Tune up your Kubernetes Application Performance with a small DNS Configuration

Jose Angel Munoz - Sep 3 '22 - - Dev Community

Introduction

One of the most common bottlenecks when communicating Kubernetes pods is DNS resolution. Analyzing network traffic within Kubernetes is not trivial.

This article will help you to understand how to analyze and debug HTTP and DNS network traffic in Kubernetes.

With your comments and help, I would like to improve this small article as much as possible to reach every profile and help any developer, network engineer, or Kubernetes enthusiast.

What is DNS?

DNS is a protocol that allows a computer to resolve a Domain Name (like dev.to) written by a human to an IP address so browsers can load Internet resources. DNS is called the phonebook of the Internet.

How to debug DNS network traffic - Wireshark

There are many tools to analyze network traffic, but needless to say, the most common one is Wireshark. It is a powerful tool that allows you to analyze network traffic and extract information from it. We will analyze DNS traffic from our computers and compare it with the traffic from a pod in a Kubernetes cluster.

What happens when calling a web page?

To understand what happens with the DNS when requesting a web page, this diagram from @wassimchegham is a good start.

The DNS request, the TCP Connection, the HTTP Request, and the HTTP Response can be displayed using different filters and the Statistics - Flow Graph:

diagram

In the next section, we will learn how to create it.

Creating the Wireshark Flow Graph

Install Wireshark on your favorite Operating System (Mac, Windows, or Linux) and open it.

The initial screen should be like this:

Wireshark Initial Screen

Preparing the filter

Before capturing traffic, let's find the IP of the domain www.example.com.

Open a Terminal or CMD Console and run:



nslookup www.example.com
Enter fullscreen mode Exit fullscreen mode

The answer will look like this:



Server: 8.8.8.8
Address: 8.8.8.8#53

Non-authoritative answer:
Name: www.example.com
Address: 93.184.216.34
Enter fullscreen mode Exit fullscreen mode

The Server field indicates the DNS server configured in the environment to which nslookup will request to resolve the name www.example.com

Find the name under Non-authoritative answer: and the IP associated. In the example 93.184.216.34.

Under using this filter: option on the main screen, input:

tcp port http or port 53 or dst host 93.184.216.34

93.184.216.34 is the address resolved with nslookup

Capturing traffic

Select the interface to start capturing network traffic. Usually, it should be the interface where the Packet/s column is constantly changing, which would indicate the presence of live traffic).

Active Interface

Press the blue fin icon to start the traffic capture.

Capture Traffic

Request the URL in the Local Client

If you are using Linux or Mac, run the following from a Terminal:



curl http://example.com
Enter fullscreen mode Exit fullscreen mode

From Windows, use Powershell and run:



Invoke-WebRequest http://www.example.com
Enter fullscreen mode Exit fullscreen mode

Stop capturing traffic by just pressing the big red button. Your screen should look like this:

http traffic

To get the promised Flow Graph, select the Statistics menu and Flow Graph

Understanding the captured traffic

When analyzing the captured traffic, there are two parts differentiated. The DNS (In cyan) and the HTTP (in green)

The first line (In cyan) shows the DNS request from our IP address to the DNS Server (8.8.8.8 in the example).

The second cyan line shows the DNS response. Select it to find the IP Address of the requested domain in the Packet Details Window. Under Domain Name System (Response) - Answers like shown in the image below.

DNS response

The 3-Way Handshaking (SYN-ACK)

Found in the following three lines (in green).

Before a client and a server can exchange data (payload), they must establish a TCP connection via the TCP 3-way handshake.

SYN - The client sends a SYN (Synchronize) packet to the server.
SYN-ACK- The server sends a SYN-ACK (Synchronize Acknowledge) packet to the client.
ACK - The client sends an ACK (Acknowledge) packet to the server.

In the image, it is after the DNS request.

3-way

The Request and Connection Close (FIN ACK)

Next to the SYN-ACK, comes the HTTP GET and the ACK from the server with the HTTP response.

And finally, the TCP requests FIN-ACK Packets to close the connection.

fin-ack

DNS on Kubernetes

To understand how a Domain Name is resolved in a pod, first, let's create a single deployment:



kubectl create deployment nginx --image nginx
Enter fullscreen mode Exit fullscreen mode

To check if the deployment is ready, run:



kubectl get deployments -o wide
Enter fullscreen mode Exit fullscreen mode 


NAME    READY   UP-TO-DATE   AVAILABLE   AGE   CONTAINERS   IMAGES   SELECTOR
nginx   1/1     1            1           26m   nginx        nginx    app=nginx
Enter fullscreen mode Exit fullscreen mode

Remember that the DNS resolution inside a container - like any Linux system - is driven by the /etc/resolv.conf config file.



kubectl exec deployments/nginx -- cat /etc/resolv.conf
Enter fullscreen mode Exit fullscreen mode

The /etc/resolv.conf file inside the container looks like this by default:



search default.svc.cluster.local svc.cluster.local cluster.local
nameserver 10.96.0.10
options ndots:5
Enter fullscreen mode Exit fullscreen mode

By default, there are three or more search Domains in a Kubernetes configuration. The example above comes from a Minikube Cluster with three local search domains specified.

Pay attention to the ndots:5 option. It is important to understand, how both search and ndots settings work together.

To understand both concepts, refer to the resolv.conf(5) Linux man page

The search represent the search path for a particular domain. Interestingly dev.to or example.com are not FQDN (fully qualified domain name). A standard convention that most DNS resolvers follow is that if a domain ends with a dot (.) (representing the root zone), the Domain is FQDN. Some resolvers try to act smart and append the dot (.) themselves. So dev.to. is an FQDN, dev.to is not.

One important point from the resolv.conf(5) Linux man page For environments with multiple subdomains please read options ndots:n to avoid unnecessary traffic for the root-dns-servers. Note that this process may be slow and can generate a lot of network traffic if the servers for the listed domains are not local and that queries will time out if no server is available for one of the domains.

The ndots represents the threshold value of the number of dots in a query name to consider it a "fully qualified" domain name.

If ndots is five (5) (the default in Kubernetes), and the name contains less than five (5) dots inside it, the syscall will try to resolve it sequentially through all local search domains first and - in case none succeed - it will resolve as an absolute name only at last. For instance, the domain name www.example.com contains two dots (.), and the number of dots (.) is less than the value of ndots.

Therefore, when querying a domain name, the DNS query iterates through all search paths until the answer contains a NOERROR code.

  1. www.example.com.<namespace>.svc.cluster.local
  2. www.google.com.svc.cluster.local
  3. www.google.cluster.local
  4. www.google.com

It is important to remark that the DNS requests the A and AAAA records in parallel. The single-request option in /etc/resolv.conf has a default configuration to perform parallel IPv4 and IPv6 lookups. This option can be disabled using the configuration option single-request in the /etc/resolv.conf configuration file.



option single-request
Enter fullscreen mode Exit fullscreen mode

Capturing Traffic in a Kubernetes Pod

There are different ways to capture traffic in a Kubernetes Pod. All the examples are based on the latest Kubernetes functionality using Ephemeral Debug Containers

First, let's create a tcpdump debug image. Wireshark will use it to display the attached pod traffic. The build process creates a multi-platform image to be able to run in linux/amd64 and linux/arm64

The Dockerfile 



ARG build_for=linux/amd64,linux/arm64
FROM alpine:latest as base

LABEL maintainer="<yourusername>"

# Packages to build image requirements
RUN apk add --no-cache \
    tcpdump

ENTRYPOINT [ "tcpdump" ]
CMD [ "-i", "any" ]
Enter fullscreen mode Exit fullscreen mode

The build command

To create a multi-platform image in dockerhub let's create a builder instance and use the extended build capabilities with BuildKit.



docker buildx create --name buildx --driver-opt network=host --use
docker buildx inspect --bootstrap
docker buildx build -t <yourusername>/tcpdump:v1.0.0 --platform linux/amd64 --platform linux/arm64 --file Dockerfile --push .
docker buildx imagetools inspect <yourusername>/tcpdump:v1.0.0
docker buildx rm buildx
Enter fullscreen mode Exit fullscreen mode

The output shows both images with their platforms:



Name:      docker.io/<yourusername>/tcpdump:v1.0.0
MediaType: application/vnd.docker.distribution.manifest.list.v2+json
Digest:    sha256:9dd8cb1d4b77b7d02d41ff8418cd442c01badfe8ecd0c0a3a58f43f528eba378

Manifests:
  Name:      docker.io/<yourusername>/tcpdump:v1.0.0@sha256:0c341c671566dbc3cdded9da05120bb2216142f46516c14cf3d10b6c38997195
  MediaType: application/vnd.docker.distribution.manifest.v2+json
  Platform:  linux/amd64

  Name:      docker.io/<yourusername>/tcpdump:v1.0.0@sha256:c6de3ab95521c9e7e07a05d99935d19686b8d6e81ab85ce631312cffe57d2ce3
  MediaType: application/vnd.docker.distribution.manifest.v2+json
  Platform:  linux/arm64
Enter fullscreen mode Exit fullscreen mode

Attaching an ephemeral container to the Nginx Pod

Now that the tcpdump image is ready, create an ephemeral container called debugger:



kubectl debug --image <yourusername>/tcpdump:v1.0.0 -c debugger $(kubectl get pod -l app=nginx -o name)
Enter fullscreen mode Exit fullscreen mode

Connecting Wireshark to the ephemeral container

Once created, run Wireshark, and connect it to the just-created container.



kubectl exec -c debugger deployments/nginx -- tcpdump -s 0 -n -w - -U -i any | Wireshark -kni -
Enter fullscreen mode Exit fullscreen mode

Request the URL from the POD

As in the local machine, do curl http://example.com in the Nginx Pod.



kubectl exec deployments/nginx -c nginx -- curl http://example.com
Enter fullscreen mode Exit fullscreen mode

Analyzing the Traffic

As expected, there are eight (8) requests to the DNS (A and AAAA) with a negative No such name answer for every pair request until reaching the end of the search list and trying with example.com

Kubernetes DNS Request

Why ndots:5?

ndots:5 generates unnecessary DNS queries. So why five (5) is the default value?

Two comments in the Kubernetes code explain why ndots should be five (5) in Kubernetes:

  1. The tradeoff between functionality and performance
  2. SRV lookup names

The reason ndots is five (5) is to allow SRV record lookups to be relative to the cluster domain.

A typical SRV record has the form _service._protocol.name. and in Kubernetes, the name has the form service.namespace.svc. The formed record will look like _service._protocol.service.namespace.svc. This query contains four dots. If ndots is four (4), it would be considered an FQDN and will fail to resolve. With ndots to five (5), it won't be considered an FQDN and will be searched relative to cluster.local.

Testing the theory with ndots less than five (5).

Let's see in practice how ndots behaves with the DNS service, and how it breaks when is less than five (5).

Restart the nginx deployment to start from scratch:



kubectl rollout restart deployment nginx
Enter fullscreen mode Exit fullscreen mode

Prepare a file to patch the current deployment with the new ndots configuration and apply.



---
spec:
  template:
    spec:
      dnsConfig:
        options:
          - name: ndots
            value: '4'
Enter fullscreen mode Exit fullscreen mode 


kubectl patch deployments.apps nginx --patch-file ndots-patch.yaml
Enter fullscreen mode Exit fullscreen mode

Install nslookup in the pod for testing purposes



kubectl exec deployments/nginx -c nginx -- bash -c "apt-get update && apt-get install dnsutils -y"
Enter fullscreen mode Exit fullscreen mode

Preparing the ephemeral container for ndots testing

Following the same procedure with Wireshark create the ephemeral container called debugger:



kubectl debug --image <yourusername>/tcpdump:v1.0.0 -c debugger $(kubectl get pod -l app=nginx -o name)
Enter fullscreen mode Exit fullscreen mode

Connecting Wireshark for ndots testing

Once created, run Wireshark, and connect it to the just-created container.



kubectl exec -c debugger deployments/nginx -- tcpdump -s 0 -n -w - -U -i any | Wireshark -kni -
Enter fullscreen mode Exit fullscreen mode

Resolve the DNS Service

Run nslookup _dns._udp.kube-dns.kube-system.svc in the Nginx Pod.



kubectl exec deployments/nginx -c nginx -- nslookup _dns._udp.kube-dns.kube-system.svc
Enter fullscreen mode Exit fullscreen mode

In the console, the error is clear:



** server can't find _dns._udp.kube-dns.kube-system.svc: NXDOMAIN
Enter fullscreen mode Exit fullscreen mode

Also in the Wireshark traffic:

k8s-dns-ndots4

The test with ndots:5 passes by changing the patch to"



---
spec:
  template:
    spec:
      dnsConfig:
        options:
          - name: ndots
            value: '5'
Enter fullscreen mode Exit fullscreen mode

Apply the same way:



kubectl patch deployments.apps nginx --patch-file ndots-patch.yaml
Enter fullscreen mode Exit fullscreen mode

Install nslookup in the pod for testing purposes



kubectl exec deployments/nginx -c nginx -- bash -c "apt-get update && apt-get install dnsutils -y"
Enter fullscreen mode Exit fullscreen mode

and repeat the process:



kubectl exec deployments/nginx -c nginx -- nslookup _dns._udp.kube-dns.kube-system.svc
Enter fullscreen mode Exit fullscreen mode

Create the ephemeral container called debugger:



kubectl debug --image <yourusername>/tcpdump:v1.0.0 -c debugger $(kubectl get pod -l app=nginx -o name)
Enter fullscreen mode Exit fullscreen mode

Launch Wireshark:



kubectl exec -c debugger deployments/nginx -- tcpdump -s 0 -n -w - -U -i any | Wireshark -kni -
Enter fullscreen mode Exit fullscreen mode

And run nslookup _dns._udp.kube-dns.kube-system.svc in the Nginx Pod.



kubectl exec deployments/nginx -c nginx -- nslookup _dns._udp.kube-dns.kube-system.svc
Enter fullscreen mode Exit fullscreen mode

The console shows:



Name:   _dns._udp.kube-dns.kube-system.svc.cluster.local
Address: 10.96.0.10
Enter fullscreen mode Exit fullscreen mode

And Wireshark loops properly over DNS search until reaching cluster.local

k8s-dns-ndots5

ndots:5 can negatively affect performance

The default ndots configuration is perfect for default Kubernetes services, but it shouldn't be for later deployed microservices.

Cluster and applications, when connecting with other external components may suffer a negative performance impact and slowness. The DNS can become a bottleneck in case of heavy traffic.

Testing CoreDNS Performance with multiple requests

So far, we have a rich theory but no data. Observability needs both enough data and a theory within which that data can be refined.

Let's create the test inside the Kubernetes Cluster using a small pod. It will launch between 60 and 80 requests per minute. This request range is enough to avoid Cluster overloading and to test the ndots behavior.

Building the application

The python code used is:



#!/usr/bin/env python
# -*- coding: utf-8 -*-

import asyncio
import concurrent.futures
import logging
import sys
import os
import requests

logging.basicConfig(format="%(asctime)s %(levelname)s:%(name)s: %(message)s",
                    datefmt="%d-%b-%y %H:%M:%S", stream=sys.stdout, level=logging.INFO)

logger = logging.getLogger('httprequests')
logging.getLogger("chardet.charsetprober").disabled = True

benchuri = os.environ.get('BENCHURI', 'http://www.example.com')


async def main():

    with concurrent.futures.ThreadPoolExecutor(max_workers=20) as executor:

        fetch = asyncio.get_event_loop()
        futures = [
            fetch.run_in_executor(
                executor,
                requests.get,
                benchuri
            )
            for i in range(20)
        ]
        for request in await asyncio.gather(*futures):
            logger.info(request)

if __name__ == '__main__':

    while True:
        loop = asyncio.get_event_loop()
        loop.run_until_complete(main())
Enter fullscreen mode Exit fullscreen mode

And the Dockerfile:



ARG build_for=linux/amd64,linux/arm64
FROM python:alpine3.10 as base

LABEL maintainer="<yournamehere>"

RUN mkdir /app
COPY dnsrequest.py /app/dnsrequest.py
ADD requirements.txt requirements.txt

RUN pip install --upgrade pip
RUN pip install -r requirements.txt && chmod +x /app/dnsrequest.py && adduser -S none
USER none

WORKDIR /app

ENTRYPOINT ["python", "dnsrequest.py"]
Enter fullscreen mode Exit fullscreen mode

The python file is called dnsrequest.py, and the requirements.txt contains one line:



requests
Enter fullscreen mode Exit fullscreen mode

The process to build the Dockerfile is the same as explained with tcpdump above.



docker buildx create --name buildx --driver-opt network=host --use
docker buildx inspect --bootstrap
docker buildx build -t <yourusername>/dnsbench:v1.0.0 --platform linux/amd64 --platform linux/arm64 --file Dockerfile --push .
docker buildx imagetools inspect <yourusername>/dnsbench:v1.0.0
docker buildx rm buildx
Enter fullscreen mode Exit fullscreen mode

Launching the Application

Minikube under Docker is enough to perform the test. To create the deployment run:



kubectl create deployment dnsbench --image <yourusername>/dnsbench:v1.0.0
Enter fullscreen mode Exit fullscreen mode

Once launched, the CPU from CoreDNS will increase immediately with the default configuration. The DNS log shows all the work performed from a single request.

Note: To activate the CoreDNS Logs, edit the DNS configmap and add the word log in the coredns configuration:



[INFO] 10.244.0.7:33509 - 64525 "AAAA IN www.example.com.default.svc.cluster.local. udp 59 false 512" NXDOMAIN qr,aa,rd 152 0.000022583s
[INFO] 10.244.0.7:33509 - 64400 "A IN www.example.com.default.svc.cluster.local. udp 59 false 512" NXDOMAIN qr,aa,rd 152 0.00005575s
[INFO] 10.244.0.7:35864 - 8087 "AAAA IN www.example.com.svc.cluster.local. udp 51 false 512" NXDOMAIN qr,aa,rd 144 0.000044s
[INFO] 10.244.0.7:35864 - 8004 "A IN www.example.com.svc.cluster.local. udp 51 false 512" NXDOMAIN qr,aa,rd 144 0.000053792s
[INFO] 10.244.0.7:35569 - 21607 "AAAA IN www.example.com.cluster.local. udp 47 false 512" NXDOMAIN qr,aa,rd 140 0.00004025s
[INFO] 10.244.0.7:35569 - 21524 "A IN www.example.com.cluster.local. udp 47 false 512" NXDOMAIN qr,aa,rd 140 0.000065667s
[INFO] 10.244.0.7:53735 - 22419 "AAAA IN www.example.com. udp 33 false 512" NOERROR qr,aa,rd,ra 76 0.000035041s
[INFO] 10.244.0.7:53735 - 22335 "A IN www.example.com. udp 33 false 512" NOERROR qr,aa,rd,ra 64 0.000027542s
Enter fullscreen mode Exit fullscreen mode

By changing the DNS ndots:5 option to ndots:1 with



---
spec:
  template:
    spec:
      dnsConfig:
        options:
          - name: ndots
            value: '1'
Enter fullscreen mode Exit fullscreen mode 


kubectl patch deployments.apps dnsbench --patch-file ndots-patch.yaml
Enter fullscreen mode Exit fullscreen mode

Or changing the request URI from http://www.example.com to http://www.example.com. (Note the dot (.) at the end)

CPU decreases radically. CoreDNS Log shows:



[INFO] 10.244.0.11:54982 - 63488 "A IN www.example.com. udp 33 false 512" NOERROR qr,aa,rd,ra 64 0.000057167s
[INFO] 10.244.0.11:54982 - 63571 "AAAA IN www.example.com. udp 33 false 512" NOERROR qr,aa,rd,ra 76 0.000119333s
Enter fullscreen mode Exit fullscreen mode

Find the CPU load in the CoreDNS CPU diagram below:

CoreDNS-CPU

The application CPU doesn't change, and the requests have the same frequency:

Application-CPU

Regarding network traffic in the CoreDNS, there is also exhaustion when ndots:5:

CoreDNS-net

And in the application:

Application-net

Conclusion

Use specific ndots for your application under spec - dnsConfig. Remember that ndots:1 ignores the search list because the query name satisfies the ndots threshold (At least one dot).

The aggressive ndots:1 forces to use the full domain for every intra-node communication. Using fully qualified names can be described as a "workaround" in different resources. I see it as a proper implementation.

When the application has many DNS requests, ndots:2 increases DNS performance and latency.

Image
20 Reasons To Believe Locks 4 Vans Hook Lock Cannot Be Forgotten
Image
The Ultimate Guide to Car Leasing in Chicago: Flexible Options for Every Driver
Image
이브벳: 승리의 시작, 베팅의 새로운 패러다임!
Image
A Step-By-Step Guide To Choosing The Right Black Fabric Corner Sofa
Image
7 Things About Drip Coffeee You'll Kick Yourself For Not Knowing
Image
Proxmox: Resize Your Local (pve) Disk

proxmox, devops, beginners, homelab
Image
Understanding the Differences Between Authentication and Authorization

security, sauthentication, authorization
Image
keuntungan main di dalam agen slots online mempunyai pengalaman
Image
Reducing Delivery Times and Costs: How Machine Learning Optimizes Delivery Routes Efficiently

machinelearning, commerce, bigdata
Image
Fitting a Cat Flap in a UPVC Door
Image
THE FUTURE OF AI IN DIGITAL MARKETING

webdev, ai, career
Image
Przestrzeń Marzeń: Jak Stworzyć Idealne Mieszkanie
Image
Enhancing Your Home's Outdoor Appeal with a Professional Patio Contractor in Fayetteville
Image
Installing a Cat Flap in a UPVC Door Cost
Image
Why Filtered Coffee Makers Is The Next Big Obsession
Image
BUILDING A CYBERSECURITY DETECTION AND MONITORING LAB BY LEVERAGING LOCAL VIRTUAL MACHINES (VMs) AND MICROSOFT AZURE

security, networking, cybersecurity, homelab
Image
Understanding the Differences Between @Controller and @RestController Annotations

spring, controller, api
Image
Power BI Conditional Formatting: Enhancing Your Visuals with Conditional Formatting

powerbi
Image
The Biggest Problem With Repairing Bifold Doors And How To Fix It
. . . . . . . . . .