Ajay Kattepur

Resource Constrained Offloading in Fog Computing

When focusing on the Internet of Things (IoT), communicating and coordinating sensor--actuator data via the cloud involves inefficient overheads and reduces autonomous behavior. The Fog Computing paradigm essentially moves the compute nodes closer to sensing entities by exploiting peers and intermediary network devices. This reduces centralized communication with the cloud and entails increased coordination between sensing entities and (possibly available) smart network gateway devices. In this paper, we analyze the utility of offloading computation among peers when working in fog based deployments. It is important to study the trade-offs involved with such computation offloading, as we deal with resource (energy, computation capacity) limited devices. Devices computing in a distributed environment may choose to locally compute part of their data and communicate the remainder to their peers. An optimization formulation is presented that is applied to various deployment scenarios, taking the computation and communication overheads into account. Our technique is demonstrated on a network of robotic sensor--actuators developed on the ROS (Robot Operating System) platform, that coordinate over the fog to complete a task. We demonstrate 77.8% latency and 54% battery usage improvements over large computation tasks, by applying this optimal offloading.

Service demand modeling and performance prediction with single-user tests

Abstract Performance load tests of online transaction processing (OLTP) applications are expensive in terms of manpower, time and costs. Alternative performance modeling and prediction tools are required to generate accurate outputs with minimal input sample points. Service Demands (time needed to serve 1 request at queuing stations) are typically needed as inputs by most performance models. However, as service demands vary as a function of workload (load dependent service demands), models that input singular service demands produce erroneous predictions. The alternative, which is to collect service demands at varying workloads, requires time and resource intensive load tests to estimate multiple sample points—which defeats the purpose of performance modeling for industrial use. In this paper, we propose a service demand model as a function of concurrency that can be estimated with a single-user test. Further, we analyze multiple CPU performance metrics (cache hits/misses, branch prediction, context switches and so on) using Principal Component Analysis (PCA) to extract a regression function of service demand with increasing workloads. We use the service demand models as input to performance prediction algorithms such as Mean Value Analysis (MVA), to accurately predict throughput at varying workloads. This service demand prediction model uses CPU hardware counters, which is used in conjunction with a modified version of MVA with single-user service demand inputs. The predicted throughput values are within 9 % deviation with measurements procured for a variety of application/hardware configurations. Such a service demand model is a step towards reducing reliance on conventional load testing for performance assurance.

Model Driven Software Performance Engineering: Current Challenges and Way Ahead

Performance model solvers and simulation engines have been around for more than two decades. Yet, performance modeling has not received wide acceptance in the software industry, unlike pervasion of modeling and simulation tools in other industries. This paper explores underlying causes and looks at challenges that need to be overcome to increase utility of performance modeling, in order to make critical decisions on software based products and services. Multiple real-world case studies and examples are included to highlight our viewpoints on performance engineering. Finally, we conclude with some possible directions the performance modeling community could take, for better predictive capabilities required for industrial use.

Resource optimization in fog enabled IoT deployments

Internet of Things (IoT) devices are typically deployed in resource (energy, computational capacity) constrained environments. Connecting such devices to the cloud is not practical due to variable network behavior as well as high latency overheads. Fog computing refers to a scalable, distributed computing architecture which moves computational tasks closer to Edge devices or smart gateways. As an example of mobile IoT scenarios, in robotic deployments, computationally intensive tasks such as run time mapping may be performed on peer robots or smart gateways. Most of these computational tasks involve running optimization algorithms inside compute nodes at run time and taking rapid decisions based on results. In this paper, we incorporate optimization libraries within the Robot Operating System (ROS) deployed on robotic sensor-actuators. Using the ROS based simulation environment Gazebo, we demonstrate case-study scenarios for runtime optimization. The use of optimized distributed computations are shown to provide significant improvement in latency and battery saving for large computational loads. The possibility to perform run time optimization opens up a wide range of use-cases in mobile IoT deployments.

Distributed optimization in multi-agent robotics for industry 4.0 warehouses

Robotic automation is being increasingly proselytized in the industrial and manufacturing sectors to increase production efficiency. Typically, complex industrial tasks cannot be satisfied by individual robots, rather coordination and information sharing is required. Centralized robotic control and coordination is ill-advised in such settings, due to high failure probabilities, inefficient overheads and lack of scalability. In this paper, we model the interactions among robotic units using intelligent agent based interactions. The autonomous behavior of these agents requires task/resource allocation to be performed via distributed algorithms. We use the motivating example of warehouse inventory automation to optimally allocate and distribute delivery tasks among multiple robotic agents. The optimization is decomposed using primal and dual decomposition techniques to operate in minimal latency, minimal battery usage or maximal utilization scenarios. These techniques may be applied to multiple deployments involving coordination and task allocation between autonomous agents.

Distributed Optimization Framework for Industry 4.0 Automated Warehouses

Robotic automation is being increasingly proselytized in the industrial and manufacturing sectors to increase production efficiency. Typically, complex industrial tasks cannot be satisfied by individual robots, rather coordination and information sharing is required. Centralized robotic control and coordination is ill-advised in such settings, due to high failure probabilities, inefficient overheads and lack of scalability. In this paper, we model the interactions among robotic units using intelligent agent based interactions. As such agents behave autonomously, coordinating task/resource allocation is performed via distributed algorithms. We use the motivating example of warehouse inventory automation to optimally allocate and distribute delivery tasks among multiple robotic agents. The optimization is decomposed using primal and dual decomposition techniques to operate in minimal latency, minimal battery usage or maximal utilization scenarios. These techniques may be applied to a variety of deployments involving coordination and task allocation between autonomous agents. Received on 29 May 2018; accepted on 22 July 2018; published on 13 August 2018

A-Priori Estimation of Computation Times in Fog Networked Robotics

Mobile robots and drones have limited onboard computation power, which severely restricts mission planning. With the emergence of Fog Computing, computations may be offloaded to robotic peers, smart gateway devices and remote Cloud virtual machines. In order to effectively make use of such resources, a-priori estimation of execution times of offloaded computational programs is necessary. In this paper, we make use of profiling tools to accurately measure execution times of runtime computations on development testbeds. By exploiting performance benchmarks, we estimate the processing times on heterogeneous robot/Fog/Cloud deployment hardware as well as with varying data sizes. This allows us to determine optimal computational offloading strategies with differing computational complexities, data sizes, parallel processing and heterogeneous devices. We demonstrate our approach on multiple image, video and map processing algorithms deployed using OpenCV. Such design time analysis is crucial for mission planning of autonomous networks of robots and drones.

Service Demand Modeling and Prediction with Single-user Performance Tests

Performance load tests of online transaction processing (OLTP) applications are expensive in terms of manpower, time and costs. Alternative performance modeling and prediction tools are required to generate accurate outputs with minimal input sample points. Service Demands (time needed to serve 1 request at queuing stations) are typically needed as inputs by most performance models. However, as service demands vary as a function of workload, models that input singular service demands produce erroneous predictions. The alternative, which is to collect service demands at varying workloads, require time and resource intensive load tests to estimate multiple sample points -- this defeats the purpose of performance modeling for industrial use. In this paper, we propose a service demand model as a function of concurrency that can be estimated with a single-user performance test. Further, we analyze multiple CPU performance metrics (cache hits/misses, branch prediction, context switches and so on) using Principal Component Analysis (PCA) to extract a regression function of service demand with increasing workloads. We use the service demand models as input to performance prediction algorithms such as Mean Value Analysis (MVA), to accurately predict throughput at varying workloads. This service demand prediction model uses CPU hardware counters, which is used in conjunction with a modified version of MVA with single-user service demand inputs. The predicted throughput values are within 9% deviation with measurements procured for a variety of application/hardware configurations. Such a service demand model is a step towards reducing reliance on conventional load testing for performance assurance.