Aravind Srinivasan is a distinguished university professor and professor of computer science at the University of Maryland, College Park. Srinivasan’s research interests lie at the intersection of algorithms, continuous and combinatorial optimization, and machine learning. The unifying theme in Srinivasan’s work is the deployment of probabilistic methods across a wide variety of areas, from cloud computing, resource allocation, machine learning, and e-commerce, to developing computational approaches in epidemiology and for sustainable growth in electrical and power networks.
Srinivasan is an Amazon Scholar, a select group of academics who work at Amazon on large-scale technical problems, while also conducting research at their universities. As an Amazon Scholar, Srinivasan develops algorithms for Amazon EC2, AWS Lambda, and Amazon Elastic Block Store that allocate customer workloads in the cloud in a matter of milliseconds.
Srinivasan’s involvement with Amazon began prior to him joining the company as a Scholar in 2019. Srinivasan received an Amazon Research Award in 2018. With the grant, he is working on how to improve fairness in a broad range of resource-allocation contexts, including cloud computing and digital marketing.
Srinivasan and his colleagues also conduct research to innovate on the traditional susceptible-infected-recovered (SIR) model used to model the spread of contagious diseases, like the coronavirus pandemic. Solutions to SIR and related models were traditionally applied to homogenous populations. The work done by Srinivasan’s team enables epidemiologists to develop more nuanced networked models that will allow for the identification of super-spreaders, design interventions for high-risk populations like first-responders, and determine the segments of the population that should receive vaccination first so as to slow down the spread of the disease.
Amazon Science sat down with Srinivasan to find out more about his work related to combinatorial optimization, trends he sees emerging in the field, and his work at Amazon.
Q. What got you interested in the field of combinatorial optimization?
At its very essence, combinatorial optimization seeks to use mathematical methods to identify the best solution for a problem from a large set of possibilities. I found the mathematical elegance that is at the heart of combinatorial optimization highly appealing. In addition, I was drawn to the field by the growing practical importance of combinatorial optimization.
Some of the most interesting problems being solved in a wide variety of fields from cloud computing to energy and medicine have to do with combinatorial optimization. Across all these sectors, we are dealing with scenarios that involve a large number of workloads coming in every second. This gives rise to some really interesting questions. How will you balance the loads? How will you place them on different servers? How will you plan for failures?
These are fundamentally discrete optimization problems, where the variables you want to compute aren’t from a continuous range. For example, for each workload that comes in you have a decision to make: should I place this on a particular server or elsewhere? And when you have a large number of discrete variables, things begin to get really interesting.
Q. What are you working on at Amazon?
I work with several teams within the AWS organization. My work involves developing the most optimal ways to deal with the high volume of workloads – be it configuring capacity for EC2 or provisioning servers for running jobs on Lambda – all in a matter of milliseconds.
This can be a really complex problem at the high volumes that are typical with these services. Workers or servers handling the workloads have different capacities in terms of both processing power and memory. Then you have other variables like daily or weekly patterns and spikes having to do with seasonality.
Amazon offers a really rich playing field to work in this area. The range of challenges is large. And everything’s taking place at massive scale. There’s a real opportunity to have very meaningful customer impact. I’ve only been at Amazon for a year -- and I’ve already had the opportunity to work with so many different groups, go deep into issues, and develop prototypes at a rapid clip.
Some of the most interesting problems being solved in a wide variety of fields from cloud computing to energy and medicine have to do with combinatorial optimization.
Q. What are some of the interesting challenges that teams you work with are addressing with combinatorial optimization methods?
As we were just talking about, a lot of the work in the industry is related to discrete optimization. But the scale and complexity of problems at Amazon forces you to look beyond what’s typical, and address discrete optimization problems using continuous optimization methods.
Take the example of workload balancing that we were talking about earlier. You have a workload and you can allocate this workload across multiple servers. If you view the decision as a problem to do with probabilities, you can infer an optimal distribution. Now, our situation becomes a continuous distribution problem where all these probabilities add up to one – subject to the constraint that they are all non-negative.
The probabilistic assignment problem is a continuous optimization problem, while our real optimization problem is as a discrete optimization problem. The interplay of these two seemingly disparate fields is really interesting, however, these are just the kind of scenarios that you encounter across a wide range of business scenarios at Amazon.
Another interesting area I’m working on is figuring out how to make neural networks more efficient. As you can imagine, there’s no shortage of people interested in figuring out how neural network inference can be made without utilizing so much time and memory.
Q. How do you see the field of combinatorial optimization evolving?
I already touched upon the blending of continuous and discrete methods that can be used to make probabilistic determinations about future data. Methods from continuous optimization such as interior-point methods (where we move in the interior of the feasible space to an optimal solution) continue to have much impact on combinatorial optimization and machine learning. The blending of continuous and discrete optimization methods is a very interesting trend in combinatorial optimization.
Q. How did you come to join the Amazon Scholars program?
I received an Amazon Research Awards grant in 2018. A short time after that, I spoke to a recruiter at Amazon. We had a conversation about areas within the company that were a good fit for me. Things happened very quickly after that.
I was intrigued at the possibility of joining a company where I had the opportunity to make real-world impact. To say that this impact happens at a large, planetary scale would not be hyperbole.
Working on problems like load balancing was also appealing to me from a sustainability perspective. If we make our data centers even more efficient for customers, we are at the same time decreasing our energy footprint and creating a greener planet. Part of my work some years prior to Amazon was focused on developing computational models for energy-efficient computation and sustainability. I view my work at Amazon as a natural progression of this work, where I can make a contribution to bettering the planet.
The opportunities to make a real-world impact in so many ways were what ultimately drew me to working at Amazon.
Q. What’s the “must read” research paper or book a student interested in combinatorial optimization should read?
It is hard to name a single resource, but if I had to choose only one, I would say Alexander Schrijver’s classic book on Combinatorial Optimization.
