Georgia Institute of Technology
Toward Adaptive Modular Computing Systems Through More Expressive and Elastic Hardware and Software Interfaces
Abstract
dc:description.abstractModern computing systems are built around a modular design philosophy in which hardware and software components are developed independently, interact through well-defined interfaces, and are provisioned in isolation based on assumptions about the behavior of their neighbors. This modular approach has enabled decades of productive specialization, allowing processors, memory technologies, network interfaces, and software stacks to evolve on independent trajectories without requiring tight coordination. The cost of this independence, however, is that each component is designed and provisioned based on static assumptions that the evolving technological landscape increasingly disrupts. Memory systems in particular have grown deeply heterogeneous — across technology characteristics, application demands, and the scaling trajectories of compute, memory, and network bandwidth — and the uniform abstractions that once made this heterogeneity manageable now actively prevent the coordination that performance requires. This thesis examines three specific manifestations of this challenge, each arising at a different layer of the memory stack and each characterized by the same underlying structure: a component that cannot make globally efficient decisions because the information it needs does not cross the abstraction boundary that separates it from the rest of the system. The first contribution addresses the interaction between high-bandwidth network interfaces and the cache hierarchy, where the cache has no visibility into the lifecycle of network data and cannot distinguish live from dead cache lines, forcing unnecessary writebacks that saturate memory bandwidth. A small interface change between the network stack and the cache hierarchy communicates data liveness across this boundary, eliminating the unnecessary traffic and recovering the memory bandwidth it consumed. The second contribution addresses persistent memory applications, which enforce crash-consistency ordering constraints that the hardware applies bluntly and uniformly, serializing operations that could safely overlap because the software has no channel to communicate which orderings are semantically necessary. A coroutine-based software interface paired with minimal hardware modifications exposes this information at the hardware-software boundary, recovering the memory-level parallelism that conservative enforcement destroys. The third contribution addresses the static provisioning of MSHRs, located at the boundaries of adjacent cache levels to track pending misses. MSHRs are sized at design time and cannot adapt to the memory-level parallelism requirements of the application or the latency characteristics of the memory system at runtime, despite the hardware observing the relevant behavior directly. A dynamic allocation mechanism that repurposes existing in-cache resources adjusts MSHR capacity at runtime, adapting to conditions that static provisioning could not anticipate. The contributions of this thesis are united by the common design philosophy of modularity and selectively enriching the interfaces between components to expose precisely the information that the abstraction currently hides. Richer interfaces restore coordination where the abstraction prevents it, and elastic interfaces replace static design-time decisions with dynamic runtime adaptation that responds to behavior the hardware already observes but previously could not act upon. The results demonstrate that investing in interface richness and elasticity at precisely scoped component boundaries is sufficient to recover substantial performance without compromising the composability and generality the principle of modularity provides. The information gaps they address are not isolated design oversights but a structural consequence of the modular philosophy, one whose cost grows as memory heterogeneity increases and the independent evolution of system components continues to outpace the expressiveness of the interfaces through which they interact.
Degree
thesis:*- Name thesis:degree_name
- Computer Science, PhD
- Department thesis:degree_department
- Computer Science
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Vemmou, Marina
- Advisor dc:contributor.advisor
-
- Daglis, Alexandros
- Committee members dc:contributor.committeemember
-
- Kim, Hyesoon
- Qureshi, Moinuddin
- Said Mohamed Tawfik Issa, Ahmed
- Grot, Boris
- ORCID person:identifier.orcid
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82244