Moset: An anonymous remote mobile cluster computing paradigm

M. A. Maluk Mohamed, A. Vijay Srinivas, D. JanakiramJ. Parallel Distributed Comput 2005 · DOI

TL;DR

This paper introduces Moset, a novel paradigm for anonymous remote mobile cluster computing (ARMCC) that integrates mobile devices with traditional static clusters. Moset aims to harness the idle computing power of both mobile and static nodes, providing resource-poor devices with access to significant computational power and enabling seamless parallel computing for mobile users.

What Problem Does This Paper Address?

The rapid advancement of mobile technology has led to an explosion of powerful and ubiquitous mobile devices. This trend, coupled with the increasing prevalence of wireless networks, presents an opportunity to integrate these mobile devices into existing cluster computing frameworks. However, traditional cluster computing models, designed for static and homogeneous environments, struggle to accommodate the unique challenges posed by mobile clusters. These challenges include:

  • Asymmetric Connectivity: Wireless networks offer lower bandwidth and higher error rates compared to wired networks, with significant variations depending on the mobile device's location and connection type.
  • Mobility of Nodes: Mobile devices constantly change their location, impacting data locality and requiring dynamic updates to network configurations and resource discovery.
  • Disconnectivity: Mobile nodes can temporarily disconnect from the network due to roaming or power-saving modes, which needs to be handled gracefully without halting computations.
  • Timeliness Issues: Mobile devices may experience delays during handoffs between network cells or when re-establishing connections, which can be critical for real-time applications.
  • Dynamic Load and Availability: The load on participating nodes and their availability fluctuate significantly due to multi-user environments, application demands, and the inherent mobility of devices.
  • Heterogeneity: Mobile clusters often comprise devices with diverse architectures, operating systems, computing capabilities, and memory capacities.

Existing mobile cluster computing proposals have addressed only a subset of these issues, often neglecting the complexities of parallel programming on such dynamic and heterogeneous environments. Moset aims to provide a comprehensive solution by abstracting these complexities from the user and enabling efficient parallel computation.

Methods

Moset proposes an Anonymous Remote Mobile Cluster Computing (ARMCC) paradigm. Its core principles include:

  • Abstraction of Heterogeneity: Moset hides the underlying hardware, bandwidth, operating system, and network differences from the user, presenting a unified computing platform.
  • Cluster-Subgroups: Nodes are grouped based on their memory capabilities, allowing for efficient task allocation. For example, subgroups are defined for tasks requiring <10MB, <50MB, <100MB, and >100MB of memory.
  • Horse Power Factor (HPF): A dynamic HPF is calculated for each participating node (static and mobile) to measure its computing power, current load, and network bandwidth. This factor normalizes the capabilities of diverse machines, allowing tasks to be allocated proportionally.
  • Dynamic Load Balancing: A non-preemptive load balancing mechanism is employed, using upper and lower thresholds on mobile hosts to manage the distribution of computing entities (threads).
  • Anonymous Remote Computing: The model extends the Anonymous Remote Computing (ARC) model, providing anonymity to applications and nodes.
  • Exactly-Once Reliable Multicast Protocol: Moset is built upon a reliable multicast protocol to ensure data delivery despite the inherent unreliability of wireless networks and mobile devices.
  • Hierarchical Coordination: A multi-level coordination system is implemented:
  • Local Coordinators (LC): Run on both static and mobile hosts, managing local activities and communication.
  • Co-Coordinators (CC): Reside on Mobile Support Stations (MSS) and coordinate activities of mobile nodes within a cell, handling mobility and acting as a proxy for the system coordinator.
  • System Coordinator (SC): Manages the overall cluster, task distribution, and migration history. It is designed for fault tolerance through replication.

Datasets

The paper does not explicitly mention specific datasets used for training. However, for evaluation, they used a CT scan dataset of a human head with a partially removed skull to reveal the brain. This dataset consists of 84 slices of 128x128 samples, which are converted into a 256x256 pixel image for rendering. This dataset is used to demonstrate the performance of the image-rendering application.

Key Insights and Ideas

  • Harnessing Idle Power: The core insight is that significant computational power lies dormant in mobile devices, which can be effectively utilized by integrating them into cluster computing.
  • Dynamic Resource Normalization: The Horse Power Factor (HPF) is a crucial mechanism for abstracting the heterogeneity of computing resources, allowing for fair and efficient task allocation across diverse nodes.
  • Mobility as a Feature, Not a Bug: Moset's architecture, with its hierarchical coordination (MSS, CC, SC), is designed to gracefully handle the mobility of nodes, ensuring continuity of computation even when devices move between cells.
  • Anonymity for Security and Simplicity: Providing anonymity to applications and nodes simplifies the programming model and potentially enhances security by obscuring direct node identities.
  • Reliable Multicast is Key: The reliance on an "exactly-once reliable multicast protocol" is fundamental to ensuring data integrity and task completion in the face of unreliable wireless links.

Key Results

The paper presents a performance analysis using a distributed image-rendering application. Key findings include:

  • Super-linear Speedup: Under no-load conditions, super-linear speedups were observed for a small number of hosts (2-4) on both SPARC 333 MHz and 500 MHz machines. This is attributed to reduced page fault overheads when the dataset is more readily available in memory across multiple nodes.
  • Scalability Limitations: As the number of hosts increased (to 8 and 16), communication delay became dominant, limiting further speedup. This was partly due to the simulation of multicast by multiple unicasts on a single host.
  • Load Impact: The study showed that the effect of processor load (simulated by cpuhog) on execution time and speedup was marginal, with speedup remaining relatively constant.
  • Memory Impact: The effect of memory usage by cpuhog also had a marginal impact on execution time and speedup up to a certain threshold. Beyond 200MB, increased memory usage by cpuhog led to more page faults, allowing the image-rendering application more time to run, thus reducing execution time.
  • Super-linear Speedup with 4 Hosts: In one instance, with 4 hosts and significant memory usage by cpuhog, a super-linear speedup of 4.21 was observed.

Limitations and Future Work

Limitations:

  • Communication Overhead: The paper acknowledges that problems with a high communication-to-computation ratio might not be ideal for this model.
  • Multicast Simulation: The performance study was affected by the simulation of multicast by multiple unicasts on a single host, which introduced overhead. A distributed or physical multicast implementation would likely improve performance.
  • Non-preemptive Load Balancing: The load balancing mechanism is non-preemptive, meaning already executing tasks are not migrated, which could be a limitation for highly dynamic scenarios.

Future Work:

  • Inter-subtask Communication: Extending the model to support communication between subtasks is a key area for future research, especially for problems with high communication needs.
  • Mobile Grid Computing: The authors are exploring extending Moset to a mobile grid, which would involve a global collection of resources.
  • Generic Model Evolution: There is a need to evolve a generic model for cluster computing and communication for mobile systems.

Summary updated

BibTeX
@article{mohamed2005moset,
  author = {M. A. Maluk Mohamed and A. Vijay Srinivas and D. Janakiram},
  title = {{Moset: An anonymous remote mobile cluster computing paradigm}},
  journal = {J. Parallel Distributed Comput.},
  year = {2005},
  doi = {10.1016/j.jpdc.2005.04.012},
}
Figure