A Survey on Deadline Constrained Workflow Scheduling Algorithms in Cloud Environment

Cloud Computing is the latest blooming technology in the era of Computer Science and Information Technology domain. There is an enormous pool of data centres, which are termed as Clouds where the services and associated data are being deployed and users need a constant Internet connection to access them. One of the highlights in Cloud is the delivering of applications or services in an on-demand environment. One of the most promising areas in Cloud scheduling is Scheduling of workflows which is intended to match the request of the user to the appropriate resources. There are several algorithms to automate the workflows in a way to satisfy the Quality of service (QoS) of the user among which deadline is considered as a major criterion, i.e. Satisfying the needs of the user with minimized cost and within the minimum stipulated time. This paper surveys various workflow scheduling algorithms that have a deadline as its criterion.

💡 Research Summary

Cloud computing has become the dominant paradigm for delivering on‑demand services through a vast pool of geographically distributed data centres. In this environment, users submit complex scientific or business workflows that consist of many inter‑dependent tasks. The scheduler’s challenge is to allocate virtual resources such that the entire workflow finishes before a user‑specified deadline while keeping the monetary cost as low as possible. The surveyed paper provides a comprehensive taxonomy of the algorithms that address this deadline‑constrained workflow scheduling problem, analyzes their design principles, and highlights the trade‑offs among performance, cost, and scalability.
The authors first formalize the problem as NP‑hard, which explains why exact optimal solutions are impractical for large‑scale, real‑time workloads. Consequently, three major families of heuristic solutions have emerged. The first family comprises list‑based heuristics such as classic List Scheduling, HEFT (Heterogeneous Earliest Finish Time) and its deadline‑aware extensions. These methods compute a priority order for tasks (often based on critical‑path length or a weighted combination of execution time and cost) and then greedily assign each task to the cheapest virtual machine (VM) that can meet its earliest‑finish requirement. Their strength lies in low computational overhead, making them suitable for very large workflows, but they cannot guarantee optimality.
The second family consists of meta‑heuristic approaches—Genetic Algorithms (GA), Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and similar population‑based techniques. By encoding a complete schedule as a chromosome, particle, or pheromone trail, these algorithms explore a much broader search space and can often find solutions that are close to the global optimum. They incorporate deadline constraints either as penalty terms in the fitness function or as hard feasibility checks during the generation of new individuals. While powerful, meta‑heuristics require careful parameter tuning (population size, mutation rate, inertia weight, etc.) and may suffer from slow convergence when the deadline is very tight.
The third family is constraint‑driven exact or near‑exact models such as Mixed‑Integer Linear Programming (MILP) and Constraint Programming (CP). In these formulations, start and finish times of each task are decision variables, and the precedence relations, resource capacities, and deadline are expressed as linear or logical constraints. Cost minimization is added as an objective function that accounts for on‑demand, spot, and reserved instance pricing. These models guarantee optimality for the given instance, but the number of variables grows with the product of tasks and candidate VM types, leading to exponential solving times for realistic workflow sizes.
Beyond algorithmic categories, the paper discusses resource allocation strategies. Exclusive (dedicated) allocation provides stronger deadline guarantees at higher cost, whereas shared allocation improves utilization but introduces performance variability. Horizontal scaling (adding more VMs) and vertical scaling (selecting more powerful VMs) are examined in relation to workflow characteristics—data‑intensive pipelines benefit from many modest VMs to parallelize I/O, while compute‑intensive kernels gain from fewer high‑CPU instances. The authors also analyze the impact of cloud pricing models. Spot instances, while cheap, can be reclaimed by the provider; therefore, many deadline‑aware algorithms incorporate checkpointing and slack‑time calculations to mitigate the risk of pre‑emption.
The survey identifies several research gaps. Most existing works assume static workflows known a priori, leaving a shortage of adaptive schedulers that react to dynamic task arrivals or runtime performance feedback. Multi‑deadline scenarios, where different sub‑workflows have distinct completion targets, are rarely addressed. Energy and carbon‑footprint considerations are largely absent, despite growing interest in green cloud computing. Finally, realistic network constraints—bandwidth limits, latency spikes, and data‑transfer costs—are often simplified or ignored, which can lead to overly optimistic deadline predictions.
To advance the state of the art, the authors suggest integrating reinforcement‑learning agents that continuously learn optimal placement policies, developing multi‑objective optimization frameworks that jointly handle time, cost, and energy, and constructing hybrid models that combine fast heuristics for initial placement with exact solvers for fine‑tuning. Such directions would enable cloud providers and users to meet stringent service‑level agreements while keeping operational expenditures under control.
In summary, the paper offers a detailed, well‑structured overview of deadline‑constrained workflow scheduling algorithms, compares their computational complexity, scalability, and suitability for different cloud pricing schemes, and points out open challenges that future research must tackle to achieve truly efficient, deadline‑guaranteed cloud execution.