By Sramana Mitra and guest author Shaloo Shalini
MA: We do have lot of data that needs to be managed; it is almost close to petabytes of data. Most of it is coming from next-generation sequencers, high-resolution imaging, and while folks at the hospital use high-performance computing, it is not fully set up for the HIPAA data. So, they can’t bring in patient data, but they can bring over basic science data like genetics data.
SM: I see. Given that context can you help me understand what is your organization activity in the realm of cloud computing?
MA: The primary business of HMS is research. We get in the order of close to a billion dollars in research funding when it comes to the medical school. I call the med school a land of thousand CIOs because each individual lab can pretty much set up however it wants to set up its high-performance computing (HPC) infrastructure. In order for this research to be more efficient, we looked at cloud computing. It is not efficient for everyone to deploy infrastructure independently and manage it as it grows. It is more cost effective to have centralized infrastructure. In essence, we provide HPC capabilities in terms of research software, the storage in petabyte range, and other services related to helping them get research done which is everything from looking at modeling and simulations and trying to find the cure for cancer.
[Note to readers: You may want to refer to this post by HMS CIO John D. Halamka’s where he gives a set of details about the data center and infrastructure available for research. They are gearing for green and power-efficient systems and technologies. Details are available here.]
SM: How do you visualize the progress of cloud computing within the Harvard Medical School today? Has it gone beyond pilots and evaluation? Are you in full production deployment?
MA: We started about five years ago, even before the term cloud was thrown around to describe this form of computing. We said that we have a heterogeneous environment. We have researchers who run all kinds of different code and software applications. What we need is a cloud infrastructure that allows the faculty to have some additional capacity that they have access to when needed, without putting in any money in terms of buying the infrastructure for individual labs. We would like them to be able to purchase additional capacity in the cloud, if they required. So, we got initial funding to do a small pilot and it has been extremely successful. We have gone from the cloud having something like a few hundred processors to over 2,000 processors today. We are recently awarded an IAH grant for almost $4 million to expand the cloud even further.
SM: What kind of workloads have you moved to the cloud, and what are you planning to move as you go along?
MA: Most of our workloads are pipelines of research so you have things such as Matlab and survival analysis techniques (SAS) are the most common things being run in the cloud. We also have custom-written code, lot of open source stuff that people earlier used to run on their workstations but now they run those workloads on the cloud and things like that. Then we have R, which is an open source statistical language that is popular with researchers. A lot of sequencer pipeline and managed data that comes out of the next generation sequencer in the order of terabytes a day is also part of that workload. You need to move in place capacity to meet the needs of the researchers. We have everything from people hand-coding custom applications to pipes and C++ stuff to more commercial applications. But because the business of what we are trying to accomplish is to get research done most of the applications are research pipeline type.
SM: So it is largely analytics oriented and simulation applications, right?
SM: You mentioned your researchers buying a certain amount of capacity to begin with and then supplementing with cloud capacity. Can you explain a bit about how the business model works? Is there any transfer pricing going on?
MA: We allow faculty or staff and researchers to purchase nodes that become part of the cloud, so if they want to purchase 500 CPUs worth of capacity that gets put into the cloud, they get guaranteed access to that capacity. What this means is when they want to use it they have it. They don’t need to preempt any work that is running on the cloud at that point of time. But if their capacity is idle, then other people’s jobs can run on that idle capacity on that part of the cloud at that time. So they physically own that piece of hardware. Now they can buy only the hardware off the list of systems that we will support within the cloud. But if they are ever not happy being a part of the cloud for some reason, then they can take their nodes and infrastructure away and use them as standalone servers in their labs. So far this model works well for us. Since much of the funding is via National Institutes of Health (NIH) funding, there is a lot of restriction about how we can charge back. We have to show that we are not double billing for any infrastructure. I would actually like to have a sort of lease program for the cloud where people purchase dedicated capacity. But it is more a virtual purchase than a physical purchase. It is easy to go buy a server, slap an asset tag on it, and say it was bought from a grant and as such set up an NIH qualified service center.
SM: Interesting! So the whole cloud infrastructure that you have set up – is that a private cloud that you own and manage?
MA: That is correct.