Research Focus: Large-Scale Desktop Grids

The core research of the XtremWeb project centers on the design, optimization, and simulation of Large-Scale Distributed Systems (LSDS). By gathering the unused capacity (CPU idle cycles, volatile disk space, and network bandwidth) of heterogeneous desktop PCs and volunteer nodes, XtremWeb provides a highly scalable platform for High-Performance Computing (HPC) and distributed data storage.

1. General Architecture: Client-Coordinator-Worker

The XtremWeb platform is structured around a three-tier topology that ensures security, simplicity, and scalability in highly volatile environments:

    [ Client ] (Submits Tasks & Parameter Sweeps)
        │  ▲
        ▼  │ (Indirect transfer of data/results)
  [ Coordinator ] (Scheduler / Job Registry / Middle-Tier)
        ▲  │
        │  ▼ (Pulls jobs, updates sandbox parameters)
    [ Workers ] (PC Grids / Volatile LAN & WAN Nodes)

1. The Client: Submits compute tasks, inputs, and executable applications to the Coordinator. The Client and Server typically act as the unified command interface.
2. The Coordinator: Acts as the middle-tier scheduling server. It maintains a secure registry of jobs, parameter sets, and active worker nodes. It distributes work packages and aggregates incoming results.
3. The Worker: Lightweight daemons running on volunteer PCs, workstations, or servers. Workers periodically poll the Coordinator to retrieve pending jobs. Once execution completes, the Worker uploads the results back to the Coordinator.

Centralized P2P Computing Mode

XtremWeb can also run in a centralized Peer-to-Peer (P2P) mode. In this topology, any Worker node can temporarily behave as a Client to submit jobs directly to the Coordinator. These are scheduled dynamically across other available Workers. To prevent rogue scripts from compromising worker host security in this trustless design, our current research integrates sandboxing execution layers to isolate remote code execution.

2. Programming Models & Applications

The Grand-Large research team focuses on high-performance computing using two main programming models:

• Remote Procedure Calls (RPC): Used to execute applications concurrently across different parameters (parameter sweeps, bag of tasks, and workflows) without modifying the source code.
• Message Passing Interface (MPI): Configured to run SPMD (Single Program, Multiple Data) parallel algorithms across cooperative grid resources.

Beyond CPU-intensive tasks, XtremWeb projects leverage worker nodes to manage distributed disk space and stress-test network infrastructures. By coordinating volunteer nodes across different geographic regions, we simulate real-life workloads to analyze network bottlenecks and stress-test web servers.

3. Data Storage & Movement

Data transfer is a primary bottleneck in volunteer computing, as many jobs require large files for input parameters or generate massive outputs. We analyze and compare three primary transfer models:

Distributed Storage & Replication

Because participant nodes in a desktop grid are highly volatile, persistent storage requires active replication and redundancy. XtremWeb research investigates segmenting files into blocks and distributing them with redundant parity (resembling RAID-like properties) across volunteer nodes. We evaluate how cache affinity, replication rates, and persistence protocols affect overall job scheduling performance.

4. Asynchronous Scheduling & top-level Topologies

Traditional grid schedulers rely on a global state, which is not scalable in volunteer networks with thousands of nodes. Our research focuses on asynchronous, fully distributed scheduling algorithms, where nodes make local scheduling decisions based solely on their own capacity and communication with immediate topological neighbors.

Simulation via V-Grid

To validate these distributed scheduling algorithms, we developed V-Grid, a simulation framework built on top of the Swarm multi-agent simulator. V-Grid allows us to:

• Simulate the behavior of tens of thousands of volatile nodes.
• Analyze workload distribution in a 2D spatial coordinate system.
• Visualize the real-time dynamics of network topology and node failures in 3D.

Experimentation through simulation remains a crucial step for modeling and understanding the emergent behavior of large-scale decentralized systems.