Example PhD

Accelerated Clouds – Application Acceleration As A Service (A4S)

Supervisor: Professor N.J. Avis

Keywords: Cloud Computing, Application Acceleration, GPUs, Virtualization

Cloud Computing is an emerging computing paradigm which can take many guises such as Infrastructure As A Service (IAAS) and Software As A Service (SAAS) which presents different levels of virtualization to the end users. Present cloud computing systems (including commercial offerings such as Amazon's ECS and S3) are built using traditional homogeneous commodity clusters with limited internode communications infrastructures on top of which virtualization software is run. These attributes allow very large physical infrastructures to be created at reasonable costs compared to more exotic supercomputing architectures, but limit the scope of present Cloud computing infrastructures to "high throughput computing" applications. Recently there has been interest in moving towards a heterogeneous physical layer by deploying GPGPU accelerator devices within specialised cloud "super nodes". These super nodes would allow applications to be accelerated by making use of a dense array of computational cores which can deliver large numbers of computational operations (c. 0.5TFlops/s) at modest cost. Whilst not offering a generic HPC Cloud solution, such a system would allow very efficient intra-node parallelisation and application acceleration to those codes which exhibit data flow fine grained parallelism contained within a single "super node".

However, major obstacles to this approach centre around how to virtualise and expose the "super node" containing the accelerator system and how to ensure user applications can make use of these accelerator devices. In the latter case this often results in the hand porting of the code using device specific APIs and language constructs, which is a time consuming and expensive task, with no absolute guarantee of significant performance increases when this task is commenced.

We have already developed an intelligent application porting system (IAPS) for application accelerators which overcomes many of the latter barriers. This project will explore how such an intelligent, automatic code porting system can be extended to allow users to easily use Cloud systems that utilise heterogeneous devices, such as GPGPUs and other accelerators. The outcome will be an "Application Acceleration As A Service" (A4S) prototype cloud. The research programme will create a prototype A4S cloud and characterise its performance when running typical scientific codes.

Key Skills/Background: Open to Computing Graduates and Postgraduates

Contact: Professor N.J. Avis to discuss this research topic.