A Multi-Agent System Approach to Load-Balancing and Resource Allocation for Distributed Computing

In this research we use a decentralized computing approach to allocate and schedule tasks on a massively distributed grid. Using emergent properties of multi-agent systems, the algorithm dynamically creates and dissociates clusters to serve the changing resource demands of a global task queue. The algorithm is compared to a standard First-in First-out (FIFO) scheduling algorithm. Experiments done on a simulator show that the distributed resource allocation protocol (dRAP) algorithm outperforms the FIFO scheduling algorithm on time to empty queue, average waiting time and CPU utilization. Such a decentralized computing approach holds promise for massively distributed processing scenarios like SETI@home and Google MapReduce.

💡 Research Summary

The paper presents a decentralized, multi‑agent based resource allocation protocol (dRAP) designed for large‑scale distributed computing grids such as those used by SETI@home or Google MapReduce. Recognizing the scalability, robustness, and latency challenges inherent in centralized schedulers, the authors propose a system in which each computer is modeled as an autonomous agent that can form or dissolve clusters with physically proximate peers to satisfy the CPU requirements of tasks drawn from a global queue.

The problem is formalized by assuming a global task queue Q where each task declares its required number of CPUs (CPU_req) and the number of threads it can parallelize (TH_n). A “cluster” is defined as a set of agents whose combined CPUs exactly match a task’s CPU_req, and clusters are created dynamically based on local interactions without any central dispatcher. Four operational modes govern agent behavior: (1) idle agents scan Q and pick the task whose CPU_req is closest to a single CPU; (2) agents executing a task that still needs more CPUs request neighboring agents to join the cluster; (3) agents already in a cluster but without a task scan Q for the best‑fit task; (4) agents that finish a task leave their cluster and return to mode 1. This rule set mirrors cellular automata such as Conway’s Game of Life, allowing complex global behavior to emerge from simple local rules.

Complexity analysis shows that each time step requires traversing the global queue, yielding an O(n·m) cost where n is the average number of clusters and m is the queue length. The worst‑case cost remains O(n·m), which scales more favorably than a monolithic scheduler that would incur O(N) where N is the total number of nodes. To further reduce search overhead, the authors draw inspiration from the biological immune system and propose an “artificial lymph node” hierarchy. Each lymph node manages a sub‑set of clusters and a local task queue; the total cost becomes O(n²) (local search) + O(N/n) (inter‑node communication). Optimizing this expression leads to n = O(N¹ᐟ³), indicating that the number of clusters per lymph node should grow sub‑linearly with the overall system size, achieving near‑linear or sub‑linear scaling depending on the exponents of local and global communication costs.

Implementation is carried out using the MASON multi‑agent simulation toolkit (Java). Experiments involve 100 nodes and 1,000 randomly generated tasks (CPU_req ranging from 1 to 5, execution times from 25 to 125 time units). The dRAP protocol is benchmarked against a naïve first‑in‑first‑out (FIFO) scheduler. Results show a ~20 % reduction in total completion time (845.6 s vs. 1,071.2 s) and a ~25 % reduction in average waiting time (342.5 s vs. 475.3 s). Moreover, because dRAP assigns tasks such that the cluster’s CPU count exactly matches the task’s requirement, CPU utilization within clusters approaches 100 %, whereas FIFO often leaves excess CPUs idle.

The paper situates its contribution among prior work on grid resource allocation, economic and agreement‑based scheduling, and other multi‑agent applications (e.g., swarm robotics, distributed radar). It argues that the emergent, decentralized nature of dRAP offers superior scalability and fault tolerance. Limitations are acknowledged: the evaluation is simulation‑based, does not address heterogeneous resources (memory, I/O), and relies on a simple distance metric for “proximity.” Future work is outlined to include real‑world deployment, handling of task dependencies and priorities, richer resource models, and deeper exploration of the lymph‑node hierarchy with learning or memory mechanisms.

In summary, the dRAP protocol demonstrates that a locally‑driven, multi‑agent approach can effectively balance load and allocate resources in massive distributed systems, achieving measurable performance gains over traditional FIFO scheduling while offering a path toward scalable, resilient computing infrastructures.