Lead Image © orson, 123RF.com

A Container for HPC

Fusion

Article from ADMIN 34/2016

By Jeff Layton

Working with containers in high-performance computing can be difficult; a new container named Singularity comes to the rescue.

Containers are an important part of the IT industry because (1) they are more efficient than full (hardware-level) virtualization and (2) the container workflow readily supports development and operations (DevOps). In high-performance computing (HPC), the priority is application portability (mobility of compute), so researchers can run their workflows anywhere. Although some efforts have been made to use Docker in this context, it lacks some features to operate in an HPC world and has other features that complicate a deployment. Singularity, a container for HPC, uses namespaces to achieve operating system virtualization.

In full virtualization, a system has a host operating system (OS) that runs something called a hypervisor (Figure 1, right side). The hypervisor is capable of communicating directly with the host kernel, which can be very useful for some purposes, such as Peripheral Component Interconnect Special Interest Group (PCI-SIG) single root I/O virtualization (SR-IOV) [1]. Above the hypervisor sits the virtual hardware layer, which sometimes has a bypass to the physical layer (if desired), and above that is the kernel for the OS installed on the virtual machine (VM). Next are the root filesystem and applications that run in the VM. Notice that the applications running on the host are much closer to the physical host hardware (Figure 1, left) than the applications running within the virtualized context. This generally results in a performance penalty when running applications within a virtualized stack.

...

Use Express-Checkout link below to read the full article (PDF).