Source-linked AI summary
A proactive autoscaling and energy-efficient VM allocation framework using online multi-resource neural network for cloud data center
Deepika Saxena, Ashutosh Kumar Singh
TL;DR
Dynamic cloud workloads can cause resource waste, excess power consumption, and performance or QoS degradation. The paper combines concurrent multi-resource prediction, cluster-based VM autoscaling, and energy-efficient VM placement. Against WPWA-VMP, PA-VMP improves power saving by 86.71% and resource utilization by 13.52%.
Problem
Changing cloud resource demand can lead to excess power consumption, resource wastage, and performance or QoS degradation.
Source: §1 · pp. 2–3Method
The framework uses OM-FNN to predict multiple resource demands concurrently, clusters tasks for VM autoscaling, and places scaled VMs on energy-efficient physical machines.
Source: §1 · p. 4, §1.1 · p. 4Results
Against WPWA-VMP, PA-VMP improves power saving by 86.71% and resource utilization by 13.52%.
Source: §8.5 · p. 43Takeaways & Limitations
The framework consolidates workloads on as few physical machines as possible while seeking to preserve application performance.
Source: §9 · p. 44The number of node sets in OM-FNN's input and output layers must be selected manually.
Source: §9 · p. 45
Abstract
from arXiv · showhide
This work proposes an energy-efficient resource provisioning and allocation framework to meet the dynamic demands of future applications. The frequent variations in a cloud user's resource demand lead 'to the problem of excess power consumption, resource wastage, performance, and Quality-of-Service degradation. The proposed framework addresses these challenges by matching the application's predicted resource requirement with the resource capacity of VMs precisely and thereby consolidating the entire load on the minimum number of energy-efficient physical machines. The three consecutive contributions of the proposed work are: Online Multi-Resource Feed-forward Neural Network to forecast the multiple resource demands concurrently for future applications; autoscaling of VMs based on the clustering of the predicted resource requirements; allocation of the scaled VMs on the energy-efficient PMs. The integrated approach successively optimizes resource utilization, saves energy and automatically adapts to the changes in future application resource demand. The proposed framework is evaluated by using real workload traces of the benchmark Google Cluster Dataset and compared against different scenarios including energy-efficient VM placement with resource prediction only, VMP without resource prediction and autoscaling, and optimal VMP with autoscaling based on actual resource utilization. The observed results demonstrate that the proposed integrated approach achieves near-optimal performance against optimal VMP and outperforms rest of the VMPs in terms of power saving and resource utilization up to 88.5% and 21.12% respectively. In addition, the OM-FNN predictor shows better accuracy, lesser time and space complexity over a traditional single-input and single-output feed-forward neural network predictor.
1. Introduction
Changing cloud workloads can waste power and resources or degrade performance and QoS, while prediction and placement have often been treated separately. The paper combines multi-resource prediction, cluster-based VM autoscaling, and energy-efficient placement, then evaluates the framework on Google Cluster benchmark data.
1. Introduction
- The introduction identifies workload variation as a source of inefficient VM placement, resource waste, excess power use, and performance or QoS degradation.Prior work considered prediction and resource allocation separately, motivating a combined autoscaling-and-placement approach.Source: §1 · pp. 2–4
1.1. Our contributions
- The framework combines OM-FNN prediction, task clustering for VM autoscaling, and multi-objective placement of scaled VMs on energy-efficient physical machines.Its stated aim is to improve performance, resource utilization, and power saving.Source: §1 · p. 4, §1.1 · pp. 4–5
- OM-FNN forecasts multiple resource demands concurrently, while Error-Driven Padding accompanies the predictor's enhanced accuracy over traditional single-resource feed-forward neural networks.The proposed Tri-adaptive Differential Evolution algorithm optimizes OM-FNN by exploring the search space and exploiting the population of candidate networks.Source: §1.1 · pp. 4–5
- Evaluation on the real Google Cluster benchmark dataset reports that the framework outperforms state-of-the-art approaches in prediction, resource utilization, and power-consumption reduction.The contributions describe successive optimization during task assignment, VM scaling, and placement as reducing active servers and VM migrations.Source: §1.1 · p. 5
1.2. Organization
- The paper is organized from problem definition and related work through the framework, autoscaling, placement, algorithm and complexity, evaluation, and conclusion.The introduction assigns these topics to Sections 2–9.Source: §1.2 · p. 5
2. Problem definition
The problem definition frames cloud resource management as predicting VM resource needs and using those estimates to guide allocation, scaling, and placement. It identifies the overhead of separate predictors for each resource and sets out objectives for a concurrent predictor and adaptive resource management.
2. Problem definition
- The resource manager distributes resources, scales and schedules VMs, and deploys requested VM capacities on servers for user jobs.The problem represents a VM's actual requirement as a set of demands across multiple resources.Source: §2 · p. 6
- The proposed predictor is intended to estimate multiple resource demands concurrently, avoiding the overhead of deploying a separate predictor for each resource.The paper states that separate prediction raises time and space complexity in proportion to the number of resources.Source: §2 · p. 6
- The stated management objectives are adaptive VM autoscaling with reduced resource waste, energy-efficient placement, and avoidance of SLA violations.Predicted requirements are used to support load balancing, scale adequate VM types and quantities, and deploy on fewer active servers.Source: §2 · p. 6
3. Recent Key Contributions
Prior work addresses prediction, VM scaling, and placement, but the cited review identifies their interactive operation on a unified platform as missing. The proposed framework brings these operations together.
3.1. Workload prediction approaches based on Neural Network
- Evolutionary neural-network approaches improve on backpropagation in cited comparisons, while LSTM-RNN models offer high accuracy but incur long training time.The review also notes complex-valued and multi-valued-neuron alternatives, and VM predictors that use CPU alone or separate models per resource.Source: §3.1 · pp. 7–8
- Several prediction approaches have specific trade-offs: longer forecast intervals reduce backpropagation accuracy, and cross-validation for per-VM model selection is described as time-consuming.The review also says that one cited approach predicts CPU but not memory, while another uses separate predictors for each resource.Source: §3.1 · pp. 7–8
3.2. Scaling of VMs
- Fuzzy Q-learning and Fuzzy SARSA-learning handle sudden and periodic traffic, while anomaly-aware autoscaling uses isolation trees and cause identification.The review also describes malware detection, queue-guided scaling and packing, threshold-free event-driven tuning, memory-based vertical scaling, and Xen-based control.Source: §3.2 · pp. 9–10
- VM autoscaling research also addresses launch delays and resource adaptation, with cited approaches using cloud functions, proactive memory scaling, or dynamic vCPU allocation.The Xen controller is reported to reduce total CPU time compared with static allocation while minimizing SLA violations and stabilizing response time for high-priority VMs.Source: §3.2 · pp. 9–10
3.3. Multi-objective VM Placement
- VM-placement methods include GA, PSO, ACO, and Firefly approaches, but the review notes premature GA convergence and ACO’s difficulty recursively improving utilization.The cited SEA-LB method ignores VM migration; HGAPSO is described as unsuitable for heterogeneous placement because its encoding does not represent VM number and type.Source: §3.3 · pp. 10–11
- The proposed framework combines online multi-resource prediction, clustering-based VM scaling, and multi-objective placement, addressing the lack of interactive resource-management operations on a unified platform.The predictor is periodically retrained with live data and adds Error-Driven Padding to predicted outputs.Source: §3.3 · p. 11
4. Proposed resource management framework
The framework prepares historical and live task data for per-VM, multi-resource prediction, then clusters predicted requirements to autoscale and place VMs on energy-efficient servers. This workflow repeats over successive prediction intervals.
4. Proposed resource management framework
- Predicted task requirements are clustered by computational and storage complexity to estimate the number and configuration of VMs needed for future workloads.The scaled VMs are then placed on energy-efficient servers, and the provisioning process repeats over consecutive prediction intervals.Source: §4 · pp. 12, 14
- Figure 2’s workflow uses historical and live task data to predict future resource requirements, with a distinct multi-resource predictor dedicated to each VM.CPU and memory attributes are extracted, aggregated per unit time, and normalized; the prediction input includes the previous error score as Error-Driven Padding.Source: §4 · p. 12, Figure 2 · p. 13
5. Proactive VM autoscaling
The framework predicts multiple resource demands online, then clusters predicted task requirements to scale VMs. Error-driven padding and adaptive training address prediction uncertainty and changing workloads.
5.1. Online Multiple Resource Prediction
- OM-FNN predicts multiple resource utilizations concurrently, using a dedicated predictor for each VM and joint classification and prediction during optimization.Its input combines historical and live utilization to forecast the next instance.Source: §5.1 · pp. 14–15
- TaDE adapts mutation, crossover, and control parameters while training OM-FNN networks, evaluating fitness with RMSE and minimizing prediction error.Mutation and crossover strategies are selected adaptively, and control rates are regenerated to guard against premature convergence.Source: §5.1 · pp. 15, 17, 20
5.1.1. Error-driven padding (EDP):
- Error-driven padding adds a weighted recent prediction error to predicted demand to provide precaution margins against over- or under-loading and SLA violations.The padding weights the latest error more heavily; α is constrained to 0.5 < α ≤ 1.Source: §5.1.1 · p. 23
5.1.2. An Illustration
- The illustration traces network fitness evaluation, parameter initialization, mutation and crossover, and selection of successful offspring for the next iteration.The example uses an OM-FNN with CPU and memory nodes.Source: §5.1.2 · pp. 23–24
5.2. Autoscaling of VMs
- K-means clusters predicted task resource utilization, assigning tasks to groups with similar utilization while keeping clusters distinct.The Elbow method selects the number of clusters; assignments minimize squared distance to cluster centroids.Source: §5.2 · pp. 24–25
- Each cluster maps to a suitable VM type based on its maximum resource requirement, and the number of VMs equals the number of tasks in that cluster.The framework considers small, medium, large, and extra-large VM types.Source: §5.2 · pp. 25–26
6. VM placement
VM placement balances resource utilization and power consumption while respecting server capacity. A multi-objective evolutionary procedure searches feasible VM allocations for high utilization and lower power use.
6. VM placement
- VM placement seeks maximum resource utilization and minimum power consumption, subject to each VM's resource requirements fitting within its assigned server's available capacity.The formulation considers CPU and memory and can extend to additional resources.Source: §6 · p. 26, §6.1 · p. 26
6.1. Resource utilization
- The resource-utilization objective measures allocated CPU and memory against server capacity, with server and VM mappings indicating active placements.A server is active when it hosts one or more VMs.Source: §6.1 · p. 26
6.2. Power consumption
- Server power is modeled separately for idle and busy states, with busy-state consumption depending on application assignment and CPU utilization.Total data-center power is computed over a specified time interval.Source: §6.2 · p. 27
6.3. Optimized VM-Allocation approach
- The allocation algorithm ranks candidate placements by Pareto dominance, applies crossover and mutation, repairs infeasible allocations, and selects better solutions for subsequent generations.VMs are migrated from non-optimal to selected servers during the search.Source: §6.3 · pp. 27–28
7. Resource management algorithm and complexity analysis
The resource-management algorithm periodically forecasts task needs, scales VMs, and places them without external intervention. Its stated overall time complexity depends on the VM-placement and Pareto-ranking search parameters.
7. Resource management algorithm and complexity analysis
- Algorithm 4 periodically forecasts resource use, clusters predicted tasks, maps clusters to VM types, determines VM counts, and calls multi-objective placement.When tasks arrive, they are assigned to the selected autoscaled VMs.Source: §7 · p. 30
- The reported overall time complexity is O(on^2pqKGmax), incorporating placement, Pareto sorting, clustering, and their workload and search parameters.The passage defines o as the number of objectives, K as clusters, and q as tasks.Source: §7 · pp. 30–31
8. Performance evaluation
The evaluation uses Google Cluster Dataset workloads to assess prediction, VM autoscaling, and energy-efficient placement. The integrated placement approach improves power saving and resource utilization over a no-prediction/no-autoscaling baseline, while approaching optimal placement performance.
8.1. Experimental set-up
- The evaluation uses Google Cluster Dataset traces covering CPU, memory, and disk-I/O information from 672,300 jobs across 12,500 servers over 29 days.CPU and memory are consolidated over intervals from 5 to 1,440 minutes, and also aggregated every five minutes over 24 hours as task resource requirements.Source: §8.1 · pp. 31, 33
8.2. Comparative data conceptual analysis:
- The comparisons include prediction methods BP and SaDE, placement heuristics Random-Fit and Best-Fit, and secure energy-efficient placement SEA-LB.The evaluation assesses prediction accuracy, VM autoscaling, resource utilization, power consumption, and active-server counts.Source: §8.2 · pp. 33–34
8.3. Multiple resource prediction
- OM-FNN prediction error is no greater than SISO-FNN's in almost every case, while its operational time and space use are always lower.The reported reductions reach 46.8% in time and approximately 8.4% in space; larger prediction windows reduce time and space requirements but increase prediction error.Source: §8.3 · pp. 34, 36, Table 10 · p. 36
- For GCD CPU and memory prediction at five-minute intervals, TaDE-trained OM-FNN reduces forecast error by up to 95.18% and 85.29% versus BP and SaDE for CPU, and 96.25% and 78.04% for memory.The authors attribute the accuracy improvement to TaDE's three-phase adaptation of mutation, crossover, and control-parameter tuning.Source: §8.3 · p. 36, Table 11 · p. 37
- The proposed TaDE-based predictor converges in fewer than 45 iterations, and its adaptive control-parameter tuning updates crossover and mutation rates when evolution shows no significant improvement.The paper contrasts this adaptive learning period with SaDE's fixed mutation-rate learning period, which it says leads to slower and premature convergence.Source: §8.3 · pp. 37–38
8.4. Proactive VM Autoscaling
- Predicted-resource autoscaling selects VM types and quantities close to those obtained by scaling from actual resource utilization, yielding near-optimal scaling for future tasks.The comparison uses actual CPU and memory utilization as the optimal-scaling reference.Source: §8.4 · p. 38, Table 12 · p. 39
8.5. VM Placement
- Against WPWA-VMP, OA-VMP improves power saving by up to 88.5% and resource utilization by up to 21.12%.Figure 9 compares overall improvement per data center; the power-saving percentage is highest for data centers with 200 VMs and decreases as the data center grows.Source: §8.5 · p. 43, Figure 9 · p. 44
9. Conclusion
The paper concludes that its integrated prediction, autoscaling, and placement framework reduces resource wastage and saves energy on Google Cluster Dataset traces. It also identifies manual selection of OM-FNN input- and output-layer node-set counts as a limitation.
9. Conclusion
- The framework combines multi-resource OM-FNN prediction, cluster-based VM autoscaling, and multi-objective placement on energy-efficient servers.The authors report testing the integrated approach with real resource-utilization traces from the Google Cluster Dataset.Source: §9 · p. 44
- OM-FNN still requires manual selection of the number of node sets in its input and output layers.The authors identify scheduling interdependent tasks on closer VMs to reduce communication-related network traffic and energy expenditure as future work.Source: §9 · p. 45