HPC on-demand: the impact on training
Stephane Raynaud reports on the evolution in training content and style for HPC on-demand users
By its nature, high-performance computing (HPC) has a narrowly focused market of scientific and one-time projects funded by large supercomputing centres and academic/research institutions. Recently, traditional HPC centres, IT providers, and hardware vendors have begun offering on-demand access to HPC infrastructures and are thus expanding this market.
From a financial standpoint, HPC on-demand is possible because the upfront investments of the systems are now spread across a larger population, thereby removing a long-standing hurdle that previously prevented small and medium enterprises access to HPC resources. This means new sectors, such as small- to medium-sized manufacturers, can now benefit from advanced HPC-based technologies, such as modelling, simulation, and analysis. However, these benefits come with new challenges. One such challenge is related to training these new types of users. Since this expansion is making an HPC environment available to users who have a different approach to computing, their usage requirements need to be matched by a modified training.
Until recently, most HPC users worked in national labs and were a relatively consistent set of employees involved in long-term research programmes. They were well supported by internal champions and Independent Software Vendors (ISVs) that continuously contributed time to training users and maintaining knowledge on the available tools. In contrast, HPC on-demand users have very different characteristics – they use HPC facilities intermittently and for short bursts of time. While these new users may be familiar with traditional cloud computing, which provides them with a bare-bones infrastructure to scale applications, they are less familiar with the HPC on-demand framework, which offers remote systems with pre-installed environments that are configured and targeted toward scientific computation.
Even though some uniformity does exist in the HPC world in terms of software stacks, the environments remain largely heterogeneous and therefore the users will be exposed to toolsets that they have never used before. As a result, these new users are not well versed in the tools or systems, nor do they have a desire to gain in-depth, working knowledge of the infrastructure. Furthermore, these users will rarely want to participate in full-scale training about a system they will only be using sporadically. This creates new challenges for developing successful training procedures for HPC on-demand offerings.
In the typical HPC world, the most common type of training requests are either for an introductory course at the original purchase and delivery of the system, or for a set of advanced trainings for fine tuning the skillsets of seasoned users. As training specialist in the HPC market, we have begun to notice that the increase of on-demand users is causing a change in the type of training requests. There is a noticeable shift from users looking to learn all the bells and whistles to users wanting to maximise the use of tools for a specific problem. This is because the goal of HPC on-demand users is to fix a specific problem – immediately. A good analogy is when customers rent a tool from a home improvement store – they are not interested in learning the craft, but instead want a few pointers on how to optimise their experience with the tool in order to shorten a specific renovation project.
As HPC on-demand users become more common, HPC facilities will be challenged to provide more frequent and accessible training. In essence, they will have to take on a consumer-oriented model, in which the advice is narrowly focused to the task at hand, versus the model of a technical college that provides a full curriculum for developing an expert.
During one of my recent visits to a large HPC centre, one of its performance computational scientists explained that training infrequent users is already, to some degree, an existing challenge faced by the traditional HPC market. Innovative programmes, similar to HPC on-demand, have created a recent trend of hosting temporary programmes that could represent up to dozens of projects that are awarded sizable allocations of HPC systems every year.
These projects typically run for one or two years and participation ranges from advanced science and academic research to commercial applications. Having a short time line, these programmes basically create a ‘revolving door’ of new faces every year, from all walks of the development community at large. In many cases, the facilities have to reinstate training every year in order to ‘show the ropes’ to new users. While this recent trend helps poise HPC centres to be better prepared for new training requests, HPC on-demand tends to bring an even shorter window of time on the systems, thus requiring even greater flexibility and reactiveness.
In order to prepare HPC facilities for the influx of inexperienced users, our recent strategy has been to help HPC on-demand vendors create their own introductory training packages for the set of tools hosted on their systems. While some HPC facilities are still reluctant to adopt this new approach, many facilities have embraced this proactive method. In addition, many HPC facilities are finding our newly-developed training packages, which can be easily integrated into the HPC offering and resold to the end user, to be a useful education tool. Ultimately, this new paradigm shifts the responsibility for training the end user from the ISV to the HPC on-demand provider.
As mentioned earlier, there is a large evolution in training content and style for HPC on-demand users. Due to the requirements of this new group, it is more productive to allow trainees to bring their current projects to the training versus leveraging the traditional classroom experience with pre-defined labs. While this new type of training is an abridged version of larger, more customised packages for traditional HPC users, the HPC on-demand offering is more time and cost efficient – fitting the practical needs of the HPC on-demand user. Because this type of training forces trainers away from a traditional teacher role to a role in which they get their hands dirty during the sessions, it is important to have trainers that have hands-on experience. This type of trainer will be able to better address the training needs of the new HPC on-demand users, which will help ensure a cost-efficient and productive use of the facilities.
After identifying this shift in users, on-demand HPC vendors and ISVs have tailored their offerings, training schedules, materials, and even the skillsets of instructors to ensure that on-demand HPC users achieve all their objectives, while getting maximum value from the runtime they buy. Accordingly, for the HPC on-demand model to be successful, HPC facilities need to put in place new and appropriate training procedures and create partnerships with ISV providers to develop new offerings tailored to address these new challenges.
Stephane Raynaud is technical services manager at Rogue Wave