These considerations are illustrated in a typical workload
deployment as shown in Figure 3.
An example of deployment with Kubernetes is an
orchestrator using the open-source CNI [8] and deployed
in VMs. The host infrastructure has the Open vSwitch
data plane development kit (OVS-DPDK) for switching,
which can be extended for VMs to leverage single-root
I/O virtualization (SR-IOV). Fortio [9] can be the load
* Second: load generator running as a bare metal process
on master-host.
Figure 3: An enterprise workload deployment example
A. Automating Performance Measurements
Meshery is ideal tooling in that it provides lifecycle
management of a large number of service meshes
and sample applications which need to be provisioned,
configured, and deprovisioned in the process of analyzing
service mesh performance. Meshery is capable of
generating load, baselining, and comparing performance
results. The canonical implementation of this specification
is implemented in Meshery, see Figure 4.
Figure 4: Meshery's load generators can be deployed in the same cluster
under test or outside of the cluster test.
Acknowledging the living nature of user deployments,
integration of automated performance testing into
continuous integration systems helps users deploy new
versions of their applications or new configurations of
their infrastructure (including service mesh configuration)
HKN.ORG
19
http://www.HKN.ORG
The Bridge - Issue 3, 2021

Table of Contents for the Digital Edition of The Bridge - Issue 3, 2021
