Computing infrastructure—algorithms, abstractions, interfaces, silicon—embodies the assumptions of its time. These assumptions, over time, harden into conventions and percolate across the stack. When technology and workloads change, yesterday’s conventions become today’s binding constraints. Revisiting these assumptions and conventions can enable architectures that were previously impossible.
Distributed systems & algorithms
OS memory management
OS storage stack
OS network stack
Network fabric & protocol
Host interconnect & silicon
Operating points
A server as the unit of computation. A host as the endpoint of a distributed system. Best-effort delivery as the core principle underlying all data transfers. Each shaped the entire stack. Each looked permanent. None of them was.
We identify these assumptions and rearchitect the stack they shaped—from algorithms to silicon. The result is systems that reach previously unattainable operating points in predictability, safety, and efficiency.
I love teaching! My views on education are best captured in: “Creativity is as important in education as literacy!”
Research Programs
-
Distributed systems & algorithms
OS memory management
OS storage stack
OS network stack
Network fabric & protocol
Host interconnect & silicon
Safety
Machines That Move
New research program · 2026–Existing computing infrastructure treats the physical dynamics of a machine as external to the computational substrate—it essentially assumes that machines stay still. The stacks are built around that assumption.
Machines that move—a robot in a kitchen, a tractor in a field, a rack in a datacenter—break that assumption. The physical dynamics decide the machine’s connectivity; the connectivity decides the substrate and the feasible computation; the computation competes for a shared energy budget with the machine’s actuators; and the actuators decide the amount of time left for the machine to finish the computation. In short: the physical dynamics of a machine change the feasibility, predictability, efficiency, and safety of the computation itself.
Building a stack for machines that move—one that treats physical dynamics as part of the computational substrate—requires rethinking distributed systems and algorithms, operating systems, networking, interconnects, and silicon. This is the next instance of the research programs below: the physical substrate (machines in this case) has changed, and so must the stack built around the old one.
Distributed systems & algorithms
OS memory management
OS storage stack
OS network stack
Network fabric & protocol
Host interconnect & silicon
Efficiency
Your Host Is a Distributed System
Ongoing research program · 2021–A host has traditionally been considered the endpoint of a distributed system. This was the right abstraction when everything inside the host looked uniform—the assumption was that hosts are composed of uniform compute devices, uniform memory devices, and a unified coherent interconnect that provides nanosecond latencies, ample bandwidth, no loss. This assumption is no longer true: modern hosts have multiple heterogeneous compute devices (CPUs, GPUs, accelerators) and multiple heterogeneous memory devices (DRAM, HBM, CXL, etc.) with multiple interconnects each forming a different coherence domain. Continuing with this abstraction has now become the binding constraint in today’s stacks: nanosecond-scale inefficiencies within the host can percolate through hardware and software stacks to create microsecond- or even millisecond-scale inefficiencies at the application layer.
In this research program, we are rearchitecting the stack with the lens of the host as a distributed system: multiple devices joined by processor, memory, and peripheral interconnects, with queueing, routing, scheduling, flow control—and congestion—within the host. Preliminary results suggest that this perspective enables previously unachievable operating points in predictability, safety, and efficiency.
Recognition: ACM SIGCOMM doctoral dissertation award · Cornell Bowers CIS dissertation award · IRTF applied networking research prize · SIGCOMM ’24 best student paper award · Deployed at scale
-
Distributed systems & algorithms
OS memory management
OS storage stack
Network fabric & protocol
Host interconnect & silicon
Efficiency
Resource Disaggregation
Ongoing research program · 2015–A server as the unit of computation was the right assumption when interconnects within the server were orders of magnitude faster than the network: resources that worked together had to be bought together. Datacenter networks have since closed that gap. Workloads whose peaks dwarf their averages now pay for the old assumption continuously—in stranded memory, idle cores, and capacity that is provisioned for the peak rather than the average.
Resource disaggregation—where each resource is built as a standalone blade, and various resource blades are interconnected via a network fabric—enables overcoming these inefficiencies via improved elasticity and statistical multiplexing. While beneficial, resource disaggregation alters several assumptions that once guided the design and optimization of existing applications, systems, and networks (e.g., CPU-storage colocation, high CPU-memory bandwidth, storage hierarchy, data locality, failure models). Capitalizing on the benefits of resource disaggregation thus requires rearchitecting the entire stack—from distributed systems and algorithms, to operating systems, to network fabrics and protocols, to silicon. Our decade-long work on resource disaggregation establishes that such an architecture enables previously unachievable operating points in terms of predictability and efficiency.
Recognition: Google faculty research award · Deployed at scale
-
Distributed systems & algorithms
OS memory management
OS storage stack
OS network stack
Network fabric & protocol
Host interconnect & silicon
Predictability
Best-effort Delivery as a Choice, Not the Only Choice
Ongoing research program · 2015–The network infrastructure has famously been built around the principle of best-effort delivery. This principle is at the very core of network design turning into an empirical craft: one builds it, measures it, tunes it, and hopes to get good performance. For the Internet, a fundamental assumption led to this principle: since no single entity owned the entire infrastructure, best-effort was the strongest promise available and empirical evaluation seemed to be the only way to reason about network properties and performance. For general-purpose datacenter networks as well as modern AI/ML clusters, this assumption breaks: a single entity now owns the entire network infrastructure from network stacks to protocols to fabrics to hardware. We can now ask a different question—not just how a design behaves and performs when deployed, but rather what can be proven about network properties and performance before deployment. Best-effort delivery as the only choice has now become the binding constraint.
Over the past decade, we have been revisiting this assumption across the entire network infrastructure: distributed network algorithms that are optimal rather than merely good; network fabrics and protocols that make congestion impossible rather than merely rare; and designs whose properties and performance are established rather than merely measured. The result is a network infrastructure that is easy to reason about in terms of predictability and efficiency. And a network infrastructure where best-effort becomes one of the choices, and not the only choice.
Recognition: SIGCOMM ’18 best student paper award · Google research scholar award · NSF CAREER award · Deployed at scale
Earlier research programs
Students
| Omar Eqbal | IIT KGP Gold Medalist | PhD student |
Alumni
| Midhul Vuppalapati | Assistant Professor, Purdue | PhD, 2026 |
| Qizhe Cai | Assistant Professor, UVA | PhD, 2024 |
| Saksham Agarwal | Assistant Professor, UIUC | PhD, 2024 |
| Sujaya Maiyya | Assistant Professor, University of Waterloo | Postdoc, 2022 |
| Mina Tahmasbi Arashloo | Assistant Professor, University of Waterloo | Postdoc, 2020–2022 |
| Jaehyun Hwang | Assistant Professor, Sungkyunkwan University | Postdoc, 2019–2021 |
| Anurag Khandelwal | Assistant Professor, Yale University | Postdoc, 2019 |
| Shreyas Kharbanda | Founding Engineer, Deft Robotics | MS, 2025 |
| Katie Gioioso | PhD student, Stanford | MS, 2021 |
Supported by the National Science Foundation and Alfred P. Sloan Foundation, with research gifts from Google, Samsung, and Snowflake.