On-Premises Private Cloud Architecture - HyperCloud

Where do you even begin building a unified cloud technology?

Given the multidimensional nature of the challenge, it’s a big question. At SoftIron, four primary design goals informed virtually every decision we made about HyperCloud—our very own on-premises private cloud.

Simplicity

Make operators’ experiences as simple as possible by abstracting away as much cloud computing, storage, networking, and compute complexity as possible, as elegantly as possible.

Unification

Ensure every part of the stack works seamlessly together and can be supported by a single cloud provider. Doing this in the design stage makes for a superior support experience.

Scalability

Support private cloud deployment from less than a rack to hundreds of racks with minimal configuration burdens, enabling operators to comfortably provide SLAs to tenants and stakeholders without worrying about resilience in cloud infrastructure.

Freedom

As much as is technically possible, prioritize user freedoms for easily moving workloads in a hybrid cloud environment, utilizing cloud services wherever it makes the most sense for them to reside.

Starting from the ground up

In cloud environments, partial system failures in hardware tend to precipitate the ugliest and most unexpected issues. Faulty error correction, memory corruption, CPU failure, and kernel panics can all lead to serious problems that are difficult to debug. Not to mention that when you’re dealing with the intricacies of all the different generic technologies in composed private cloud environments and managed services, ongoing management is already a significant challenge.

The task-specific hardware options in HyperCloud are designed for consistency and to work as one regardless of network, storage, or compute function. This approach not only eliminates compatibility issues and the associated snowballing management challenges in private cloud environments, but also simplifies management of the technology “fleet” and opens unprecedented possibilities for improving cloud operations, cloud infrastructure, and private cloud computing. And every layer of HyperCloud technology is designed and built accordingly.

Multi-tenancy

HyperCloud is service-provider ready, which means it supports strict tenant separation and granular access controls and permissions. With advanced cloud security, owners can take advantage of:

User and Group ACLs: Granular access control lists for users and groups for every resource available across virtualization, containerization, storage, virtual private cloud, virtual networking, and marketplaces.

Accounting toolset: Consistent hypervisor reporting of CPU, memory, storage, and network bandwidth usage per user, which can be linked to cost and billing systems for better cloud resource management.

Showback: Built-in showback systems that enable chargeback and billing for users, groups, and tenants within a private cloud environment.

Resource quotas: Configurable quotas for all resource types, including rate limiting, to restrict user overconsumption of specific constrained cloud resources.

Host overcommitment: As with most modern cloud platforms, you can configure the compute layer to overcommit CPU and memory resources according to a user’s workload.

Marketplace

While users can manually create and upload or import guest instances from other clouds, HyperCloud also supports the concept of ‘cloud apps’ and a marketplace for guest images. Teams can build and deploy virtual machines and containers from:

HyperCloud’s public marketplaces
Docker Hub
Linux Containers
Owner-created and hosted marketplaces

Users can also snapshot live virtual machines and add them to existing marketplaces where they can then be used by tenants.

Workloads

The virtualization layer in HyperCloud enables the deployment of:

Traditional virtual machines
LXC containers
Multi-VM or multi-container services

All the above live within the context of a strictly tenanted environment, with the option to build underlying consumable storage and templates that can optionally be shared across tenants as operator-provided functionality. Templates enable the configuration of capacity, networking, persistent storage, and granular instance configuration (boot order, memory, SSH-keys), with full support for cloud-init. We design all of our CPU architectures with feature parity so you can mix and match various CPU architectures in a single cluster, providing optimal cloud service and private cloud infrastructure. Our managed private cloud solutions integrate seamlessly with data security, offering operative and successful cloud monitoring.

Storage

In a HyperCloud environment, the cloud itself and all the layers above it can leverage the distributed storage layer, which provides three main services:

Designed to be highly flexible, the storage layer enables teams to build and deliver storage across multiple performance and cost tiers. While under the hood, it employs several industry-standard storage tenets:

Compute layer storage: Image and snapshot storage for the compute layer and marketplace.

Object storage: A featureful S3-compliant object storage API consumable by apps.

Guest layer block storage: Persistent storage for virtual machines and containers.

Calculated placement: Deterministic data placing using hashing algorithms that enable compute nodes and guests to be cluster-aware and read and write directly to storage nodes boost performance and scalability.

Journaling and caching: In many cases, storage nodes have multiple differing media types to help accelerate performance for tricky workloads and use technology such as on-flash hybrid volume caching and in-kernel caching. We can’t break the laws of physics but we do our best to get as close as possible.

Efficient cloning and snapshotting: All volumes and snapshots are copy-on-write clones of their base image, ensuring that private cloud storage across tenants is used extremely efficiently. Only deltas are written and regular guest snapshots cost next to nothing. Our cloud platform supports public cloud infrastructure and integrates seamlessly with private cloud services.

Networking

From configuration challenges to performance issues, L2 and L3 switching and architecture are well-known sources of pain. HyperCloud relies on a simple and easy-to-support physical/virtual appliance that delivers high-speed 10/25/100G Ethernet networking and advantages on multiple levels.

Scalable interconnect: Full mesh where possible and leaf-spine thereafter, the HyperCloud interconnect is resilient and automated.

Intelligent data center operations: The control plane streamlines life cycle activities like deployment and maintenance by automating low-level networking configuration, firmware deployment, and OS installation for all resource nodes, helping to simplify DCOps and break/fix.

Automated endpoint configuration: Stateless bring-up automatically pulls the working network configuration from the control plane and provisions new nodes accordingly. Host networking is as simple as plugging in a node and turning it on.

Virtual networking: High-speed networking is piped all the way through to guest instances, enabling VMs, containers, and multi-VM services to use as much bandwidth as an operator will allow the tenant. Tenant security groups enable granular firewall management.

Hardware

The flexible building-block-style hardware range includes networking, storage, and compute-focused appliances. Everything runs on a hardened custom OS based on GNU/Linux that is minimized to include only the components required to power host machines. A comprehensive hardware management platform simplifies firmware and OS deployment across the range and underpins:

Standardized delivery: A single image powers interconnect, compute, and storage nodes.

Stateless provisioning: Bare-metal provisioning is entirely stateless, making dealing with failed nodes or upgrading machines as simple as running ‘reboot.’

AI/ML

HyperCloud is designed to handle the most demanding AI and machine learning workloads. With built-in support for GPU acceleration, HyperCloud enables the deployment of AI/ML models at scale, offering high-performance compute resources to support intensive training and inference processes. HyperCloud’s flexible architecture allows for:

GPU Passthrough & Virtualization: Maximize AI/ML performance with direct access to GPUs or by sharing GPU resources across multiple VMs or containers.
Seamless Integration with AI Frameworks: Deploy your AI/ML applications in containers or VMs using popular frameworks like TensorFlow, PyTorch, and more.
Scalable Infrastructure: Easily scale up or down to meet the fluctuating needs of AI/ML workloads without compromising performance or availability.
Multi-Architecture Support: Mix and match CPU and GPU nodes across architectures in a single cluster, ensuring optimal performance for AI tasks while managing resource efficiency.
Cloud-Native Efficiency: With HyperCloud’s cloud-native infrastructure, manage large datasets, conduct high-speed data processing, and enable multi-tenant environments where AI/ML models can be deployed securely.
Discrete Resource Allocation: HyperCloud ensures that resources like GPUs, RAM, and storage are used exclusively for their designated workloads. GPU power can be dedicated solely to AI/ML tasks, while other compute and storage resources remain isolated, optimizing performance and avoiding resource contention across the cluster.

Whether you’re working on data-heavy AI models or developing cutting-edge ML applications, HyperCloud ensures optimal performance, resource efficiency, and ease of management.