Bursting the Clouds: SC10 Survey Sheds Light on HPC User Experiences
The results of a survey conducted at this year’s Supercomputing Conference in New Orleans provided key insights into how HPC users are making use of cloud computing.
Platform Computing, the company behind the survey, put a number of questions to over 100 IT executives from government, research, manufacturing and other industries to get to the heart of why organizations with high-performance computing needs are looking at clouds as a possibility and what, if any, experiences they had with current implementations.
According to the results, sixty-two percent of HPC users have at least experimented with private or public cloud computing models.
Platform also noted that there are a number of ways that the respondents were evaluating private cloud infrastructures, “including building shared infrastructures (36 percent) and bursting existing workloads (15 percent).” The survey also revealed that public clouds being used as a primary mode for running some applications with 23 percent stating that they had used such a resource.
As Randy Clark, Chief Marketing Officer for Platform said of the results, “For cloud computing, we see private clouds with and without public cloud bursting (or dedicated shared public resources) and then also applications that are just run on the public cloud.” He also stated that he knows of no customer or prospect that has “fully” adopted either model—whether public or private.
Cloud Bursting Model Still in Question
When it comes to sentiments about cloud computing for HPC users, there are some tentative signs that the tide might be shifting overall. The survey revealed that of the 100 IT professionals questioned, 74 percent reported that their experience with cloud computing was positive.
Interestingly, however, for those who had experimented with cloud bursting as a model, only 45 percent reported that their experiences were positive.
While the survey didn’t include a field to allow respondents to discuss why they were dissatisfied with cloud bursting experiences, Clark did provide some projections about why this might be the case in an interview. He states that it could be the “paradox that it is technically much easier to burst predictable workloads but it is much more economical to burst unpredictable workloads; this is forcing customers to experiment and really think about their workload scheduling.”
There is certainly some degree of skepticism about the viability for cloud computing, no matter what model (public, private, cloud bursting, etc.) for HPC applications due to performance and other perceived application-specific issues.
Of those surveyed during SC10 who had not yet experimented with cloud computing, only 26% stated that they planned to test it in the next year.
When asked about the cautious nature of this figure for the coming year’s experimentation, Platform Computing CMO Randy Clark stated, “We expect that cloud computing (including cloud bursting) will grow. Developments that are making it more feasible include improved integrations and optimizations between workload management, VMs, and AMIs (and other public cloud interfaces), more public cloud resources that are targeted for HPC applications as well as more collaboration between ISVs, middleware/management vendors and hosting/public cloud providers to offer focused or bundled solutions.”
Finding a Pulse for HPC Cloud Use
While only 26 percent of those who have not yet experimented plan on taking steps to do so over the coming year, it is worth noting that the survey did find that there were some key drivers behind the move that might force some to reconsider.
The main reasons respondents cited for evaluating cloud solutions for HPC included, “ability to have a more flexible infrastructure (33 percent); the ability to provision resources more quickly (19 percent); and the ability to address problems that were previously difficult to solve (16 percent)” and finally, cost effectiveness as a driver, with 12 percent citing this concern.
When asked if there was any way to determine differences between research and enterprise experiences with cloud computing, Clark admitted that one of the limitations of the dataset was that there was no way to determine any differences between cloud use or experimentation between research or scientific users and enterprise customers.
Randy Clark notes that “This might have been a valuable way to gauge how the two communities were making use of cloud computing for HPC applications” although the survey was brief and lacked fields for detail. Hopefully with the release of a new report from IDC later this year that will provide similar indicators there will be a clearer picture for 2011 in HPC cloud use.
It should be noted that this was by no means an exhaustive study; it is based on only 100 respondents. Still, it will be interesting to see how this data compares to more comprensive survey figures based on HPC cloud users in 2011.
While the survey results were certainly not as detailed as analyst figures (nor should we expect them to be) it does provide a firm starting point for looking at how use of cloud computing in HPC might occur over the following year.