Intelligent strategies for DAG scheduling optimization in Grid environments

The paper presents a solution to the dynamic DAG scheduling problem in Grid environments. It presents a distributed, scalable, efficient and fault-tolerant algorithm for optimizing tasks assignment. The scheduler algorithm for tasks with dependencies uses a heuristic model to optimize the total cost of tasks execution. Also, a method based on genetic algorithms is proposed to optimize the procedure of resources assignment. The experiments used the MonALISA monitoring environment and its extensions. The results demonstrate very good behavior in comparison with other scheduling approaches for this kind of DAG scheduling algorithms.

💡 Research Summary

The paper addresses the challenging problem of dynamically scheduling tasks with dependencies (modeled as a Directed Acyclic Graph, DAG) in heterogeneous Grid environments. It proposes a two‑layer solution that combines a heuristic‑driven list scheduling algorithm with a cooperative genetic algorithm (GA) for resource assignment, all tightly integrated with a real‑time monitoring infrastructure (MonALISA).

In the first layer, the authors adopt the Cluster‑Ready Children First (CCF) algorithm, a dynamic list‑scheduling approach that processes the DAG in topological order. Each task is characterized by three metrics: t‑level (the weight of the longest path from a source node to the task), b‑level (the weight of the longest path from the task to an exit node), and ALAP (As Late As Possible), which quantifies how much a task’s start time can be delayed without increasing the overall makespan. These metrics are used to prioritize tasks in two priority queues: RUNNING‑QUEUE (tasks whose predecessors have completed) and CHILDREN‑QUEUE (tasks awaiting predecessor completion). The algorithm continuously dequeues the highest‑priority task, checks its readiness, and immediately attempts to assign a resource.

Resource assignment is performed by the second layer, a cooperative GA. Each cluster node maintains its own population of chromosomes, where a chromosome encodes a possible mapping of a batch of tasks to available resources. The fitness function incorporates task requirements (memory, CPU speed in MIPS, execution time) and communication costs derived from network latency and bandwidth (τ = latency + size/bandwidth). Nodes evolve their populations independently but exchange their best individuals after each generation (migration), ensuring rapid convergence toward a globally optimal solution. Chromosome length is fixed; if a batch contains fewer tasks, the algorithm waits a predefined time before starting the GA, padding the chromosome if necessary. Conversely, excess tasks are placed in a waiting queue for later scheduling. After a predetermined number of generations (e.g., 100), each node selects its optimal chromosome, broadcasts it to all peers, and adopts the consensus best solution.

The monitoring subsystem, based on MonALISA and its extensions, supplies up‑to‑date information on CPU load, memory availability, and network bandwidth for every Grid node. This data feeds both the CCF priority calculation (e.g., adjusting ALAP values) and the GA fitness evaluation, enabling the scheduler to react to dynamic changes such as node failures or load spikes.

Experimental evaluation uses a nine‑task DAG (the example illustrated in the paper) and a cluster of three processors. The authors compare three configurations: (1) a static resource‑assignment baseline, (2) CCF alone, and (3) CCF combined with the cooperative GA. Results show that the GA‑enhanced approach reduces the makespan by approximately 16 % relative to the static baseline, with notable improvements in tasks that have high inter‑task communication costs. The system also demonstrates fault tolerance (no single point of failure) and near‑linear scalability as the number of clusters increases.

Key contributions of the work are:

1. Integration of t‑level, b‑level, and ALAP metrics into a dynamic CCF scheduler for DAGs, providing a solid heuristic foundation for makespan minimization.
2. Introduction of a distributed, cooperative genetic algorithm for resource allocation that leverages migration of elite solutions to accelerate convergence.
3. Seamless coupling with a real‑time monitoring framework (MonALISA), allowing the scheduler to adapt to fluctuating Grid conditions.

Overall, the paper presents a comprehensive, scalable, and fault‑tolerant strategy for DAG scheduling in Grid environments, demonstrating measurable performance gains over traditional static or purely heuristic methods.