A team of Croatian researchers published a paper in the January 2013 edition of Scientific-professional Journal of Technical Faculties of University in Osijek examining the present state of high-performance computing in the cloud with an emphasis on currently-available solutions.
The authors provide a review of the key benefits of cloud: namely better resource utilization, efficient charge back mechanisms, on-demand provisioning and dynamic reallocation, and the potential to resolve peak workload demands (aka bursting).
In spite of these many advantages, however, HPC cloud still faces an uphill battle as it seeks to address technical and cultural barriers. Accordingly, the majority of HPC solutions continue to be the traditional, earth-bound variety. The biggest challenge is I/O-related. The slow network speeds and commodity interconnects that characterize most clouds present significant bottlenecks to data-intensive applications. But there are additional contributing factors underlying the slow rate of adoption, the authors observe. Virtualization, for example, adds a layer of complexity that is an anathema to the most latency-sensitive HPC workloads, while a growing reliance on co-processors presents its own set of challenges for HPC scaling.
HPC vendors are in the interesting position of being able to either support or retard the HPC cloud model. When it comes to developing cloud-friendly software models, the major ISVs have dragged their feet, not wanting to risk the cannibalization of proven income streams. But there are signs of momentum with most vendors now offering some level of cloud offering.
The researchers make the case that HPC vendors have indeed begun offering “fully functional HPC cloud solution[s].” In light of this, they recommend a set of helpful questions for would-be adopters:
- Where does virtualization fit in?
- Which type of clouds do they support: private, public or both, i.e., hybrid clouds?
- How well do HPC applications scale on their cloud solutions?
The paper includes an overview of several current first, second and third-tier HPC cloud solutions.
The authors hope that their work will act as a “helpful compass for someone trying to shift from standard HPC to large computations in cloud environments.”