Ma[R]s
Memory as
a [Reconfigurable]
service
Ma[R]s
Memory as a
[reconfigurable]
service
for Distributed Application
as a Service
as a Service
advancing STRONGLY CONSISTENT DISTRIBUTED SHARED MEMORY SYSTEMS (DSM)
Supervised by experts from academia (University of Cyprus) and industry (Algolysis Ltd), our PhD Student is dedicated to advancing research and industrial development in Strongly Consistent Distributed Shared Memory Systems (DSM).
Dual focus
Ma[R]s in Theory:
To advance the knowledge in the field of DSM by exploring efficient, robust, and practical solutions in highly dynamic environments.
Ma[R]s in Practice:
Creating a Memory-as-a-Service (MaaS) platform for deploying and managing DSM, supporting next-gen distributed applications. Real-World Validation: Testing DSM algorithms on stationary servers and less-powerful peer-to-peer devices for practical testing and validation.
Motivation - How Ma[R]s makes a difference
The data explosion and the increasing complexity of computational problems have driven the demand for distributed applications in recent years. Distributed applications enable organizations to harness the power of multiple computers or nodes in a distributed system, enhancing scalability, fault tolerance, and performance. Decentralized computation, exemplified by technologies like blockchain, promotes transparency and trust. Designing and building dependable distributed applications is challenging due to communication complexities like asynchrony, message delays/losses, and node failures. DSMs are fundamental for creating complex, decentralized, cloud applications in emerging technologies (e.g., IoT, VR/AR), they may offer a transparent cloud memory space where distributed applications can store, retrieve, and coordinate over shared data. An Atomic DSM (ADSM) provides the illusion of a sequential memory space over asynchronous, fail-prone, message passing nodes, simplifying the development process. An ADSM keeps data copies in multiple network locations (replica hosts or servers) to ensure data availability and survivability. The main challenge comes when those copies are accessed concurrently by different processes or clients: how can we ensure that data copies remain consistent? Several ADSM algorithms have been proposed, some for static replica host sets and others for dynamic sets. Although these algorithms theoretically proven to satisfy atomic guarantees, only a handful concern about practicality issues (e.g., support for large data objects, operation speed and liveness), and their scalability has not been closely examined.
Main Outcomes
ADSM algorithms
Project Objectives
Building on state-of-the-art dynamic ADSM algorithms (e.g., CoARESf), the first objective is to devise methodologies to reduce the latency of read and write operations, making dynamic ADSM services more practical and attractive for commercial use. To reach our goals we will: (i) use distributed tracing tools within existing PoC implementations to identify performance bottlenecks during executions on the Emulab testbed, and (ii) propose solutions for eliminating (or at least reducing) the identified performance bottlenecks; ideas for the optimizations can be retrieved from other works that use reconfiguration mechanisms, like RAMBO.
Utilizing the reconfiguration mechanism offered by the ADSM services, in this objective we will design and analyze (smart) Reconfiguration Orchestration Strategies (ROS) on when reconfigurations should be invoked on the ADSM service and how the membership of the ASDM service should change. Proposed approaches will specify which environmental parameters may affect the decisions on when and how to reconfigure, and reconfiguration decisions will utilize (existing) tools that monitor these parameters, eliminating or minimizing human intervention. These parameters will involve, for example, network connectivity, server’s health, and storage capacity. The developed service will be designed to interact with any dynamic ADSM service via the reconfiguration mechanism.
The focus of this objective is to bring the developments in Objectives O1 and O2 together, yielding a dynamic ADSM service with ROS support. More precisely, we will implement an ADSM service based on the algorithms proposed in O1, and a Reconfiguration Orchestration Module (ROM) as an external service given the outcomes of O2. Integration of the two services will be accomplished by implementing an API that exposes the reconfiguration operation of the ADSM service. An API will also be used for the interaction of the ROM service with the different monitoring tools, which will be deployed to collect the data necessary for the ROM to generate reconfiguration decisions. The integrated service will be deployed on a real testbed (e.g., Emulab) and evaluated against five (5) characteristics: (i) scalability in terms of concurrent users and object sizes the service may support, (ii) operation speed (especially compared with previous ADSM algorithms), and (iii) fault-tolerance and liveness, mainly testing the reaction of the ROM service during harsh environmental conditions.
