Double-Bypass Virtualization for HPC vNIC Access
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional high performance computing (HPC) environments face challenges in multi-tenant cloud environments due to resource competition, overhead, and performance loss associated with conventional virtualization techniques, making them impractical for HPC applications.
Innovation Solution
Double-bypass virtualization techniques that allow HPC user space applications to directly access virtual network interface cards (vNICs), bypassing the operating system and hypervisor, thereby reducing latency and increasing throughput by offloading network operations to hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional virtualization techniques are used in HPC environments, then resource sharing and multi-tenant capability are improved, but performance and throughput deteriorate due to OS kernel and hypervisor overhead
Solution Approach 1:
The patent extracts and removes the OS kernel and hypervisor layers from the HPC application execution path. By using containerization technology, HPC applications are isolated in user space containers that directly access underlying virtualization infrastructure, eliminating the need for full OS kernels and hypervisors. This extraction resolves the contradiction by maintaining multi-tenant isolation through containers while removing the performance overhead of traditional virtualization layers.
Solution Approach 2:
The patent introduces a container runtime and resource manager as an intermediary between HPC applications and the physical infrastructure. This intermediary provides the necessary isolation and resource management functions traditionally performed by OS kernels and hypervisors, but with minimal overhead. The container runtime enables multi-tenant capability while maintaining high performance by acting as a lightweight mediator rather than a full virtualization stack.
2Ease of operation
If batch scheduling with large time-slices is used, then resource allocation simplicity is improved, but scheduling flexibility and responsiveness deteriorate
Solution Approach 1:
The patent implements dynamic scheduling by allowing HPC applications to be scheduled and executed as isolated containers with fine-grained resource control. Instead of fixed batch time-slices, the system can dynamically allocate and revoke resources based on application needs and priority. Containers can be started, stopped, and migrated independently, providing both simplicity through automated management and flexibility through granular control.
Solution Approach 2:
The patent segments the HPC workload into independent container instances that can be scheduled individually. Each container represents a discrete, manageable unit that can be allocated resources independently. This segmentation enables simple resource allocation at the container level while maintaining flexibility to schedule multiple containers with different priorities and resource requirements simultaneously.
3Productivity
If multiple HPC applications share the same OS kernel, then resource utilization is improved, but system stability and isolation deteriorate
Solution Approach 1:
The patent segments each HPC application into its own container instance with isolated file systems, process spaces, and resource allocations. While multiple containers share the underlying physical infrastructure, each container maintains strong isolation boundaries through containerization mechanisms. This enables high resource utilization through sharing while maintaining reliability through isolation, as failures or security issues in one container cannot affect others.
Solution Approach 2:
The container runtime acts as an intermediary layer that provides isolation between multiple HPC applications sharing the same infrastructure. It manages resource allocation, enforces security boundaries, and handles inter-container communication. This intermediary enables multiple applications to share resources efficiently while maintaining the isolation and stability needed for reliable HPC workloads.
Data Source
AI summary
Double-bypass virtualization techniques are provided to enable use of high performance computing (HPC) application in a virtualized environment. In one example, a HPC user space application running on a virtual machine obtains direct access to virtual network interface card (vNIC) on a host-computing device associated to that virtual machine. The HPC user space application is configured to transmit and/or receive one or more data frames via the vNIC while bypassing the operating system of the virtual machine and the virtual machine hypervisor of the host-computing device.