|
David Garlan Professor of Computer
Science
Home
¦ Projects
¦ Publications
¦ Academics
¦ External
Activities
|
|||
|
|
|||
|
Current
Projects Carnegie
Mellon University's ABLE Project conducts research leading to an engineering
basis for software architecture. Components of this research
include developing ways to describe and exploit architectural styles,
providing tools for practicing software architects, and creating formal
foundations for specification and analysis of software architectures and
architectural styles. Furthermore, the ABLE group is researching how to
cope with emerging computing challenges of ubiquity, pervasiveness,
heterogeneity, mobility, and naive users. The
CAMELOT (autonomiC plAtform
for MachinE Learning using anOnymized
daTa) project aims at developing a highly
innovative machine learning platform that will tackle three key challenges: ·
Ensuring real-time constraints
during both the training and inference phases of machine learning models,
while minimizing operational costs deriving from the utilization of
computational resources from public cloud providers. ·
Enabling learning over anonymized
data, thus circumventing the privacy issues that arise when treating
sensitive data that currently prevent the reuse of information across models
trained over datasets belonging to different entities (e.g., the customers of
different financial institutions in fraud detection’s applications). ·
Modern machine-learning based
applications generate heterogeneous workloads that are supported by a large
ecosystem of diverse data platforms, each specialized to cope with specific
use cases (key-value stores, relational databases, graph-databases). Ensuring
that these various data stores are kept in sync, while ensuring an adequate
level of consistency and performance, is a time-consuming and error-prone
process that hinders the productivity of data scientists, as well the
efficiency of existing machine learning platforms. The
overarching goal of the AIDA (Adaptive, Intelligent and Distributed Assurance
Platform) project is to devise and implement a new version of the current
WEDO platform in which some of the pipeline phases can be dynamically moved
to the edges of the information system. As of today, the RAID platform is
fully deployed in homogenous and physically colocated
servers, either on premises or on the cloud. While some phases, such as
notification and discovery, are likely to always be kept under the platform's
service owner, with AIDA data collection and monitoring, and even actuation
phases should be prepared to run in diverse hardware architectures (e.g.
end-user devices, routers and gateways, telcos base stations, cloudified microservices
etc.) outside the platform's service owner physical or even administrative
control. AIDA should still be able to allow highly configurable and rich data
collection and monitoring while preserving the current real-time, security
and dependability guarantees of the RAID platform. More information can be
found here. DARPA has supported mobile
robotics research for decades, resulting in exciting advances and incredible
demonstrations—but remarkably few successful long-term deployments, because
it is so difficult to adapt robotics software in response to even small
changes in the ecosystem. The underlying problem is the low level of
abstraction at which robotics code is written, making code difficult to
evolve. Local band-aids in the code enable small-scale adaptation but make
the code even more brittle to larger-scale changes. In this project, we are exploring
a program of transformative research that will fundamentally raise the level
of abstraction at which we build and evolve mobile robotics software. In
order to adapt to a broad set of ecosystem changes, we will explicitly model
the software ecosystem as a software architecture, capturing the high-level
intent of the system and its components in domain-specific languages tailored
specifically for that purpose. Informed by variability-aware analysis that
can discover how software properties vary within a multidimensional
configuration space, we apply architecture-based self-adaptation to compute
optimized adaptations in response to a change, then apply those adaptations
using novel program transformation and repair techniques. To reduce the cost and improve the
reliability of making changes to complex systems, we are developing new
technology supporting automated, dynamic system adaptation via architectural
models, explicit representation of user tasks, and performance-oriented
run-time gauges. This technology is base don
innovations in three critical areas: 1)
Detection: the ability to determine dynamic (run-time)
properties of complex, distributed systems, 2) Resolution: the ability
to determine when observed system properties violate critical design
assumptions, and 3) Adaption: the ability to automate system adaptation
in response to violations of design assumptions. These new capabilities
will provide both (a) the ability to handle system changes with respect to
the specific (performance-oriented) gauges supported by our technology, and
(b) an extensible framework to handle additional gauges and system adaptation
strategies produced by others. In aggregate, the capabilities will
dramatically reduce the need for user intervention in adapting systems to
achieve quality goals, improve the dependability of changes, and support a
whole new breed of systems that can perform reliable self-modification in response
to dynamic changes in environment. We will demonstrate these
improvements in the context of complex real time information systems
supporting distributed collaboration and planning. Specifically, we
will show how our technology enables automatic system adaptation in the
presence of significant variations in processing and network capabilities,
and for dynamically evolving workloads, while maintaining critical
architectural constraints. |
Past Projects
The most precious resource in a
computer system is no longer its processor, memory, disk or network.
Rather, it is a resource not subject to Moore's law: User Attention.
Today's systems distract a user in many explicit and implicit ways, thereby
reducing his effectiveness. Project Aura will fundamentally rethink
system design to address this problem. Aura's goal is to provide each
user with an invisible halo of computing and information services that persists
regardless of location. Meeting this goal will require effort at every
level: from the hardware and network layers, through the operating system
and middleware, to the user interface and applications. Project Aura will
design, implement, deploy, and evaluate a large-scale system demonstrating the
concept of a "personal information aura" that spans wearable,
handheld, desktop and infrastructure computers.
RADAR (Reflective Agents with
Distributed Adaptive Reasoning) is a flagship research project within Carnegie
Mellon University to develop a personal cognitive assistant that integrates
with current desktop and applications, and helps users to carry out routine
tasks, such as organizing meetings, answering routine emails, managing web
pages, etc. RADAR is composed of several specialist components that have
knowledge about how to do a task, and which learn over time user preferences
and idiosyncrasies when performing tasks. The ABLE group is researching
the software architectural style that is required to put together such a
system, and also in providing task management support within RADAR.
Specification and Verification Center
Our center focuses on the formal specification
and verification of hardware and software systems. We invent new
mathematically-based techniques, languages, and tools to model the behavior of
systems and to verify that these models satisfy desired properties. We
also use our tools to find bugs in hardware and software designs. Thus,
our approach of using formal methods complements the more traditional
approaches of simulation and testing. Our challenges are in modeling
large, complex systems and in verifying behavioral properties of concurrent,
distributed, real-time, and resource-constrained systems. To meet these
challenges, we do fundamental research on data structures and algorithms, data
and control abstractions, specification logics, and compositional proof
techniques; we build tools such as model checkers, proof checkers, and
combinations of the two; we apply our methods to a diverse range of
applications: automotive controllers, circuit designs, communication
protocols, disk arrays, distributed simulation architectures, file systems,
networked systems, robots, security protocols, and spacecraft.