Learning based Scalable Multi-Objective Scheduling of Workflows across UD-IaaS-FaaS Systems

arXiv:2610.15886v1 [cs.DC] 11 Oct 2026

Learning based Scalable Multi-Objective Scheduling of Workflows across UD-IaaS-FaaS Systems G Vasantha Reddy

Ram Bansal∗

Krashna ∗

IIT Guwahati Guwahati, India r.vasantha@iitg.ac.in

IIT Guwahati Guwahati, India ram.bansal@iitg.ac.in

IIT Guwahati Guwahati, India krashna@iitg.ac.in

Suvarthi Sarkar

Aryabartta Sahu

University of Western Australia India Mumbai, India suvarthi.sarkar@india.uwa.edu.au

IIT Guwahati Guwahati, India asahu@iitg.ac.in

Abstract

1

Workflow scheduling in user device-centric edge–cloud computing has emerged as a challenging problem due to the need to jointly optimize execution performance, energy consumption, and monetary cost. Modern workflows consist of heterogeneous tasks, where compute-intensive components are well-suited for Infrastructureas-a-Service (IaaS) resources, while fine-grained, event-driven tasks are more efficiently executed using the Function-as-a-Service (FaaS) model. We formulate this scheduling problem for application workflows over a hybrid User Device (UD)-IaaS-FaaS computing continuum as a multi-objective Mixed-Integer Linear Program (MILP). The formulation models the workflow execution mechanism by incorporating cumulative capacity constraints for IaaS resources and elastic resource provisioning for FaaS. To overcome the intractability of MILP, we propose a learning-based framework that combines graph neural networks, surrogate performance modeling, and a multi-sample REINFORCE approach. The proposed approach is workflow-aware, leveraging task pairing to improve data locality and structured task placement decisions. Agent training is organized into three stages: small-scale learning of UD-IaaS trade-offs, intermediate locality aware co-location, and large-scale handling of parallelism and aggregation bottlenecks in workflows. The proposed framework is validated through real-world experiments using the Pegasus Montage workflow on a custom UD-IaaS–FaaS testbed comprising a user device, multiple IaaS nodes, and serverless FaaS instances. Extensive experimental results demonstrate the improved trade-offs across makespan, energy consumption, and cost. Compared with the genetic algorithm and greedy scheduling approaches, the proposed approach reduces makespan by 10-17.5%, energy consumption by 3-9%, and cost by 4-10%, while remaining within 3.5-5.6% of the MILP optimum.

In the domain of edge-cloud computing, workflow scheduling is a critical and emerging research problem driven by the need to balance performance, energy consumption, and cost. The hybrid edge-cloud is a well-known model that strategically combines the extensive, scalable power with local private resources of cloud services in the public domain [1]. Contemporary applications are progressively built with intricate parts that are appropriate for various Infrastructure-as-a-Service (IaaS) virtual machines (VMs) that are frequently used for large, monolithic, or persistent tasks (on VMs), whereas Function-as-a-Service (FaaS) platforms are used to offload fine-grained, event-driven, or transient functions [2]. Resource optimization is possible with this hybrid IaaS-FaaS approach. For example, by running smaller jobs on FaaS and long running tasks on IaaS, the system saves more IaaS resources for computationally demanding tasks. This hybrid IaaS-FaaS approach enables optimal resource utilization. Notable platforms providing these services including Amazon EC2 (IaaS) and AWS Lambda [3]. An user device (UD) centric framework extends the hybrid model by incorporating the user’s own hardware as both a local compute resource and the central workflow orchestrator [4]. In this architecture, the UD coordinates task dispatch, manages resource provisioning, and controls data flow across the distributed cloud environment. Given its central role, optimizing task offloading and scheduling decisions for the UD has become a key research challenge. This UD centric model is distinguished from traditional cloudonly environments by a heterogeneous collaboration between untrusted public resources, trusted long term hired IaaS servers (managed by users) and a trusted private device (the UD) [5–7]. A complex trade-off between application completion time (makespan), user device energy consumption, and the financial costs to users of cloud services can result from using different offloading strategies, which can produce wildly disparate performance outcomes. But in user device centric model adds a number of special and noteworthy complications. Schedulers need to concern with the limited, shared CPU capacity on IaaS VMs, where multiple tasks may run concurrently, requiring cumulative resource constraint modeling rather than simple one job per node allocation. These factors collectively make the function offloading and scheduling problem in this environment significantly more complex than in standard cloud models. The IaaS servers are provisioned for longer durations and remain under user control, making them comparatively more trustworthy.

CCS Concepts • Computer systems organization → Cloud computing; • Software and its engineering → Allocation and scheduling.

Keywords Reinforcement Learning, IaaS, FaaS, DAG, Cloud, Scheduling ∗ These authors contributed equally to this work.

1

Introduction

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

G Vasantha Reddy, Ram Bansal, Krashna, Suvarthi Sarkar, and Aryabartta Sahu

In contrast, FaaS resources are ephemeral and managed by third party service providers, and are therefore generally treated as unreliable. Consequently, temporary input-output data can be safely retained on IaaS servers to exploit data locality and reduce transfer overhead. However, such data persistence is not available on FaaS platforms. Instead, interactions with FaaS resources should follow a two-step data transfer mechanism, where data is securely staged through trusted storage before and after execution. To develop an optimal offloading strategy, extensive research has investigated workflow scheduling in cloud and edge environments. However, most existing works focus on optimizing a single performance objective, such as minimizing makespan or minimizing energy consumption or cost. This is often not enough because users usually have to balance all three competing goals based on their own needs. For example, users might want to finish as quickly as possible for latency sensitive tasks or reduce cost and energy when they are on a limited budget. In such cases, single-objective optimization methods are inadequate, underscoring the need for multi-objective function offloading approaches. Multi-objective optimization approaches are well-suited for decision making scenarios involving multiple conflicting objectives. A common and effective technique is to scalarize the disparate objectives (e.g., time, energy, and cost) into a single objective function using the weighted sum method. While formulating this scheduling problem as a Mixed-Integer Linear Program (MILP) provides a formal path to a true optimum, the underlying problem is known to be NP-hard. The specific complexities of this model particularly the pseudo-polynomial nature of the discrete-time formulation, the cumulative IaaS resource constraints, and FaaS resources make finding an exact solution computationally intractable for large-scale, real world applications. This highlights the urgent need for research into high performance heuristic and metaheuristic solutions tailored to this specific problem. In this paper, we propose a multi-objective optimization approach to address the hybrid UD-IaaS-FaaS workflow offloading problem. We model the application as a Directed Acyclic Graph (DAG) [9, 10] and focus on optimizing application completion time, UD energy consumption, and user cost. Because these goals are incommensurable (measured in dollars, joules, and seconds), a normalized weighted-sum approach is used to combine them, which is essential to guaranteeing that user preferences are accurately reflected. The primary contributions of this work are summarized as follows:

• We develop a custom UD–IaaS–FaaS testbed comprising a user device, multiple IaaS nodes, and serverless FaaS instances to validate the proposed framework, and validated it using the Pegasus Montage workflow. Extensive experiments demonstrate the effectiveness of the proposed approach by achieving improved trade-offs among makespan, energy consumption, and execution cost.

2

Literature Review

Research in hybrid edge-cloud and serverless computing has extensively investigated trade-offs among performance, energy efficiency, and cost across multiple complementary directions. Significant research has focused on reducing energy consumption in IaaS data centers through VM consolidation and predictive scheduling techniques. Zakarya et al., [11] extends backfilling approach to cloud systems by utilizing idle host intervals for lower-priority VMs, thereby improving resource utilization and reducing energy usage. However, these approaches typically assume homogeneous infrastructure and stable request arrivals, limiting their applicability in heterogeneous, multi-tier systems with DAG-based dependencies. Similarly, queueing theoretic methods, such as Anselmi et al., [12], aim to minimize latency through balanced server grouping, achieving near-zero wait times without requiring job-size data. Despite their efficiency, these models rely on assumptions of homogeneous and static environments and do not account for task dependencies or inter-node communication, making them less suitable for hybrid computing scenarios. Hybrid orchestration frameworks, such as Sheshadri et al., [13], integrates IaaS and FaaS by assigning baseline workflows to virtual machines and handles burst demand through serverless platforms, often leveraging machine learning for improved latency and cost minimization. However, these approaches are typically limited to service function chains, and lacks explicit multi-objective optimization, and do not support dynamic task placement across multiple computing tiers. Federated learning frameworks, such as Zhao et al., in [14], address data and system heterogeneity through adaptive aggregation, improves model accuracy and convergence. However, they primarily focus on learning performance and do not incorporate resource scheduling or system-level optimization, leaving energy and communication trade-offs largely unaddressed. Similarly, workload placement approaches for edge-cloud systems, such as Wang et al., [1], employs optimization techniques to minimize cost and latency through joint VM placement and routing. These approaches assumes static environments and rely on offline optimization, limiting their adaptability to dynamic conditions and only partially accounting for energy efficiency. Yang et al., in [15] proposes multi objective RL approach for serverless offloading, balancing latency, cost, and energy. While effective for function-level scheduling, these methods do not address end-to-end workflow scheduling across heterogeneous computing resources. Wu et al. [16] propose a Double Deep Q-Network (DDQN) based approach to mitigate cold starts and improve QoS by jointly optimizing execution time, energy consumption, and monetary cost in serverless edge computing. However, the approach is limited to scheduling independent tasks and does not consider workflow applications with task dependencies.

• We formulate workflow scheduling across user devices, hybrid IaaS and FaaS computing continuum as a linearly combined multi-objective optimization problem, modeled as a Mixed Integer Linear Program (MILP). • We propose a learning based framework that integrates Graph Neural Networks (GNNs), surrogate performance modeling and a multi-sample REINFORCE (MSR) based reinforcement learning approach to learn high quality placement strategies efficiently. • We propose a workflow-structure aware scheduling method that exploits task pairing to improve data locality and scalability across heterogeneous workflow sizes. 2

Learning based Scalable Multi-Objective Scheduling of Workflows across UD-IaaS-FaaS Systems

3.2

User Devices (UDs)

IaaS

VM

VM

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

Workflow Model

An application workflow is modeled as a DAG 𝐺 = (V, E), where 𝑉 is the set of tasks (nodes), and E ⊆ V×V is the set of directed edges, represents the dependencies among tasks. An edge (𝑡𝑖 , 𝑡 𝑗 ) ∈ E denotes that task 𝑡 𝑗 depends on task 𝑡𝑖 . A workflow task 𝑡 𝑗 is eligible for execution only after all its predecessor tasks have completed thereby satisfying its data dependencies. The system is modeled using a discrete-time formulation, where the time horizon 𝐻 is partitioned into a set of equal length time slots T , each of duration Δ. This approach is often more computationally tractable for problems with complex, cumulative resource constraints compared to continuous time models. However, the choice of Δ represents a fundamental trade-off. A smaller Δ provides a more precise schedule (approaching a continuous time solution) but exponentially increases the size of the model and computational complexity, potentially making it intractable to solve.

FaaS

VM 3.3

Figure 1: System Model of Hybrid UD-IaaS-FaaS Computing Continuum

Constants (derived from parameters)

These constants are derived to translate the continuous physical parameters (cycles, bytes, seconds) into the integer-based “slot” domain required by the discrete-time MILP formulation. We discretize execution durations into integer slot counts. For a slot duration Δ (seconds):    cold    𝑡 + 𝑐𝑖 /𝜔𝑛 𝑐𝑖 /𝑓𝑘 𝑐𝑖 /𝜔𝑢 𝐼 𝐹 ; 𝐷𝑖,𝑚,𝑘 ; 𝐷𝑖,𝑛 = 𝑛 . (1) 𝐷𝑖𝑈 = = Δ Δ Δ

Overall, prior work has advanced energy-aware scheduling, hybrid orchestration, federated learning, and edge-cloud placement [1, 11–17]. However, most of these approaches are limited to static or single tier settings and do not jointly optimize latency, energy, and cost across hybrid infrastructures. This work addresses these gaps with MSR-GNN, a learning-based framework that jointly optimizes makespan, energy, and cost for DAG workflows across UD, IaaS, and FaaS resources.

where 𝐷𝑖𝑈 denotes the execution duration of task 𝑖 on the user 𝐼 device, 𝐷𝑖,𝑚,𝑘 denotes its execution duration on IaaS node 𝑚 with 𝐹 denotes its execution duration on FaaS resource tier 𝑘, and 𝐷𝑖,𝑛 instance 𝑛. The FaaS execution duration explicitly includes the cold-start delay 𝑡𝑛cold , thereby accounting for the additional slot occupancy incurred during function initialization.

3 System Model and Problem Statement 3.1 System Model

3.4

Figure 1 illustrates the proposed hybrid IaaS–FaaS system model, comprising a heterogeneous computing architecture with User Devices (UDs), Infrastructure-as-a-Service (IaaS), and Function-as-aService (FaaS) resources. The IaaS layer consists of multiple virtual machines (VMs) managed by a cloud service provider, whereas the FaaS layer offers containerized execution environments with heterogeneous resource configurations hosted in the cloud. User applications are modeled as Directed Acyclic Graph (DAG) workflows, where tasks are scheduled across the available UD, IaaS, and FaaS resources. In this architecture, UDs serve as a dual role and act both as local compute resources capable of executing tasks and also, as a central orchestrator responsible for coordinating task scheduling and resource allocation across the hybrid distributed system. Upon receiving workflow tasks, the UD scheduler determines placement of tasks based on workflow requirements and system resources. Typically, the IaaS resources are utilized for compute intensive and long running tasks (as IaaS servers are hired for long term basis and managed by the users), whereas FaaS platform is suitable for lightweight and short-term executions (completely outsource as simple task execution). This hybrid design enables efficient workload distribution across tiers, thereby improving overall system performance and resource utilization.

Decision Variables

Binary placement variables are defined as follows: 𝑥𝑖𝑈 ∈ {0, 1}, 𝐼 where 𝑥𝑖𝑈 = 1 if task 𝑖 is executed on the user device; 𝑥𝑖,𝑚 ∈ {0, 1}, 𝐼 𝐹 ∈ the value of 𝑥𝑖,𝑚 = 1 if task executes on IaaS server 𝑚; and 𝑥𝑖,𝑛 𝐹 = 1 if task executes on FaaS endpoint 𝑛. The {0, 1}, the value 𝑥𝑖,𝑛 scheduling (time-slot) variables are defined as follows. • Start indicator: 𝑢𝑖,𝑡 ∈ {0, 1}: 1 if the task 𝑖 starts at slot 𝑡. It is a time-indexed variable, and it creates a binary variable for each possible start time 𝑡 for each task 𝑖. Í • Start slot integer: 𝑠𝑖 ∈ {0, . . . , 𝐻 − 1} such that 𝑠𝑖 = 𝑡 𝑡 · 𝑢𝑖,𝑡 . • Duration in slots (chosen): 𝐷𝑖choice ∈ {0, . . . , 𝐻 } (integer). • Finish slot: 𝐹𝑖 ∈ {0, . . . , 𝐻 } (integer) with 𝐹𝑖 = 𝑠𝑖 + 𝐷𝑖choice .

3.5

Constraints

Constraints are the rules and limits that make up the “feasible set” of solutions. All of the following algebraic constraints must be met by any valid schedule. Placement exclusivity: Each task must be placed/scheduled at exactly one location either on the UD, one IaaS server, or one FaaS endpoint, and this can be expressed as ∑︁ ∑︁ 𝐼 𝐹 𝑥𝑖𝑈 + 𝑥𝑖,𝑚 + 𝑥𝑖,𝑛 = 1, ∀𝑖. (2) 𝑚∈ S 𝐼

3

𝑛∈ S 𝐹

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

G Vasantha Reddy, Ram Bansal, Krashna, Suvarthi Sarkar, and Aryabartta Sahu

IaaS tier linking: If a task is assigned to IaaS server 𝑚, exactly one tier must be selected on that server: 𝜅 ∑︁ 𝐼 𝑦𝑖,𝑚,𝑘 = 𝑥𝑖,𝑚 , ∀𝑖, 𝑚. (3)

Precedence constraint: This is the set of rules that makes sure both the logical dependencies of the workflow and data flow are followed: 𝑡 𝑠𝑗 ≥ 𝑡 𝑗 −1 + 𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡 𝑗 −1,𝑙𝑜𝑐 · Ψ𝑗 −1,𝑙𝑜𝑐 𝑓

𝑘=1 𝐼 to the speThis equation links the general placement variable 𝑥𝑖,𝑚 cific tier selection 𝑦𝑖,𝑚,𝑘 . If a task is not placed on server 𝑚 (LHS=0),

the data transfer time required to transmit the predecessor’s output from its execution location to the current execution node. Memory feasibility: A workflow task can only be assigned to a resource node with sufficient available memory.

𝑛

𝑚,𝑘

Since the exclusivity constraint in Equation 2 requires the RHS to equal 1, the model selects exactly one start time 𝑡 for each task, thereby enforcing 𝑢𝑖,𝑡 = 1 for a single time slot. Start slot definition: The start time 𝑠𝑖 is determined from the binary start indicators as: ∑︁ 𝑠𝑖 = 𝑡 · 𝑢𝑖,𝑡 , ∀𝑖. (5) 𝑡∈T

𝑚∈ S 𝐼 𝑘=1

3.6

𝑛∈ S 𝐹

∀𝑖.

𝑖

𝑘=1

𝑓𝑘 ·

𝑢𝑖,𝜏 · 𝑦𝑖,𝑚,𝑘 ≤ Ω𝑚 ,

(10)

∀𝑖, 𝑚.

(11)

Performance Models

3.6.1 Workflow Completion Time (Makespan). The workflow execution time is modeled using a continuous variable 𝑇app , constrained such that 𝑇app ≥ 𝐹𝑖 · Δ for all workflow tasks 𝑖, where execution time depends on the chosen placement. While discretized execution durations 𝐷 · are used for scheduling decisions, continuous formulations are applied for computing energy and monetary costs. The up upload and data transfer time for each task 𝑖 is given by 𝑇𝑖 = Δ · 𝐿𝑖 , where 𝐿𝑖 represents the selected communication latency.

(7)

IaaS per-slot CPU capacity: For each IaaS server 𝑚 and time slot 𝑡, the total CPU demand of all tasks executing during slot 𝑡 must not exceed the server capacity Ω𝑚 . A task assigned to tier 𝑘 contributes 𝑓𝑘 CPU cycles/sec over its execution interval, which is captured by summing its start time indicators across the corresponding active time slots. The resulting linear constraint is as follows: 𝑡 ∑︁

∀𝑖,

𝐼 𝑥𝑖,𝑚 · mem𝑖 ≤ Capmem𝑚 ,

The performance models define how the primary objectives time, energy, and cost are computed from the solver’s decision variables. While scheduling constraints (e.g., Equation 8) rely on discretized slot durations 𝐷 · , the objective functions use continuous time formulations for higher accuracy (e.g., 𝑐𝑖 /𝜔𝑢 in seconds instead of Δ · 𝐷 𝑈𝑤 ).

where the placement variables are binary and the execution durations 𝐷 · are constants, 𝐷𝑖choice equals to the execution duration of the selected resource and is therefore an integer. Finish time: Finish slot equals start plus execution duration: 𝐹𝑖 = 𝑠𝑖 + 𝐷𝑖choice,

𝑥𝑖𝑈 · mem𝑖 ≤ Capmem𝑢 ,

FaaS memory feasibility is enforced by permitting task placement only on FaaS instances that satisfy the required memory capacity of the current task. Start/finish bounds: All 𝑠𝑖 , 𝐹𝑖 , 𝑡 ud_ready, 𝑡 ud_allready are within [0, 𝐻 ] and integer. For the workflow, we model this by introducing a continuous variable 𝑇app for workflow completion time and constraints 𝑇app ≥ 𝐹𝑖 .Δ for all tasks 𝑖.

Since the exclusivity constraint in (4) ensures that exactly one 𝑢𝑖,𝑡 equals 1, the above expression uniquely assigns 𝑠𝑖 to the corresponding start time slot. Execution duration selection (location-dependent): The execution duration is determined by the selected execution resource node: 𝜅 ∑︁ ∑︁ ∑︁ 𝐼 𝐹 𝐹 𝐷𝑖choice = 𝑥𝑖𝑈 𝐷𝑖𝑈 + 𝑦𝑖,𝑚,𝑘 𝐷𝑖,𝑚,𝑘 + 𝑥𝑖,𝑛 𝐷𝑖,𝑛 ∀𝑖 (6)

𝜅 ∑︁ ∑︁

(9)

𝑓 where 𝑡 𝑠𝑗 denotes the start time of task 𝑡 𝑗 , 𝑡 𝑗 −1 denotes the finish 𝑝𝑟𝑒𝑑 time of its predecessor, and 𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡 𝑗 −1,𝑙𝑜𝑐 · Ψ𝑗 −1,𝑙𝑜𝑐 represents

it cannot select a tier (RHS=0). If it is (LHS=1), it must select exactly one tier (RHS=1). Start indicator consistency (one start if scheduled): Exactly one start slot indicator must exist if and only if the task is scheduled: ∑︁ ∑︁ ∑︁ 𝐹 𝑢𝑖,𝑡 = 𝑥𝑖𝑈 + 𝑦𝑖,𝑚,𝑘 + 𝑥𝑖,𝑛 , ∀𝑖. (4) 𝑡∈T

𝑝𝑟𝑒𝑑

3.6.2 Energy Model. Energy is modeled as conservative linear energy accounting for both local and remote resources per task 𝑖: ∀𝑚, 𝑡 .

(8)

𝐼 𝜏=max(0,𝑡 −𝐷𝑖,𝑚,𝑘 +1)

comp

𝐸𝑖loc = 𝑥𝑖𝑈 · 𝑝𝑢

This constraint is a cumulative resource constraint, formulated using a “time-bucket” or “sliding window” approach. For every server 𝑚 and every time slot 𝑡, it calculates the total CPU load. The Í𝑡 𝑢𝑖,𝜏 checks if task 𝑖 started at a time 𝜏 such that it is inner sum 𝜏=... still running during slot 𝑡. If so, the CPU demand 𝑓𝑘 is added to the total resource load and this total load must not exceed the server’s capacity Ω𝑚 . This correctly models a shared cloud resource, which is more complex than a simple IaaS model where one job occupies one node.

·

𝑐𝑖 , 𝜔𝑢

  up pred 𝐼 𝐹 𝐸𝑖remote = (𝑥𝑖,· + 𝑥𝑖,· ) · 𝑝𝑢tx · 𝑇𝑖 + 𝑝𝑢idle · (𝑇𝑖 + 𝑇𝑖exec ) ,

(12)

(13)

pred

where 𝑇𝑖 denotes the waiting time at the UD for predecessor outputs. The UD energy model considers three power states: comp computation (𝑝𝑢 ), transmission (𝑝𝑢tx ), and idle waiting (𝑝𝑢idle ) [21]. Consequently, local execution incurs computation energy, whereas remote execution incurs transmission and idle energy. To preserve this linearity, the remote energy terms are linearized using auxiliary variables and constraints that are activated only when the corresponding placement variable is selected. 4

Learning based Scalable Multi-Objective Scheduling of Workflows across UD-IaaS-FaaS Systems

𝐸𝑡𝑜𝑡 =

∑︁

𝐸𝑖loc + 𝐸𝑖remote )

the Policy Evaluation Cost (PEC) including time, energy and cost to provide immediate reward feedback. The proposed framework integrates three key components as shown below.

(14)

𝑖

where the total workflow energy consumption 𝐸 is computed as the sum of the energy consumed by local and remote executions.

(1) GNN-based Policy Model (𝜋𝜃 ): GNNs uses message passing mechanisms to encode the DAG topology by capturing task dependencies and system constraints, enabling effective node to resource placement decisions beyond purely local features. (2) GNN-based Surrogate Model (𝑓𝜙 ): We use a supervised, differentiable GNN predicts the PEC including time, energy and cost for node placement, thereby replacing computationally expensive simulations with rapid and accurate estimations. (3) Policy Gradient Optimization: The multi-sample REINFORCE (MSR) approach iteratively refines the policy parameters 𝜃 using the rewards derived from the surrogate model (for improvement), guiding the agent toward good scheduling decisions within the large combinatorial action space.

3.6.3 Monetary Cost Model. Monetary cost of each user workflow is modeled as follows:   ∑︁ ∑︁ ∑︁ ∑︁ 𝑐𝑖 𝐼 𝐼 𝐼 + 𝐶𝑡𝑜𝑡 = 𝑦𝑖,𝑚,𝑘 𝑐 cpu,𝑚 𝑐 vm,𝑚 𝑥𝑖,𝑚 · Δ · 𝐷𝑖choice 𝑓𝑘 𝑖 ∈𝑉 𝑚∈ S 𝐼 𝑘 𝑚∈ S 𝐼 ! ∑︁ 𝐹 𝐹 𝐹 exec
+ 𝑥𝑖,𝑛 𝑐 inv,𝑛 + 𝑐 mem,𝑛 mem𝑖 𝑇𝑖 . (15) 𝑛∈ S 𝐹

where, the IaaS cost (term 2) is computed as the product of the per𝐼 second VM instance price 𝑐 vm,𝑚 and the execution duration of the task. The FaaS cost (term 3) consists of a fixed per-invocation charge 𝐹 𝑐 inv,𝑛 , and resource consumption cost given by the memory pricing 𝐹 rate 𝑐 mem,𝑛 (e.g., cost per MB-second), multiplied by the allocated memory and the precise execution time of the task [22, 23]. The up network transfer money costs are added by including 𝑐 tx ·𝑇𝑖 terms.

3.7

4.1

Objective Function

4.1.1 State and Feature Encoding. A global state is represented by the state vector 𝑠 = [𝑁 , 𝑛𝑖𝑎𝑎𝑠 , 𝑤𝑇 , 𝑤 𝐸 , 𝑤𝐶 ], which encodes the workflow jobs (𝑁 , i.e., number of application workflows), available IaaS resources (𝑛𝑖𝑎𝑎𝑠 ) and the user defined optimization weights for Í makespan, energy, and cost, subject to 𝑤𝑖 = 1. At the graph level, the workflow is modeled as a DAG 𝐺 = (𝑉 , 𝐸), where each task 𝑣𝑖 ∈ 𝑉 is associated with a feature vector 𝑥𝑖 = [type𝑖 , 𝑑𝑖 , 𝑒𝑖 , 𝑁 ] in R 6 . This feature vector uses one-hot encoding to represent the task type based on DAG nodes, for example (a) Type A for parallel entry level tasks, (b) Type B for intermediate processing and filtering and (c) Type C for fan-in tasks. Additionally with its normalized topological depth 𝑑𝑖 , hardware specific assignment constraints 𝑒𝑖 , and the workflow jobs size 𝑁 .

MSR-GNN: GNN based multi-sample REINFORCE Approach

Optimizing workflow task schedule and placement in a heterogeneous compute continuum is a high dimensional NP-hard problem, where workflow task dependencies, data-flow constraints, and resource heterogeneity must be jointly considered to minimize makespan time, energy, and cost (TEC). Although conventional RL approaches such as A3C or DQN offer lower variance, they are unsuitable for this setting due to dynamic workflow sizes and high simulation overhead. To address these challenges, a learning based GNN-driven multi-sample REINFORCE (GNN-MSR) framework is proposed as the core optimization engine, enabling stable policy learning in large combinatorial action spaces by reducing variance through multi-sample estimation. The end-to-end operational pipeline is defined as follows: 𝜋𝜃

𝑓𝜙

4.1.2 Constraint aware Action Space. To prevent the REINFORCE agent from selecting invalid or infeasible mappings, a constraint validation layer is implemented as an action mask 𝑀𝑖,𝑗 that dynamically prunes the action space by enforcing three key rules based on infrastructure state 𝑠 and the workflow metadata: (1) IaaS Dynamic Availability: Since only a subset of IaaS nodes may be active at a given time, the action mask excludes devices with index 𝑗 > 𝑛𝑖𝑎𝑎𝑠 using the global state variable 𝑛𝑖𝑎𝑎𝑠 , ensuring that task placement decisions consider only currently available resources. (2) Data Sovereignty and Ingress/Egress Pinning: Certain tasks responsible for data ingestion and final aggregation are constrained to execute on user devices to avoid high-latency WAN transfers. Specifically, nodes are pinned by enforcing the action mask. (3) Task Pairing for Data Locality: In the DAG workflows, certain tasks form strict producer-consumer pairs and must be colocated to avoid data transfer overhead.

𝑅𝐿

d −−→ 𝜃 (𝐺, 𝑠) −−→ 𝐴 −−→ 𝑃𝐸𝐶

State and Workflow Representation

The workflow scheduling environment is formulated as a Markov Decision Process (MDP) that captures both the static workflow topology and dynamic infrastructural system conditions.

  𝑇app 𝐸 tot 𝐶 tot min TEC = min 𝑤𝑇 + 𝑤𝐸 + 𝑤𝐶 (16) 𝑇ref 𝐸 ref 𝐶 ref where the Equation 16 represents the objective function formulated as a weighted combination of execution time (𝑇𝑎𝑝𝑝 ), total energy consumption (𝐸 tot ), and total cost (𝐶 tot ). The parameters Tref , 𝐸 ref , and 𝐶 ref are normalization factors used to scale time, energy, and cost into a common dimension. The coefficients 𝑤𝑇 , 𝑤 𝐸 , and 𝑤𝐶 denote the relative weights assigned to time, energy, and cost respectively and these coefficients are tunable based on user requirements, subject to the constraint (𝑤𝑇 + 𝑤 𝐸 + 𝑤𝐶 = 1).

4

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

(17)

where 𝐺 denotes the workflow DAG and 𝑠 represents the global system state (e.g., number of active IaaS/FaaS nodes and optimization weights). The GNN-based policy processes (𝐺, 𝑠) to generate a task assignment 𝐴, while the surrogate GNN rapidly estimates 5

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

G Vasantha Reddy, Ram Bansal, Krashna, Suvarthi Sarkar, and Aryabartta Sahu

Algorithm 1 GNN Surrogate Training

For any compute device 𝑗 ∈ 𝐷 that is invalid for task 𝑣𝑖 , the ′ corresponding preference logit 𝑧𝑖,𝑗 is assigned with −∞, ensuring zero selection probability after Softmax normalization.

1: procedure TrainSurrogate(𝐺, 𝑁 samples , 𝑁 epochs )

16:

Initialize Dtrain ← ∅ for each workflow size 𝑁 img do Load DAG 𝐺 = (V, E) corresponding to 𝑁 img for 𝜅 = 1 to 𝑁 samples do Generate random valid assignment 𝐴 (𝜅 ) Execute simulator to obtain (𝑇𝜅 , 𝐸𝜅 , 𝐶𝜅 ) Compute 𝑃𝐸𝐶𝜅 = 𝑤𝑇 ∗ 𝑇𝜅 + 𝑤 𝐸 ∗ 𝐸𝜅 + 𝑤𝐶 ∗ 𝐶𝜅 Dtrain ← Dtrain ∪ {(𝐺, 𝐴 (𝜅 ) , 𝑠, 𝑃𝐸𝐶𝜅 )} Initialize model parameters 𝜙 for epoch = 1 to 𝑁 epochs do for each sample (𝐺, 𝐴, 𝑠, 𝑃𝐸𝐶) ∈ Dtrain do 𝑋 ← BuildNodeFeatures(𝐺, 𝐴, 𝑠) ⊲ Feature build d ← 𝑓𝜙 (𝑋, E, 𝑠) 𝑃𝐸𝐶 ⊲ Forward pass d 2 L ← (𝑃𝐸𝐶 − 𝑃𝐸𝐶) ⊲ Loss computation 𝜙 ← 𝜙 − 𝜂∇𝜙 L ⊲ Parameter update

17:

return trained surrogate model 𝑓𝜙

4.3

Optimization via Multi-Sample REINFORCE

2: 3:

4.2

Dual-GNN Architecture

4: 5:

The proposed framework employs two complementary GNN models that leverages the relational inductive bias of the workflow DAG, enabling scalable generalization across varying workflow sizes without the need for structural retraining.

6: 7: 8: 9:

4.2.1 GNN Policy Network (𝜋𝜃 ). The GNN policy network encodes the workflow topology into actionable placement probabilities using two sequential Graph Convolutional Network (GCN) layers (hidden dimension = 64), enabling each task to aggregate structural information from its multi-hop neighborhood:

10: 11: 12: 13: 14:

ℎ𝑖(𝑙+1) = 𝜎

∑︁

1

(𝑙 )

𝑊 (𝑙 ) ℎ 𝑗 √︃ ˜ ˜ 𝑗 ∈ N (𝑖 )∪𝑖 𝑑𝑖 𝑑 𝑗

15:

(18)

where ℎ𝑖(𝑙 ) denotes the embedding of node 𝑖 at layer 𝑙, 𝑊 (𝑙 ) is the corresponding learnable weight matrix, 𝑑˜𝑖 is the degree of node 𝑖 including the self-loop, and 𝜎 (·) denotes the nonlinear activation function. To incorporate global contextual information, the final node embedding is concatenated with the global state vector, yielding ℎ˜𝑖 = [ℎ𝑖(𝑙 ) ||𝑠]. The fused representation is then passed through a multi-layer perceptron (MLP) readout network to produce logits 𝑧𝑖 , which are subsequently masked and normalized using the softmax function to obtain the placement policy 𝜋𝜃 (𝐴|𝑠, 𝐺).

The core objective of the solution is to minimize the multi-objective policy evaluation cost over a large combinatorial action space. To mitigate the high gradient variance inherent in standard REINFORCE, a multi-sample estimation strategy inspired by Group Relative Policy Optimization (GRPO) is employed. 4.3.1 Group Relative Advantage Scaling. For a given state 𝑠 and DAG 𝐺, the policy generates 𝜅 independent assignment trajectories 𝐴(1), . . . , 𝐴(𝜅). Each assignment is evaluated by the surrogate š𝜅 . Instead model to obtain a PEC, defining the reward as 𝑟𝜅 = −𝑃𝐸𝐶 of using a fixed baseline, we compute the advantage score 𝐴𝑑𝑣𝜅 relative to the sample mean: 𝜅 1 ∑︁ 𝑟𝑗 (19) 𝐴𝑑𝑣𝜅 = 𝑟𝜅 − 𝜅 𝑗=1

4.2.2 Surrogate Performance Model (𝑓𝜙 ). To eliminate the high latency of exploring all action space combinations, surrogate GNN is employed as fast performance estimator. This model operates on an augmented feature representation, where each node includes its assigned placement device, (𝑥𝑖𝑠𝑢𝑟𝑟 = [𝑡𝑦𝑝𝑒𝑖 ||𝑑𝑒𝑣𝑖 ]). The features are processed through GCN layers, followed by global mean to obtain a graph-level embedding. This embedding is concatenated with the global state 𝑠 and passed through the MLP to predict the scalar PEC. This framework samples assignments 𝑎𝑖 ∼ 𝜋𝜃 via a categorical distribution, with independent sampling for sequential tasks and structure aware sampling for parallel tasks to ensure load balancing and data locality. By embedding constraints within the probability space, surrogate model mitigates sparse reward issues in constrained combinatorial optimization. Algorithm 1 describes the offline supervised training of the GNN surrogate model, consisting of data generation and model training phases. The process begins by initializing an null assignments to 𝐷𝑡𝑟𝑎𝑖𝑛 (Line 2). During assignment generation (Lines 4-9), DAGs of varying sizes are processed by sampling valid placements (𝐴 (𝜅 ) ), evaluating them with a simulator to obtain execution time, energy, and cost, aggregating these into 𝑃𝐸𝐶𝜅 , and storing the resulting samples in training model 𝐷𝑡𝑟𝑎𝑖𝑛 (Lines 6-9). After initializing model parameters, the training phase (Line 10) iterates over epochs, where each sample (𝐺, 𝐴, 𝑠, 𝑃𝐸𝐶) is encoded with augmented node features, passed through the GNN to predict PEC, and optimized using d 2 with gradient based updates 𝜙 (Lines MSE loss L ← (𝑃𝐸𝐶 − 𝑃𝐸𝐶) 13-16). Finally, the algorithm outputs a trained surrogate model capable of faster cost estimation (Line 17).

This group relative formulation provides positive reinforcement only for good or above average assignments, improving stability and eliminating scale dependent bias across varying workflow complexities. Algorithm 2 outlines the training of proposed MSR-GNN optimization of baseline approach. The procedure begins with initialization (Lines 2-3), where a running baseline 𝑏 is set and buffers are prepared to store trajectory data. At each episode (Lines 5-10), the state 𝑠 and workflow DAGs 𝐺 are sampled to construct node features, and 𝜅 constraint-aware assignment trajectories are generated using the Í current policy. The rewards are then averaged (¯𝑟 ← 𝜅1 𝜅 S𝑅 [𝜅]) and normalized (𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒 (¯𝑟 )) (Lines 9-10), and compared with the baseline 𝑏 to compute the advantage (𝐴𝑑𝑣 ← 𝑟˜ − 𝑏). In the update phase (Lines 12-13), each trajectory contributes to the total loss L, combining advantage weighted policy gradients with entropy regularization, and the GNN parameters 𝜃 are updated via gradient descent 𝜂∇𝜃 L. Finally, the baseline 𝑏 is updated using an exponential moving average of recent rewards (Line 14) to stabilize subsequent training. 6

Learning based Scalable Multi-Objective Scheduling of Workflows across UD-IaaS-FaaS Systems

Algorithm 2 GNN-Based Constraint-Aware MSR

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

Algorithm 3 GNN-MSR Inference

1: procedure TrainPolicy(𝐺, 𝑓𝜙 , 𝜃 )

1: procedure Inference(𝐺, 𝑠, 𝜃 )

Initialize baseline 𝑏 ← 0 3: Initialize replay buffers S𝐴 , S𝐿 , S𝑅 , S𝐻 4: for episode 𝑡 = 1 to 𝑇 do 5: Sample state 𝑠 = [𝑁 img, 𝑛 iaas, 𝑤𝑇 , 𝑤 𝐸 , 𝑤𝐶 ] 6: Load DAG 𝐺 = (V, E) for 𝑁 img 7: Build initial node features 𝑋 8: (S𝐴 , S𝐿 , S𝑅 , S𝐻 ) ← GenMaskedTraj(𝐺, 𝑠, 𝑋, 𝜃 ) ⊲ Generate 𝜅 trajectories Í 9: 𝑟¯ ← 𝜅1 𝜅 S𝑅 [𝜅] ⊲ Reward processing 10: 𝑟˜ ← Normalize(¯𝑟 ); 𝐴𝑑𝑣 ← 𝑟˜ − 𝑏 11: for each trajectory 𝑘 do ⊲ Policy gradient update 12: Lpolicy ← −𝐴𝑑𝑣 · S𝐿 [𝜅]; Lentropy ← −𝛽 · S𝐻 [𝜅]; 13: L ← Lpolicy + Lentropy ; 𝜃 ← 𝜃 − 𝜂∇𝜃 L

2:

2:

𝑏 ← 𝛼𝑏 + (1 − 𝛼)𝑟˜ return trained policy 𝜋𝜃

14: 15:

3: 4: 5: 6: 7:

Build node features 𝑋 H (𝜅 ) ← GCNConv(𝑋, E, 𝜃 ) 𝑍 ← MLP([H (𝜅 ) ∥ 𝑠]) Apply constraint masking: 𝑍 ′ ← 𝑍 + 𝑀 𝜋 ← Softmax(𝑍 ′ ); 𝐴∗ ← arg max 𝜋 return optimal assignment 𝐴∗

MILP, GA Greedy-TEC MSR-GNN

⊲ Baseline update

4.3.2 Policy Gradient Update. The policy parameter 𝜃 is updated via gradient descent to maximize the expected relative advantage, with an entropy regularization term H (𝜋𝜃 ) (coefficient 𝛽 = 0.2) included to encourage exploration and prevent premature convergence. 𝜅  1 ∑︁  𝐴𝑑𝑣𝜅 · ∇𝜃 log 𝜋𝜃 (𝐴 (𝜅 ) |𝑠, 𝐺) + 𝛽∇𝜃 H (𝜋𝜃 ) (20) ∇𝜃 𝐽 (𝜃 ) ≈ 𝜅 𝜅=1

Application Requests

DAG

Performance Metrics

Real-time Execution with same System

Scheduler System Configuration (#UDs, #IaaS VMs, FaaS)

Assignment Task_i -> Device_j

Figure 2: Workflow Deployment and Execution flow

Table 1: Experimental Parameters Used in the Evaluation

4.3.3 Curriculum Learning Schedule. Random initialization in large DAGs rarely yields in probability of finding efficient configurations. So, the environment state 𝑠 is sampled using a curriculum learning strategy. Training begins with small workflows to learn basic UDIaaS trade-offs, progresses to intermediate sizes to capture data locality patterns, and finally scales to larger workflows. The shared message-passing weights of GNN enables effective transfer of these learned structural patterns to larger, unseen DAGs. A predefined training schedule prevents the agent from being overwhelmed by high-dimensional action spaces before learning fundamental resource trade-offs from smaller cases. Accordingly, training is organized into three stages, each targeting key structural patterns of the scientific pipeline.

Parameter Description

Value Range

Workflow Job Size Makespan Energy Consumption Monetary Cost Normalized Weights (𝑊𝑇 ,𝑊𝐸 ,𝑊𝐶 ) CPU Requirement of Jobs Memory Requirement of Jobs

2 - 80 0.5 - 2500 Sec 5000 - 150000 Joules 1 - 100 units 0.005 - 0.99 0.5 - 3 vCPU 128 - 500 MB

selection across the masked distribution: 𝐴∗ = arg max 𝜋𝜃 (𝑎𝑖 | 𝑠, 𝐺)

(21)

𝑎𝑖 ∈𝐷

(1) Stage 1: Structural Foundations: The agent learns basic UDIaaS trade-offs, with small graphs enabling broad exploration and initial understanding of task classes. (2) Stage 2: Intermediate Locality: With increased overlap, the agent learns data locality by co-locating tasks and capturing fan-in and fan-out dependencies. (3) Stage 3: Production Scale: The agent manages high parallelism and the aggregation bottlenecks by generalizing learned patterns to larger workflow sizes.

4.4

Algorithm 3 describes the deterministic, low-latency inference phase of the MSR-GNN framework, executed after policy parameter 𝜃 and it is optimized via curriculum based training. For a given workflow 𝐺, node features (H 𝜅 ) are processed through GCN layers to capture multi-hop dependencies, combined with the global state 𝑠, and passed through the MLP to produce placement logits 𝑍 (Lines 3-4). A constraint mask ensures only valid assignments are considered (Line 5), and Softmax followed by arg max yields the final placement 𝐴∗ (Line 6). Unlike the training phase, the inference procedure is a deterministic, low-latency operation suitable for real-time scheduling in dynamic IaaS-FaaS environments. This single forward pass inference operates in 𝑂 (|E |) time enables nearoptimal scheduling while decoupling training complexity from deployment.

Real-Time Inference

Upon completion of the curriculum guided training, the MSR-GNN framework is deployed for live scheduling. During inference, the probabilistic sampling is replaced with a deterministic arg max 7

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

GA

G Vasantha Reddy, Ram Bansal, Krashna, Suvarthi Sarkar, and Aryabartta Sahu

G-T

G-C

MSR-GNN

1.75

2.0

1.50

1.5 1.0 0.5

1.25 1.00 0.75 0.50 0.25

27

51

79

0.00

MILP 1.0

Monetary Cost

2.5

0.0

G-E

×105

Energy ( J )

Time (s)

×103

G-TEC

27

51

79

Number of Workflow Jobs

×102

0.8 0.6 0.4 0.2 0.0

27

51

79

Figure 3: (a) Execution Time, (b) Energy Consumption and (c) Monetary Cost under varying number of workflow jobs Table 2: System Configuration

5

Components Controller

Count 1

IaaS Nodes

5

FaaS Nodes

3

5.2

Configuration CPU: Quad-core, RAM: 16 GB OS: Ubuntu24.0 CPU: 4-8 cores, RAM: 8-16 GB OS: Ubuntu22.04 Platform: Kubernetes Framework: Knative

Experimental Setup and Results

To evaluate the proposed MSR-GNN framework, we develop a comprehensive heterogeneous UD-IaaS-FaaS testbed that emulates a realistic compute continuum integrating IaaS VMs and FaaS resources. While serverless capabilities are implemented through a local FaaS deployment using Kubernetes cluster nodes with Knative serverless framework. The testbed comprises a centralized controller for workflow orchestration and scheduling, together with distributed compute nodes spanning both resource tiers. Real-world workflows, such as Pegasus Montage image processing workflow [26], face detection, log processing, and image processing applications are deployed to generate realistic execution traces (Ref. Table 1) and validate the proposed framework. The energy consumption on the UD (laptop) is measured by PowerTop and EnergyTop profiler of Linux system. The system architecture and hardware configuration of the testbed are summarized in Table 2.

5.1

Compared Approaches • The MILP formulation jointly determines task placement (UD/IaaS/FaaS), execution tier, and start times using binary decision variables and linear constraints, providing a lower bound on TEC due to its exhaustive search, even though its practical deployment is not considered. • Greedy methods include four variants: (a) Greedy-TEC (minimizes instantaneous TEC), (b) Greedy-T (execution time or makespan), (c) Greedy-E (energy), and (d) Greedy-C (cost), each making local decisions based on a single objective. • The Hybrid Genetic Algorithm (GA) initializes its population using greedy solutions to ensure diversity and faster convergence, with fitness evaluated via a simulation model equivalent to the MILP TEC formulation, enabling consistent comparison across all approaches.

5.3

Execution Time (Makespan) Analysis

Figure 3(a) shows the makespan results of the proposed MSR-GNN framework across different workload scales (27, 51 and 79 workflow jobs). Compared to baseline approaches, the proposed MSR-GNN approach achieves significant improvements, reducing execution time up to 10.7% relative to GA and up to 17.5% compared to GreedyTEC. Moreover, the framework consistently attains solutions within 3.5 to 4.1% of the MILP solution. These results highlight the GNNMSR capability optimize makespan with near-optimal accuracy, while avoiding the exponential computational complexity of offline MILP method.

Deployment and Execution 5.4

Figure 2 illustrates the execution of a typical workflow and the corresponding task-to-resource assignments on the testbed. The final task assignment 𝐴∗ is serialized into a NODE-ASSIGNMENT dictionary, mapping task IDs to resource labels (laptop, iaas_1, faas). This mapping is then provided to the workflow executor, which manages container orchestration and data transfers across the heterogeneous compute continuum.

Energy Consumption Analysis

Figure 3(b) illustrates the energy consumption trends across the varying DAG workflow sizes (27, 51 and 79 workflow jobs). The experimental results indicate that the MSR-GNN framework consistently reduces energy consumption across all the scenarios. Compared to baseline methods, MSR-GNN achieves overall energy savings of up to 3.1% over GA and up to 9.1% over Greedy-TEC. In 8

Learning based Scalable Multi-Objective Scheduling of Workflows across UD-IaaS-FaaS Systems

GA

Normalized TEC

Normalized TEC

GA G-TEC G-T G-E G-C MSR-GNN MILP

6 5 4 3 2 1 0

27

51

79

Number of Jobs

Figure 4: Normalized TEC across varying workflow sizes

GA

1.1

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

G-TEC

MSR-GNN

G-TEC 3.0 MSR-GNN 2.5 MILP 2.0 1.5 1.0 0.5 0.0[0.005,0.005,0.99][0.99,0.005,0.005] [0.34,0.33,0.33] [0.005,0.99,0.005] Weights [w_T, w_E, w_C] Figure 6: Normalized TEC under varying objective weights

MILP

1.0

Normalized TEC

10 12

Reward

0.9 0.8 0.7 1

2

3

4

Number of IaaS Servers

18 22 24

0

250 500 750 1000 1250 1500 1750 2000

Number of Episodes

Figure 7: Learning curve convergence: Reward vs. No. of Episodes

comparison with MILP, the framework demonstrates strong performance, producing energy-efficient placements within a margin of 5.1 to 5.5% of the MILP solution.

5.7 Cost Analysis

Impact of IaaS Resource Availability

Figure 5 shows normalized TEC values for varying numbers of IaaS nodes (1 to 5) to evaluate adaptability under different infrastructure capacities. As resources increase, the proposed MSR-GNN framework consistently outperforms baseline approaches, achieving normalized TEC reductions of up to 16.8% compared to the GA and up to 23.5% against the G-TEC approach. the MSR-GNN agent maintains highly competitive performance, yielding placement strategies that fall within a narrow margin of just 5.5 to 10.8% of absolute optimality of MILP.

Figure 3(c) shows the monetary cost results across different approaches, the proposed MSR-GNN framework consistently outperforms baseline approaches in minimizing cost across varying size of application workflows (27, 51 and 79 workflow jobs). As compared to the traditional GA and Greedy-TEC approach, proposed MSR-GNN approach achieves cost reductions up to 4.3% and 9.9%, respectively. The MSR-GNN approach effectively navigates the high dimensional scheduling space, producing near-optimal placements within 4.5-5.6% of the MILP across all workflow scales.

5.8 5.6

16 20

5

Figure 5: Normalized TEC vs. Varying number of IaaS nodes

5.5

14

Normalized TEC Analysis

Impact of Objective Weight Variation

Figure 6 shows normalized TEC values for different weight configurations ([𝑤𝑇 , 𝑤 𝐸 , 𝑤𝐶 ]): time-focused [0.99, 0.005, 0.005], balanced [0.34, 0.33, 0.33], energy-focused [0.005, 0.99, 0.005], and costfocused [0.005, 0.005, 0.99]. The results highlight significant variation in TEC due to inherent trade-offs among objectives, where optimizing one metric may adversely affect others. Despite this, the proposed MSR-GNN framework consistently outperforms baseline heuristics, achieving TEC reductions of 7.8-12.4% over GA and 10.616.9% over Greedy-TEC. Furthermore, it maintains close proximity to the MILP optimum, with a narrow gap of 4.5-7.3% across all scenarios, demonstrating near-optimal performance under diverse objective preferences.

The Figure 4 shows the normalized execution time (makespan), energy, and cost (TEC) across different workflow scales (27 nodes, 51 nodes, and 79 nodes), demonstrating the effectiveness of the proposed MSR-GNN framework. Compared to baseline methods, MSR-GNN achieves significant improvements, reducing TEC by up to 25.0% relative to GA and up to 28.8% compared to Greedy-TEC. Also, our approach consistently attains solutions within 2.0-5.7% of the MILP optimum. These results highlight the GNN-MSR capability to jointly optimize makespan, energy, and cost with near-optimal accuracy, while avoiding the exponential computational complexity of offline optimization methods. 9

Accepted in ICDCN 2027, January 4–7, 2027, Surathkal, India

5.9

G Vasantha Reddy, Ram Bansal, Krashna, Suvarthi Sarkar, and Aryabartta Sahu

Learning Behavior of the MSR-GNN Policy

[7] Qinqin Tang, Renchao Xie, Fei Richard Yu, Tianjiao Chen, Ran Zhang, Tao Huang, and Yunjie Liu. Distributed task scheduling in serverless edge computing networks for the internet of things: A learning approach. IEEE Internet of Things Journal, 9(20):19634–19648, 2022. [8] Ewa Deelman, Karan Vahi, Gideon Juve, Mats Rynge, Scott Callaghan, Philip J. Maechling, Rajiv Mayani, Weiwei Chen, Rafael Ferreira da Silva, Miron Livny, and Kent Wenger. Pegasus, a workflow management system for science automation. Future Generation Computer Systems, 46:17–35, 2015. [9] Liuyan Liu, Haisheng Tan, Shaofeng H.-C. Jiang, Zhenhua Han, Xiang-Yang Li, and Hong Huang. Dependent task placement and scheduling with function configuration in edge computing. In 2019 IEEE/ACM 27th International Symposium on Quality of Service (IWQoS), pages 1–10, 2019. [10] Shuiguang Deng, Hailiang Zhao, Zhengzhe Xiang, Cheng Zhang, Rong Jiang, Ying Li, Jianwei Yin, Schahram Dustdar, and Albert Y. Zomaya. Dependent function embedding for distributed serverless edge computing. IEEE Transactions on Parallel and Distributed Systems, 33(10):2346–2357, 2022. [11] Muhammad Zakarya, Lee Gillam, Mohammad Reza Chalak Qazani, Ayaz Ali Khan, Khaled Salah, and Omer Rana. Backfillme: An energy and performance efficient virtual machine scheduler for iaas datacenters. IEEE Transactions on Services Computing, 18(2):660–672, 2025. [12] Jonatha Anselmi and Josu Doncel. Balanced splitting: A framework for achieving zero-wait in the multiserver-job model. IEEE Transactions on Parallel and Distributed Systems, 36(1):43–54, 2025. [13] Sheshadri K R and J. Lakshmi. Hybrid serverless platform for smart deployment of service function chains. IEEE Transactions on Cloud Computing, 13(1):351–368, 2025. [14] Yunfeng Zhao, Chao Qiu, Shangxuan Cai, Zhicheng Liu, Yu Wang, Xiaofei Wang, and Qinghua Hu. Multi-granularity weighted federated learning for heterogeneous edge computing. IEEE Transactions on Services Computing, 18(1):270–287, 2025. [15] Yaning Yang, Xiao Du, Yutong Ye, Jiepin Ding, Ting Wang, Mingsong Chen, and Keqin Li. Multi-objective deep reinforcement learning for function offloading in serverless edge computing. IEEE Transactions on Services Computing, 18(1):288– 301, 2025. [16] Peihao Wu, Haiming Chen, Tianhao Wu, Kaiqi Gu, and Yinshui Xia. Coldstart-aware offloading and resource allocation by importance sampling-based double dueling dqn in serverless edge computing. IEEE Internet of Things Journal, 12(19):40190–40205, 2025. [17] Behrooz Zolfaghari, Saeid Abrishami, Abbas Rasoolzadegan, and Bahman Javadi. Dynamic Function Placement and Request Scheduling of Serverless Workflows in Edge Environment . IEEE Transactions on Services Computing, 18(05):2781–2793, September 2025. [18] Suvarthi Sarkar, Sparsh Mittal, Shivam Garg, and Aryabartta Sahu. Scientific workflow scheduling in cloud considering cold start and variable pricing model. arXiv preprint arXiv:2504.21536, 2025. [19] Suvarthi Sarkar, Ketan Singh, Aryabartta Sahu, et al. Minimizing energy in reliability and deadline-ensured workflow scheduling in cloud. arXiv preprint arXiv:2505.16496, 2025. [20] Shubhradeep Roy, Suvarthi Sarkar, and Aryabartta Sahu. Profit maximization using collaborative storage management in multi-tier edge-cloud system. In 2023 IEEE 30th International Conference on High Performance Computing, Data, and Analytics (HiPC), pages 309–318, 2023. [21] Mohammad Sadegh Aslanpour, Adel N. Toosi, Muhammad Aamir Cheema, and Raj Gaire. Energy-aware resource scheduling for serverless edge computing. In 2022 22nd IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid), pages 190–199, 2022. [22] Tarek Elgamal, Atul Sandur, Klara Nahrstedt, and Gul Agha. Costless: Optimizing cost of serverless computing through function fusion and placement. In IEEE/ACM Symposium on Edge Computing, pages 300–312, 2018. [23] AWS. [Online], 2025. Available: https://instances.vantage.sh/aws/ec2/c3.large? currency=USD&duration=annually&region=ap-southeast-1. [24] K. Hrishchenko and Oleksii Pysarchuk. Action-masked reinforcement learning technology for order scheduling. Ukrainian Journal of Information Technology, 7:35–44, 01 2025. [25] Mirco Theile, Lukas Dirnberger, Raphael Trumpp, Marco Caccamo, and Alberto L. Sangiovanni-Vincentelli. Action mapping for reinforcement learning in continuous environments with constraints, 2024. [26] Joseph C. Jacob, Daniel S. Katz, and et al. Montage: a grid portal and software toolkit for science-grade astronomical image mosaicking. Int. J. Comput. Sci. Eng., 4(2):73–87, July 2009. [27] G Vasantha Reddy, Chandrashekhar Dharmarajan, Suvarthi Sarkar, and Aryabartta Sahu. C-pcp: Correlation-aware proportional critical path scheduling for input-sensitive serverless workflows. In 2026 IEEE 19th International Conference on Cloud Computing (CLOUD), pages 500–506, 2026. [28] Shubhradeep Roy, Suvarthi Sarkar, Koushik Mukka, Tanish Verma, and Aryabartta Sahu. Data-parallel and heterogeneity-aware scheduling for geodistributed serverless scientific workflows. In 2026 IEEE 19th International Conference on Cloud Computing (CLOUD), pages 78–88, 2026.

To analyze the convergence of proposed MSR-GNN framework, training reward is monitored across all the training episodes, as shown in Figure 7. A high variance is observed during the initial phase (0 to 500 episodes) due to stochastic action space exploration. however, the RL agent gradually stabilizes by leveraging surrogate cost feedback to reinforce efficient placements. The learning curve flattens beyond 1000 episodes, which indicates the convergence to a near-optimal policy. After approximately 1,250 episodes, the reward exhibits negligible variance, suggesting diminishing returns from further training and the risk of overfitting, which may reduce generalization to unseen workflows. Therefore, early stopping within 800 to 1100 episodes provides an effective balance between convergence and generalization.

6

Conclusion and Future Work

This work proposes a scalable and adaptive workflow scheduling framework for heterogeneous IaaS-FaaS systems using a GNN based multi-sample REINFORCE approach. The problem is formulated as a MDP, where GNNs capture task dependencies and global system dynamics, and a differentiable surrogate enables faster PEC estimation, reducing reliance on expensive simulations. Experimental results show that proposed MSR-GNN outperforms heuristic baseline approaches, achieving near-optimal performance within 3.5-5.6% of MILP, and adapts effectively to varying workflow sizes, resource availability, and objective weights. Overall, the framework offers an efficient, low-latency solution for large-scale DAG scheduling in hybrid systems. In Future work we plan to extend this framework in several directions. First, incorporating real-time system feedback and online learning would enable adaptation to dynamic workload arrivals and fluctuating resource and network conditions. Second, we extends this framework to support multi-user and multi-tenant scenarios with fairness constraints could improve applicability in shared cloud environments.

References [1] Wei Wang, Massimo Tornatore, Yongli Zhao, Haoran Chen, Yajie Li, Abhishek Gupta, Jie Zhang, and Biswanath Mukherjee. Infrastructure-efficient virtualmachine placement and workload assignment in cooperative edge-cloud computing over backhaul networks. IEEE Transactions on Cloud Computing, 11(1):653– 665, 2023. [2] Nikhil Kapoor and Sougata Mukherjea. Optimizing cloud deployment: Blending of iaas and faas for microservice architecture. In 2025 IEEE International Conference on Cloud Computing Technology and Science (CloudCom), pages 1–8, 2025. [3] Jing Peng, Zhi Zhou, Kongyange Zhao, Huirong Ma, and Xu Chen. Cost-efficient cloud-edge video analytics with hybird iaas and faas resources. In 2023 IEEE 29th International Conference on Parallel and Distributed Systems (ICPADS), pages 2515–2522, 2023. [4] Hung-Chin Jang and Ting-Ching Li. Adaptive privacy-preserving task scheduling with user device collaboration in high-density edge computing environments. In 2024 IEEE Conference on Communications and Network Security (CNS), pages 1–6, 2024. [5] Mohammad S. Aslanpour, Adel N. Toosi, Claudio Cicconetti, Bahman Javadi, Peter Sbarski, Davide Taibi, Marcos Assuncao, Sukhpal Singh Gill, Raj Gaire, and Schahram Dustdar. Serverless edge computing: Vision and challenges. In Proceedings of the 2021 Australasian Computer Science Week Multiconference, ACSW ’21, New York, NY, USA, 2021. Association for Computing Machinery. [6] Renchao Xie, Qinqin Tang, Shi Qiao, Han Zhu, F. Richard Yu, and Tao Huang. When serverless computing meets edge computing: Architecture, challenges, and open issues. IEEE Wireless Communications, 28(5):126–133, 2021. 10

Related documents

• Learning based Scalable Multi-Objective Scheduling of Workflows across UD-IaaS-FaaS Systems#1327136
• FedHV: Low-Overhead Hypervolume Weighting for Federated Multi-Objective Optimization#1117121
• Active Inference-Based Adaptive Routing for Heterogeneous Edge AI Services#120477
• Scheduling in Cloud,Fog, and Edge Computing: A Comparative Analysis and Decision Framework for Technique Selection#1298326
• Scheduling in Cloud,Fog, and Edge Computing: A Comparative Analysis and Decision Framework for Technique Selection#1291765