Do Bio-Inspired Algorithms Save Energy in Data Centres? Separating Server Packing from Migration Search in Virtual Machine Consolidation

Motivation

• Idle powered hosts draw 50–70% of peak demand, making it possible to save energy by concentrating load and deactivating the remaining hosts.The paper frames consolidation as a key IT-layer response as facility-level efficiency gains approach their practical limit.Source: §1 · pp. 1–2

Why consolidation is difficult

• VM consolidation is a multi-dimensional bin-packing problem; exact methods miss online timescales, while near-linear constructive heuristics can leave unusable capacity fragments.Metaheuristics occupy the interval between these approaches and can return an available solution during search.Source: §1 · p. 2

Research gap

• Prior comparisons vary both initial VM placement and migration search, so their energy differences cannot identify the contribution of metaheuristic search alone.The study evaluates whether reported improvements persist under matched initial conditions and workload-delivery requirements.Source: §1 · p. 2

Study aim and approach

• The study reproduces ACO, PSO and a modified genetic algorithm in CloudSim Plus, comparing them with a control that keeps the same initial packing and host deactivation but disables migration search.The framework also fingerprints workloads and tracks completion across two fleet configurations and four workload regimes.Source: §1 · pp. 2–3

Energy and cost models

• The underlying consolidation rationale is that powered idle servers retain a substantial fixed power cost, and formal models describe this as bin packing with linear usage costs.The matched control is used to measure how closely ordinary best-fit approaches the resulting lower bound without search.Source: §2 · p. 3

Algorithm approaches

• The literature includes ACO methods for static placement and migration, alongside genetic and swarm methods and newer reinforcement-learning hybrids.Recent work also considers renewable- and carbon-aware objectives and container-level placement.Source: §2 · p. 3

Evaluation gap

• Prior comparisons vary placement and migration search together, so their energy difference combines both effects and cannot isolate search.An energy reduction can also reflect unfinished work; a completion count is needed to establish that arms delivered equal work.Source: §2 · p. 3

Scope of the gap

• Methodological concerns about incomparable variants across papers are distinct from the within-study ambiguity caused by unmatched controls.The related-work discussion also notes arguments that research incentives favor proposing algorithms over comparing them.Source: §2 · pp. 3–4

Comparison design

• An unmatched control limits conclusions within one study, while one additional arm differing in a single respect can address that limitation without requiring a shared benchmark.Table 1 classifies the comparison structure of representative studies, the reproduced implementation and the present study.Source: §2 · pp. 3–4, Table 1 · p. 4

3.1 Placement, migration and workload

• VM placement occurs on arrival, while migration can be reconsidered each scheduling interval; both decisions are made by the allocation policy.Memory, storage, bandwidth and aggregate processing capacity constrain admission, while processing is time-shared after placement.Source: §3.1 · pp. 4–5

3.2 Power and energy

• A powered host’s modeled power includes a fixed idle component plus a term rising with processor utilisation.The study uses an idle fraction of 0.5; for the Dell PowerEdge R740, the stated model is 300+3.00u watts, with u in per cent. Energy is integrated over active hosts at each scheduling tick.Source: §3.2 · pp. 5–6

3.3 Decomposition of the energy saving

• Total energy saving is decomposed into admission-time placement and migration-search contributions by comparing baseline, matched-control and policy energies.A negative search effect means the metaheuristic used more energy than the same placement rule with search disabled.Source: §3.3 · p. 6

3.4 Energy per unit of delivered work

• Energy per unit of delivered work is calculated as energy divided by the fraction of submitted instructions executed, with cloudlet completion also reported separately.The analysis varies the charge for undelivered work from γ = 0 to γ = 1; the flip exponent marks where a positive search gain reverses.Source: §3.4 · p. 6

4.1 Separating admission from search

• Adjacent stack levels isolate admission-time packing and host deactivation from the added migration search, so each contrast changes one component.The spreading baseline, PackOnly control, and search policies form the three levels.Source: §4.1 · p. 7

4.2 Policy arms

• PackOnly shares the active-first, fewest-free-processing-elements placement rule and host shutdown with the search arms, but returns no migrations.The spreading baseline instead chooses suitable hosts with the greatest free capacity and keeps every host powered.Source: §4.2 · p. 7, Table 2 · p. 8

4.3 The search stage

• ACO searches pheromone-weighted host-to-host edges (50 evaluations per invocation), PSO moves assignment vectors (100), and MGA evolves assignment chromosomes (1,010).MGA uses fitness-proportionate parent selection, single-point crossover, and mutation; its fitness thresholds come from a control chart.Source: §4.3 · pp. 8–9, Figure 2 · p. 8

4.4 Verification of the design premises

• Across 93 fingerprinted runs, each load regime had one VM-and-cloudlet fingerprint pair, and all 272 runs created 100% of submitted VMs.The shared admission rule was verified in code across the four metaheuristic classes and the control.Source: §4.4 · p. 9

4.5 Operating characteristics of the search policies

• ACO, PSO, and artificial bee colony classify every active host as underloaded because SLA thresholds are percentages but utilisation is a fraction.Their migration is therefore driven by the underload path; MGA derives thresholds from a control chart instead.Source: §4.5 · pp. 9–10
• MGA searches only once per simulation; in seven of eight cells it makes zero migrations and matches PackOnly energy bit-for-bit across 70 runs.A guard returns an empty map after the first invocation.Source: §4.5 · p. 10
• Artificial bee colony keeps every host powered, averaging 19.98 of 20 active hosts in every cell, matching the baseline within 10^-5 of a host.Its energy results therefore do not obtain reductions through empty-host shutdown.Source: §4.5 · p. 10

5.1 Configuration

• The two fleets share 5120 GB of memory, and the common horizon lets percentage energy comparisons use the same interval.The configuration varies processing-capacity and power-rating distributions; search policies receive ten seed-based runs per condition, while the deterministic baseline and control run once per cell.Source: §5.1 · p. 10, Figure 3 · p. 11

5.2 Verification and validation

• The tick-integrated energy measure matches the expected work-to-dynamic-energy relationship within 1.3% in homogeneous runs; nine runs below capacity floors all delivered incomplete workloads.The reference arms were bit-identical across checked seeds, and MGA's seven zero-migration cells matched control energy across 70 runs.Source: §5.2 · p. 12

5.3 Statistical protocol

• Each testable search-versus-control contrast uses a one-sample t test of ten policy energies against the deterministic control, with nine degrees of freedom.Seven of 24 comparisons are structural identities for MGA's zero-migration cells and are recorded as p = 1; the reported family retains 17 testable comparisons after Holm correction.Source: §5.3 · p. 12

6.1 Decomposition of the energy saving

• 0.82%–81.40% of baseline energy is saved by placement alone; search raises energy in 11 cells, ties in seven, and lowers it in six.Fifteen search effects are significant after Holm correction: ten increases and five reductions.Source: §6.1 · pp. 12–13, Table 5 · p. 13, Figure 4 · p. 14

6.2 Workload completion and energy normalisation

• Five of six positive search effects reverse when unfinished work is charged; only ACO under heterogeneous Scenario 1 remains positive, reversing only at γ = 27.7.Across ACO and PSO cells, search effect correlates with work lost (r = 0.536, p = 0.032); the reported mechanism is lower processor utilisation when powered hosts receive less work.Source: §6.2 · pp. 13–14, Figure 5 · p. 15

6.3 Active hosts and the capacity floor

• Processing capacity binds in all eight cells; PackOnly reaches its floor in the first three homogeneous regimes but remains two to four hosts above it in the heterogeneous fleet.ACO reaches 6.12 active hosts against PackOnly’s 7.00 at heterogeneous Scenario 1; the placement rule ranks free processing elements without accounting for their speed.Source: §6.3 · p. 15, Figure 6 · p. 16

6.4 Search gain under heterogeneous light load

• ACO under heterogeneous light load gains 3.217 percentage points (95% CI 2.962–3.472), delivering 99.65% of its work and retaining a 3.11-point gain after full normalisation.The control is four hosts above the floor in this cell, while ACO reaches 6.12 active hosts; the gain is specific, not general across the 24 comparisons.Source: §6.4 · p. 16

6.5 Friedman rank comparison across arms

• Across four arms, raw-energy ranks do not separate any pair, but energy per delivered work ranks the control best and separates it from PSO.The raw-energy test gives χ2(3) = 4.39, p = 0.223; the work-normalised test gives χ2(3) = 14.89, p = 0.0019.Source: §6.5 · p. 17, Figure 7 · p. 17

6.6 Migration without consolidation, and an energy measure that concealed it

• ABC holds 19.98 of 20 hosts powered in every cell, yet its tick-integral saving ranges from −0.35% to +2.55%; continuous migration coincides with 85.6–98.4% work delivery.A mean-utilisation estimator instead reports 1.1%–24.1% savings, despite the hosts remaining powered throughout; the arm has only one to three replications per cell.Source: §6.6 · pp. 17–18, Figure S1 · p. 18

6.7 Validation against corrected utilisation thresholds

• Correcting utilisation-threshold scale moves search effects further from zero; in full-work cells, ACO shifts from +3.22 to −37.76 points at heterogeneous Scenario 1.Under the corrected scale, the three remaining cells deliver only 0.4%–5.1% of their workload, so the comparison rests on the three work-conserving cells.Source: §6.7 · pp. 18–19

7.1 Principal findings

• Against spreading, packing delivers 0.82%–81.40% of baseline energy across eight cells, with zero migrations and every cloudlet completed; migration search increases energy in 11 of 24 comparisons, leaves seven unchanged, and reduces six.Delivered-work normalisation removes five of six reductions; the remaining reduction is ant colony optimisation under light heterogeneous load, at 3.22 percentage points with 99.65% of work delivered.Source: §7.1 · p. 19

7.2 The capacity floor and unfinished work as mechanisms

• Packing reaches the processor-derived host-count floor in three homogeneous regimes; in heterogeneous fleets, its processor-element ranking leaves two to four hosts above the floor.The remaining gap is where the normalisation-robust ACO benefit occurs; unfinished work can lower measured energy because powered hosts accumulate energy whether or not their cloudlets complete.Source: §7.2 · pp. 19–20

7.3 Implications for data centre operation

• In these experiments, packing with host deactivation captured essentially all available consolidation energy while completing all submitted work, without search or migrations.Search time ranged from 0.4 seconds for MGA under light heterogeneous load to 202 seconds for ACO under heavy heterogeneous load; the work-complete search reduction was 3.22 percentage points in one cell.Source: §7.3 · p. 20

7.4 Implications for comparative evaluation

• Comparisons against spreading alone conflate placement and search; the authors call for a matched placement control, delivered-work reporting, the simulated horizon, and actual search effort.These reporting additions comprise one additional arm and three additional recorded quantities.Source: §7.4 · p. 20

7.5 Extensions

• The decomposition was tested on a 20-host fleet; rack-scale studies could test whether the relative placement and search effects persist with topology, thermal coupling, and failure domains.The authors also identify ten replications as computation-limited and server-level energy accounting as a boundary, proposing more replications and cooling and distribution parameters as extensions.Source: §7.5 · p. 21

Conclusion

• The matched control retains the reproduced policies’ placement and host-deactivation behaviour while disabling search, across two fleet configurations and four load regimes.Workload completion was recorded in all 272 runs.Source: §8 · p. 21
• Packing and host deactivation reduce energy by 0.8–81.4% relative to the spreading baseline, without migrations and with every submitted cloudlet completed.Adding search changes savings by −71.6 to +11.8 percentage points of baseline energy: energy rises in 11 of 24 comparisons, is identical in seven, and falls in six.Source: §8 · p. 21
• Five of six search-related reductions coincide with unfinished work and reverse after normalisation by delivered computational work; only light-load heterogeneous ACO retains an advantage.That ACO advantage is 3.22 percentage points in raw energy and 3.11 points after full work normalisation, with 99.65% of work delivered.Source: §8 · p. 21
• Packing reached the processor-derived host-count floor in three homogeneous regimes, while the surviving ACO benefit partly closed a gap above that floor in the heterogeneous fleet.The ABC arm kept as many hosts active as the baseline and left work unfinished; mean-utilisation energy estimation would have credited it with savings of up to 24.1%.Source: §8 · p. 21
• The findings concern the examined implementations and configuration, whose threshold handling and search invocation materially shaped results; they do not establish a general absence of value in metaheuristic consolidation.The authors recommend matched placement controls and reporting delivered work, simulated horizon, and actual search effort.Source: §8 · p. 21

CRediT authorship contribution statement

• Aun Yichiet is credited with conceptualization, methodology, formal analysis, writing, supervision, and project administration; Woo Yu Hang with software, investigation, data curation, validation, and writing.Source: CRediT authorship contribution statement · p. 22

Data availability

• The source, configuration, seeds, per-run records, workload fingerprints, and table- and figure-regeneration scripts are deposited at Zenodo under doi:10.5281/zenodo.23632346.Source: Data availability · p. 22

• Claude assisted with drafting and editing the manuscript’s language, structure, and exposition under continuous author supervision, sentence by sentence.Source: Declaration of generative AI and… · p. 22
• No generative AI tool was used to design, configure, or run simulations, generate or analyze results, or derive the reported conclusions.Source: Declaration of generative AI and… · p. 22
• The authors reviewed and edited the content as needed and take full responsibility for the published article.Source: Declaration of generative AI and… · p. 22