Qrackmin - a minimalistic container for Qrack - OpenCL & CUDA

Qrackmin container system deployed in rancher through the ThereminQ HELM definitions

Docker deployments

• The :latest container image is meant to be used on a single node with Nvidia-Docker2 and Linux support

docker run --gpus all --device=/dev/dri:/dev/dri --privileged -d twobombs/qrackmin[:tag] [--memory 24G --memory-swap 250G]

• -v /var/log/qrack:/var/log/qrack for saving of measured results outside container

docker exec -ti [containerID] bash

• ThereminQ repo with runfiles is checked out on /root

• Windows users should install WSL2, Docker Desktop, docker.io, nvidia-docker2 to run this (CUDA only)

---

• Qrackmin:AWS - Qrackmin:BRAKET

  on demand AWS template proposals for x86 and ARM - CUDA, OpenCL and CPU powered

• :AWS boilerplate binary runtime code for Qrack as a Service - QFT RND benchmarks output

• :BRAKET boilerplate python runtime code for PyQrack as a |BraKET> container service

---

• Qrackmin:pyqrack - a python enviroment to run tests for pyqrack

• Qrackmin:qbdd - a python enviroment to run benchmarks and tests for qbdd

---

• Qrackmin:POCL - latest & POCL added for CPU-only support

:pocl container image adds the generic OpenCL-ICD and is to be used with high memory & CPU count hosts

• Simulate performance and measured results on CPU
• For validation before GPU cluster deployment

---

• Qrackmin:VCL - Qrack@VCL with node support

The :vcl tag contains VCL binaries that are copyrighted by Amnon Barak to run VCL as a backend and host

On the host the following directory structure needs to be created

sudo mkdir /var/log/vcl /var/log/vcl/etc /var/log/vcl/etc/vcl/ /var/log/vcl/etc/init.d /var/log/vcl/usr /var/log/vcl/usr/bin /var/log/vcl/etc/rc0.d /var/log/vcl/etc/rc1.d /var/log/vcl/etc/rc2.d /var/log/vcl/etc/rc3.d /var/log/vcl/etc/rc4.d /var/log/vcl/etc/rc5.d /var/log/vcl/etc/rc6.d

Run the bash in /run-scripts/ in this repository called ./1-run-nodes

You will be asked two questions:

• the amount of virtual nodes you want to create
• the NVIDIA devices you want to expose ( often 'all' will suffice, otherwise use the device number )*

When you've deployed enough backend containers to you liking you can start ./2-run-host

• the nodes' IPs will be scraped
• the host container will be started and will initialize
• you'll drop into the host-containers' bash
• then run workloads through ./vcl-1.25/vclrun [command]

(*) = other OpenCL device types such as an IntelIGP that are also recognised will also be taken into the cluster

---

A Full VDI host experience is avaliable in ThereminQ:vcl-controller

Follow the guide for running VCL in a cluster here

For multi-host setup please look Docker Swarm and Zerotier

---

Credits go to Dan Strano for creating Qrack and Qimcifa

Some rights are reserved regarding code and functionality for the Amazon AWS |BraKET> container images; they are in the git repo checkouts for the :aws and :braket container images and the aws tools for Amazon AWS as well