Reasons to keep your on-site HPC center and avoid the cloud
This is co-released with a companion post Tips for Interactive HPC.
Scientific institutions today are considering how to balance their current HPC infrastructure with a possible transition to commercial cloud.
This transition is partially motivated by data science users, who find that HPC policies make their workflows difficult. This frustration is reasonable. HPC systems weren’t designed around these use cases, but there are steps we can take to adapt HPC centers for data science use.
But should we bother? Why might an institution choose to adopt cloud, and why might they choose to stay with in-house HPC resources? Some colleagues of mine recently published an opinion paper supporting a move to cloud, so I thought I would respond with some arguments in the other direction.
To be clear, I support a move to cloud in many cases, as they would support staying on in-house HPC. There is no single right answer.
With that prelude out of the way, here are some reasons:
You already have an in-house HPC center so you might as well use it. Adopting a technology is easy, but changing the culture of an institution is a multi-year effort that is going to be more intense than you expect. Even very nimble for-profit companies have a hard time at this, and if you have an in-house HPC center, then you are probably an organization with considerable inertia.
Your data may be generated in-house if you’re running simulations or making local observations then you might have data generation on-premises. If these datasets are likely to be reused many times by an entire field then please place them onto the cloud (this is what it is best at), however if your users are each creating their own large datasets that have only ephemeral value, then it may not make sense to publish these to the cloud, so you will want to support in-house computation instead.
- A few long-lived datasets serve a large community: Cloud
- Everyone makes their own short-lived datasets: In-house
Hiring: the cloud is a new technology. Actually, it’s a whole new suite of new technologies, and people who understand these well are in high demand by well-paying employers. If you have difficulty meeting market rates, then please verify that you’ll be able to hire or retrain experienced people before making this decision.
User familiarity: In some ways HPC systems are more familiar to scientific users today than cloud environments. For example when you move to the cloud you often lose things like POSIX file systems, file formats like HDF, and more. In the long run this is fine, there are excellent alternatives to these technologies, but a lot of user code will have to change a bit.
Lockin: The clouds will sell you products that are highly productive, but will make it difficult for you to leave them in the future. They will make it very easy to adopt tooling that differentiates them from other vendors and that you will, in time, rely upon heavily so that your public institution becomes a little bit of a captive. This happens with on-site HPC machines and software too, but it’s a bit more prevalent in the cloud.
Arguably, it’s good for government funded public institutions to maintain some level of independence from commercial cloud.
Again, there are plenty of reasons to move to Cloud. It can be cheaper, easier to maintain, and open up new ways of doing business/science that aren’t possible with HPC centers. But we should acknowledge that in many cases it will also be the wrong choice, especially for some classes of compute and data intensive fields. I’ve listed a few reasons above on why sticking with an HPC center can make sense, but there are more.
blog comments powered by Disqus