Build a Hybrid Application beyond x86

Ahmad Sayed Hassan
14 min readJan 26, 2021


This article audience mainly Cloud-native engineer and architect with knowledge of containers and Kubernetes. The following section is intended to brief ISA and multiarch. Please skip ahead to section “ Options to build multiarch container images” for the technical part.

Why do we need to consider Multiarch hybrid applications?

The simple answer let us put everything on x86 and go with it. The answer to that needs a look at past, present, and the future.


Enterprises and Large system used to run on Mainframe, a large computer with a large amount of memory, but with increasing Personal Computer (a.k.a) Microcomputer, more CPU started to Mass produced to fit inside a box that lay on the Desktop.

The Microcomputer CPUs started to make their way to Enterprise applications. With the introduction of the distributed system, the x86 machines began to dominate the Enterprise applications with distributed systems’ success. Stewart Alsop II famously predicted that the last Mainframe would be unplugged on March 15, 1996. This prediction turns out to be very wrong with IBM introducing IBM Z15 in September 2020. IBM Z and Mainframe turns out to be very much alive and striving. Running a Legacy application very reliable, and the Mainframe is packed with considerable computing power in a small footprint compared to distributed x86_64 systems.


Public Cloud and Hyperscaler become the new normal for hosting Enterprise application, with apparent domination of x86_64 ISA from Intel Xeon and AMD EPYC. Still, the demands for specialized workloads and hosting from businesses make the public cloud provider introduce different Machines such as IBM power, IBM Z, and Amazon ARM.


Nowadays, we are in the Era of 5G and IoT, with increased Edge computing demands derive the demand for Power-efficient machines, also the success of Heterogeneous computing introduced initially in Mobile devices, ARM machines start to get back to the PC market and move toward Servers and Enterprise Applications.

The previous introduction is trying to demonstrate why there is a need for multiarch deployments; if you are building an application, asset, or even custom enterprise solution, you may still face a demand to deploy it on none x86 machines, so what are the other options we have now? Before answering that, just a quick detour around what is the ISA?

Introducing Instruction set Architecture (ISA)

As per

The instruction set, also called ISA (instruction set architecture), is part of a computer that pertains to programming, which is more or less machine language. The instruction set provides commands to the processor to tell it what it needs to do. The instruction set consists of addressing modes, instructions, native data types, registers, memory architecture, interrupt and exception handling, and external I/O.

In early 1960, IBM used to have a series of incompatible families, but with the introduction of IBM System 360, the grandparent of IBM Z/Architecture, IBM standardized their Mainframe. IBM System 360 is The first true ISA designed as a portable hardware-software interface.

ISA landscape today

Will list some of the CPU, there is many other such as SPARC, RISC, MIPS, .. etc. But here, I will list the machines that are easily accessible via owning a device or provided as a self-service on the public Cloud.

Note Definition in Linux on what comes when you execute “uname -m” for those ISA that is used within this article. There are other names for a different generation, not the scope to cover them here.


IBM Z is the latest generation of the System/360 introduced in 1964, currently used in IBM Z Mainframe. Still, you do not need to purchase an expensive mainframe to try. IBM Cloud introduces Hyper Protect Virtual Server, which gives access to IBM Z Machines running Linux One.

  • Strengths: Handle Intensive IO operations, such as Credit card swipes.
  • Definition in Linux: s390x
  • Trivia: In z/architecture, the Z stands to Zero downtime.


Developed by Intel and currently produced mainly by Intel and AMD, it provides a large family of processors covering a broad spectrum of mobility, laptop, workstation, and High-Performance Servers.

  • Strengths: General purpose computing, the defacto standard for all software.
  • Definition in Linux: x86_64, amd64
  • Trivia: Even Intel is the one who created x86 and licensed to other companies such as AMD and Cyrix, AMD is the one who defines x86_64, the 64-bit extension, after the limited success of IA_64 in the consumer, and intel ended up license AMD64 from AMD.

Power ISA

Developed by the AIM alliance, AIM refers to Apple, IBM, and Motorolla, to create RISC machines. PowerPC used to be the standard CPU in Apple products till 2005. PowerPC was the CPU used in many consumer products such as Playstation 3, xbox360, and Nintendo Wii. It is part of OpenPower Foundation and is introduced in IBM pSeries servers. Also, it is available on IBM public cloud as PAYGO self-services virtual servers.[]

  • Strengths: with IBM power 10, IBM claims its main advantage is to provide faster AI inference.
  • Definition in Linux: ppc64el
  • Trivia: Power2 was the CPU used in Deepblue, the first machine to win chess, Kasparov. Also, Power7 was the CPU behind the machine that won the first Jeopardy


Developed by Acorn computer, to be used in the BBC Microcomputer produced by Acorn Computers, ARM engineer designed the CPU to be power-efficient. The goal back then is to replace the expensive ceramic package and heatsink with a cheaper plastic package. To get your hands on a machine running ARM, hold your phone. or get a Raspberry Pi. Also, Amazon EC2 based ARM Graviton. []

  • Strengths: Power efficiency.
  • Definition in Linux: aarch64
  • Trivia: according to Steve Furber, the principal designer on ARM, when they received the first chip, they tested the power consumption and found it Zero, he reviewed the wiring, thinking that the meter is not connected correctly it turned out they did not connect the CPU to the power supply at all, because of the faulty board, and it is running only on the power coming from the input signal.

Options to build multiarch container images

Now it comes to what I should do to make my containerized application run on multiarch. Here will cover four different approaches.

  • Do nothing; It will work
  • Build everything on its dedicated machine
  • Cross Build and Cross-compiling
  • Userspace emulation

Options 1: Do nothing it will work

This option is the most forward way to think; currently, most of the application is written in Architecture agnostic language, Java for example come with slogan compile once run anywhere, also Python, Node.JS, all supported and ported to previous ISA and more. Bringing this to the container world should as simple as

podman run -it python python -c 'print("Hello world!")'

and this command will run correctly on mostly all commercial Machines, but this only possible because the python image on itself is multiarch,

when running it, podman will pull the proper image that matches your arch, and as soon as you build a custom image as simple as

Create a docker file just pulling Python Image

FROM python

Build and push it

buildah build-using-dockerfile -t ahmadhassan83/simplypy86 
podman push

Then try to run the same command on an ARM Machine.

podman run ahmadhassan83/simplypy86 python -c 'print("Hello World")'

Will it run or Not ???!!!

The answer is NO and gives the following error.

standard_init_linux.go:219: exec user process caused: exec format error

It will no longer work on any other different architecture because it is a custom image, pulled and created by default according to the computer you use.

So with this option, we did not do any actual building for container images; even the most straightforward image failed to run.

Options 2: Build everything on its dedicated machine

This option is a no brainer option, pick the proper machine; maybe it sounds like an expensive idea, but still, with Cloud offerings, as elaborated in the previous section, you can rent a machine at an hourly or monthly rate for the most commonly used CPU such as s390x, aarch64, ppc64el, and x86.

Still, picking a dedicated machine for the build may add operational complexity, another option to run on emulation on any x86 machine using the qemu system. for those who do not know, Qemu can emulate most of the CPU in the market. Unfortunately, this option may sound attractive, still emulating a complete machine take resources and even require an additional operation. Running an entire system in emulation will decrease the performance significantly.

To give a sense of emulation’s performance impact, find below some numbers I got from benchmarking s390x a machine on emulation vs. the machine running on the Cloud; in this test, I ran ubuntu 18.04 on s390x and benchmarked the time needed to executed different tasks.

Notice that Qemu, which is faster than Hercules still on average ~ 10 to 50 times slower than the machine provided on Cloud.

Hercules s390x emulation

CPU: SHA256-hashing 500 MB
284.661 seconds
CPU: bzip2-compressing 500 MB
271.677 seconds
CPU: AES-encrypting 500 MB
19.429 seconds

Qemu s390x emulation

CPU: SHA256-hashing 500 MB
18.163 seconds
CPU: bzip2-compressing 500 MB
66.803 seconds
CPU: AES-encrypting 500 MB
38.001 seconds

IBM Z Virtual machine on KVM

CPU: SHA256-hashing 500 MB
2.532 seconds
CPU: bzip2-compressing 500 MB
6.933 seconds
CPU: AES-encrypting 500 MB
0.351 seconds

You can bring a workable Linux machine on an emulator running on decent x86. Still, to have a more complex distributed system with compression and encryption, the system with reasonable timeouts will be a quiet challenge.

Option 3: Cross Build and Cross-compiling

In this option, we tell buildah to override the architecture of the image with the expected target architecture, which means buildah; please build me an image but bring the binaries for this specific architecture so in that case the

The Dockerfile will be the same.

FROM python

The build command will explicitly mention the target architecture.

buildah build-using-dockerfile --override-arch arm64  -t ahmadhassan83/simplypyarm  .

try running this Docker in the same x86 machine that builds it will lead to

standard_init_linux.go:219: exec user process caused: exec format error

Meanwhile, pulling and running this container in the Arm machine will run successfully.

for a compiled language such as c,c++, and golang, use the GCC cross compiler

for this sample will write a simple c code

#include <stdio.h>int main() {
printf("Hello World!\n");
return 0;

cross-compile it for aarch64 (ARM 64)

aarch64-linux-gnu-gcc test.c -o test.aarch64

Note: you need to install a cross compiler on your x86; in Ubuntu Linux, run this command

sudo apt-get install gcc-aarch64-linux-gnu g++-aarch64-linux-gnu

Note: .aarch64 at the end of the filename does not have any significant meaning to ease dealing with multiple binaries

Then create a docker file that copies the compiled binary.

FROM ubuntu
COPY test.aarch64 /usr/bin/test
CMD ["test"]

Then run buildah and podman push.

buildah build-using-dockerfile --override-arch arm64  -t ahmadhassan83/simple  .
podman push ahmadhassan83/simplec

As expected testing, this Docker on the build machine x86 will not work, but testing it on the ARM64 machine will work without issues.

This approach sounds it solve it, even if when using compiled language, cross compiler came to the rescue, but this applies to over-simplistic Docker operation, to add one level of complexity such as executing a command such as chmod, or install an additional package using yum or apt, it will end up the failure of building the container image.

For example, if the Dockerfile will look like this.

FROM ubuntu
COPY test.aarch64 /usr/bin/test
RUN chmod 777 /usr/bin/test
CMD ["test"]

we already cross-compiled our C code, but we need to chmod the binary; if we try buildah with target arm64 on x86 build machine, here is what will happen

ahmed@ubuntu:~/simple$ buildah build-using-dockerfile --override-arch arm64  -t ahmadhassan83/simplec  .
STEP 1: FROM ubuntu
Getting image source signatures
Copying blob a970164f39c1 skipped: already exists
Copying blob e9c66f1fb5a2 skipped: already exists
Copying blob 94362ba2c285 [--------------------------------------] 0.0b / 0.0b
Copying config 1c28a15891 done
Writing manifest to image destination
Storing signatures
STEP 2: COPY test.aarch64 /usr/bin/test
STEP 3: RUN chmod 777 /usr/bin/test
standard_init_linux.go:211: exec user process caused "exec format error"
error building at STEP "RUN chmod 777 /usr/bin/test": error while running runtime: exit status 1
ERRO exit status 1

Again exec format error, because buildah uses runc to run a binary within the image, this binary is aarch46 binary, so it is not executable, to make this buildah run on x86 build machine we get back that to option 2, or emulators or Proceed with Option 4

Option 4: User space emulation

Before proceeding with this option, we need to talk about userspace emulation. In option 2, we mentioned QEMU as a system emulator, but QEMU has another mode of operation which user emulation. This mode runs for a specific OS, and qemu supports two hosts, qemu-linux-user and qemu-bsd-user. With this mode, Qemu emulates only the CPU to translate the foreign machine code, fix the syscalls parameters for endianness, and then send it to the native host kernel. This mode may be similar to the concept of the virtual machine in Java.

It is still emulation but relatively faster, and no need to build, maintain, and operate a complete emulated system; this setup will also help build the container image and run and test on your x86 build machine to execute unit testing or smoke testing.

to enable this mode will run a specific container (

sudo podman run --rm --privileged multiarch/qemu-user-static --reset -p yes

With this command, all previously failed commands will run seamlessly. This container utilizes two tools, qemu-linux-user, and another Linux capability called binfmt_misc. binfmt_misc, according to Wikipedia, a capability of the Linux kernel which allows arbitrary executable file formats to be recognized and passed to certain userspace applications, such as emulators and virtual machines. It is a binary format handler in the kernel involved in preparing a userspace program to run. In our case, the emulator will qemu-linux-user.

This option has an additional advantage when it comes to compiled language, no need even to cross-compile compilation can be part of the container image itself or spawn a dedicated container for compiling. Why this important, compiling binaries during the linker for cross-compiling, you need to have all binaries and dependencies, having a container image with the compatible OS during the compiling will significantly simplify the build process.

Also, as a bonus point no need to install multiple cross compilers and deal with environment variables. The build file will be the same for all targets like it runs on the native machine.

Testing Environment.

Now it is time to put all of those theories and a simple couple of liner hello world to more realistic test, will build a microservices that expose a rest API, and this rest API query a Mysql DB get information.

These microservices will be deployed on Kubernetes and connected to Mysql DB in the same environment.

We are trying to build a mutliarch version of these microservices and deploy it in both x86 and aarch46, enabling replication between MySQL installed on aarch46 primary and x86 Slave.

The code for this rest API will be available in this github Also, the Dockerfile is used to build and create this Restful APU Microservices.

FROM ubuntu
RUN apt update
RUN apt -y install default-libmysqlclient-dev libulfius-dev uwsc gcc
COPY rest.c /root
RUN gcc rest.c -o rest -L/usr/lib/`uname -m`-linux-gnu/ -lulfius `mysql_config --cflags --libs`
ENTRYPOINT ["/root/rest"]

I wrote it in C. For this code, I will use option four and will compile the code within the buildah. For simplicity will run it in the same container image.

  • Kubernetes Cluster on ARM64 using VMware ESXi as Hypervisor
  • Mysql on ARM64 on Arm64 Virtual Machines
  • Single node Kubernetes Cluster and Mysql on x86 VM

Now we are on the fork of choosing two options to continue with buildah or move to another tool called buildx so that we will start with buildah

For Buildah to build a mutliarch container image, we need to do the following steps

  • Build the container image for AMD64 and ARM64
  • Push the container image to Docker Hub.
  • Create a manifest
  • Push the manifest

I can do those steps with the following list of commands.

buildah bud --override-arch amd64 --os linux/amd64 -t ahmadhassan83/rest-arm64 .
buildah bud --override-arch amd64 --os linux/amd64 -t ahmadhassan83/rest-amd64 .
podman push ahmadhassan83/rest-arm64
podman push ahmadhassan83/rest-amd64
podman manifest create offeringapp
podman manifest add --arch arm64 offeringapp ahmadhassan83/rest-arm64
podman manifest add --arch amd64 offeringapp ahmadhassan83/rest-amd64
podman manifest push --all offeringapp docker://

Those steps seem to work fine, but for some reason, it gives an error when trying to run it on k3s a Kubernetes with containerd

archive/tar: invalid tar header: unknown

I did not do enough investigation, but it looks like it is some issue in buildah (

Moving forward, we got the idea, from now will use a new tool provided by Docker called build. This tool makes all previous steps, but with Docker, and what is interesting about this tool no need to manipulate the manifest or build images and add them manually.

To run this tool to install Docker and build plugins, proceed with qemu-user-static and then run the initial one-time setup.

docker run --rm --privileged multiarch/qemu-user-static --reset -p yes
docker buildx create --name mybuilder
docker buildx use mybuilder
docker buildx inspect --bootstrap

Then for the build, all that is needed is one line.

docker buildx build --platform linux/arm64,linux/amd64   --tag ahmadhassan83/myapprest . --push

Handling Databases

This use case will use Mysql as databases, the setup for master-slave configuration is taken straight ahead from this tutorial.

The only modification we have here is that one DB instance runs on ARM64 and the second DB running is Slave running on x86. Having two DB’s running on different architecture will introduce another use case for AI and Analytics in CQRS[], architecture pattern

In this Pattern, the DB replicated to read-only replica for other use cases, noticing the different advantages provided by different architectures, for example, replicating the Data to Z machine for intensive IO operation, or to P Machine for AI and analytics.

Automated Pipeline using Github actions.

Last but not least, many tools help with doing build automation. The scope of this article will go with GitHub’s actions. It provides integration with container registry such as dockerub, GitHub container registry.

name: ci


runs-on: ubuntu-latest
name: Checkout
uses: actions/checkout@v2
name: Set up QEMU
uses: docker/setup-qemu-action@v1
name: Set up Docker Buildx
id: buildx
uses: docker/setup-buildx-action@v1
name: Available platforms
run: echo ${{ steps.buildx.outputs.platforms }}
name: Login to DockerHub
uses: docker/login-action@v1
username: ${{ secrets.DOCKERHUB_USERNAME }}
password: ${{ secrets.DOCKERHUB_TOKEN }}
name: Build with buildx
run: |
docker buildx build --platform linux/arm64,linux/amd64 --tag ahmadhassan83/myapprest . --push

As soon as you push the code, GitHub workflow triggered a multiarch build.


The goal of this is to highlight that, Hybrid and multi-Cloud patterns are not only about utilizing multiple cloud providers or putting your workload on your data center and burst to the Cloud. Those patterns can extend it to different CPU architecture like each cloud provider has its advantage, each Machine and ISA has its advantage.

Machines such as Z, Power, and ARM even accessible as PAYGO and self-services model, you can pick and try from a public cloud provider, same as x86 machines.

My current setup for ARM64

Raspberry PI 4 running VMware ESXi.

VMware ESXi Cluster

Using Lens IDE to Manage and operate two Kubernetes Clusters on x86 and ARM64