PaperlayerView exact revision on arXiv

Source-linked AI summary

TRACE: Tackling Real-World Resource Assignment Problems via Agentic Heuristic Design

Jose A. Ayala-Romero, Andres Garcia-Saavedra, Xavier Costa-Perez

arXiv:2610.01887v1cs.NEcs.LG

TL;DR

Dynamic resource assignment needs hardware-specific heuristics, but scalar-feedback AHD does not reveal the system behavior behind fitness, and execution logs are too large and variable for direct or fixed parsing. TRACE uses agents to extract objective-relevant insights and tools from those logs for heuristic evolution. It outperforms AHD methods across cloud and 5G vRAN scenarios, with lower energy use in 5G and more auditable heuristics.

  • Problem

    Scalar-feedback AHD reveals which heuristic scores better but not why, while real systems have unknown hardware behavior and logs that are large, objective-dependent, and heterogeneous.

    Source: §1 · p. 2

  • Method

    TRACE couples evolutionary heuristic search with agents that formulate objective-aware hypotheses and execute schema-specific code to distill logs into insights or tools.

    Source: §1 · p. 2, §2.2 · p. 3

  • Results

    TRACE outperforms state-of-the-art AHD methods across cloud and 5G vRAN scenarios, cutting 5G energy use by 65% relative to the industry standard without losing reliability.

    Source: §7 · p. 9

  • Takeaways & Limitations

    System execution logs are a valuable, largely untapped source of feedback for automatic heuristic design.

    Source: §7 · p. 9

    TRACE depends on agents capable of inferring latent dynamics and writing valid extraction code, as well as logs with informative fields; otherwise it yields no useful knowledge and falls back to scalar-feedback AHD.

    Source: §6 · p. 9

Abstract

from arXiv · show

Dynamic resource assignment, the real-time allocation of task streams to heterogeneous processing nodes, is the backbone of modern computing infrastructure. While learning-based schedulers excel in research, industrial deployments still rely on hand-written rules that operators can read, audit, and execute within tight latency budgets. LLM-based Automatic Heuristic Design (AHD) promises to automate writing such rules. However, existing AHD frameworks were developed for combinatorial problems fully specified to the LLM, and they learn only from a scalar fitness score. In real systems, the behaviour that determines a good heuristic, such as processor speeds or power consumption, is unknown a priori: the score reveals which heuristic performs better, but not why. This missing information is recorded in the system logs that every evaluation produces. Exploiting it is non-trivial: logs are massive and noisy, the relevant signals depend on the objective, and their content and format vary across hardware and software stacks, so they can neither be fed to an LLM as is nor processed by a fixed parser. We propose TRACE, which couples an evolutionary AHD loop with an agentic knowledge-extraction workflow. A Reasoner agent analyzes the log schema in light of the objective and formulates hypotheses about the system dynamics; a Coder agent writes and executes schema-specific code to test them, producing insights or executable tools for the evolved heuristics. We evaluate TRACE on a synthetic cloud benchmark and a 5G vRAN scenario built from industrial testbed measurements and operational traffic traces. TRACE consistently outperforms state-of-the-art AHD methods in resource assignment problems and yields more auditable heuristics at under 2% overhead.

1 INTRODUCTION

Dynamic resource assignment must allocate task streams across limited heterogeneous nodes, yet deployed scheduling rules require hardware-specific expertise. TRACE uses system logs to inform heuristic design and reports gains across cloud and 5G scenarios.

1 INTRODUCTION

  • Production schedulers remain predominantly hand-written because operators need rules they can inspect, certify, modify, and execute within tight latency budgets.Good rules require expertise in deployment-specific hardware, which must be reacquired for each deployment.Source: §1 · p. 1
  • Existing AHD frameworks evolve heuristics using scalar fitness, but real systems hide the node processing, energy, and queue characteristics that explain why one heuristic performs better.Their full-problem specifications make scalar feedback more suitable for the combinatorial benchmarks where these methods were developed.Source: §1 · p. 2
  • System logs record hardware behavior that could replace blind trial and error with informed heuristic design.For example, an energy-efficient policy may need to reserve a faster but power-hungry node for heavy tasks or system saturation.Source: §1 · p. 2
  • TRACE’s Reasoner identifies log information relevant to the objective, while its Coder writes and executes schema-specific extraction code to produce insights or tools for heuristic generation.This workflow addresses logs that are massive, objective-dependent, and heterogeneous in content and format.Source: §1 · p. 2
  • In cloud scheduling, TRACE is the only method with zero task loss at 50 servers and cuts cost by more than an order of magnitude versus the best AHD baseline.In 5G vRAN it matches the industry policy’s 0% loss while using 65% less energy than that policy and 25% less than SoTA AHD; its heuristics rank highest in semantic grounding and auditability, with 1.9% overhead versus EoH.Source: §1 · p. 2

2 PROBLEM FORMULATION

The formulation treats resource assignment as program design under unknown system dynamics: a task’s node determines its queueing, processing, resource cost, and possible drop. Candidate programs are evaluated on workloads using an objective-specific cost and execution logs.

2.1 DYNAMIC RESOURCE ASSIGNMENT

  • Each task arrives with a feature vector and is assigned to one of N heterogeneous nodes; the model makes no assumptions about arrival processes or feature distributions.A task waits in the node’s FIFO queue and is processed with a feature-conditioned processing-time distribution.Source: §2.1 · p. 3
  • A node has finite queue capacity, so assignment to a full queue drops the task; systems may also drop tasks that miss a deadline.Processing incurs a feature-conditioned resource cost such as energy.Source: §2.1 · p. 3
  • The processing-time, resource-cost, and queue-capacity functions are noisy and unknown a priori, reflecting undocumented hardware, firmware, and power-management characteristics.A heuristic is a program mapping task features to a node.Source: §2.1 · p. 3
  • A task’s cost combines its application-specific performance term when processed with a drop penalty Γ, chosen much larger so reliability is primary.Because waiting and drops depend on earlier assignments, a heuristic’s cost depends on the full trajectory.Source: §2.1 · p. 3
  • The objective is expected heuristic cost over a workload set, with the performance term instantiated as latency in cloud and energy in 5G vRAN.The 5G vRAN formulation also imposes a deadline.Source: §2.1 · p. 3

2.2 PROBLEM FORMULATION: HEURISTIC DESIGN FROM EXECUTION FEEDBACK

  • Designers search over programs without access to system dynamics, evaluating each candidate to obtain scalar fitness and execution logs containing task and system metrics.Logs can include processing time, drop status, queue occupancy, and energy.Source: §2.2 · p. 3
  • Log-driven design must handle growing logs that cannot fit in an LLM context, objective-dependent relevant metrics that require distillation, and system-specific metrics and formats.Distilled knowledge can be a textual insight or executable tool, such as a processing-time predictor.Source: §2.2 · p. 3

3 METHODOLOGY

TRACE combines evolutionary heuristic search with an agentic workflow that converts execution logs into knowledge for later candidates. The workflow proposes objective-aware extraction hypotheses, tests them with schema-specific code, and incorporates successful knowledge into system information.

3 METHODOLOGY

  • TRACE couples an evolutionary generation loop with an agentic knowledge-extraction workflow that distills logs into insights and executable tools for the generation prompt.The agents see only a log excerpt or schema; generated code processes the full log buffer, and extraction is tailored to the objective and system schema.Source: §3 · p. 3

3.1 HEURISTIC GENERATION LOOP

  • Each generation step prompts the LLM with the problem, current system information, two fitness-biased parent heuristics, and an evolutionary operator.Valid candidates enter a fixed-size population and their logs enter a bounded buffer spanning evaluations; a random-policy log initializes the buffer so all nodes are observed.Source: §3.1 · pp. 3–4

3.2 AGENTIC WORKFLOW

  • At each epoch, the workflow forms a diagnostic hypothesis, has a Coder test it on buffered logs, then merges the resulting insight or surrogate model into system information.The Reasoner chooses hypotheses from the objective, current information, and log excerpt or schema; the Coder produces schema-specific extraction code or a low-parameter predictor, and merging keeps the information concise. Helpful knowledge can propagate through high-fitness offspring, while unhelpful knowledge can be ignored.Source: §3.2 · pp. 4–5, Figure 3 · p. 5

3.3 SCHEDULING AND COST

  • The Agentic Workflow runs at the start of epochs until its successful-run budget is reached, after which system information is frozen and evolution continues alone.The asymmetric agent design keeps the workflow under 2% wall-clock overhead relative to EoH at an equal generation budget.Source: §3.3 · p. 5

4 EVALUATION

The evaluation tests TRACE’s scaling, transfer to a measurement-driven 5G system, computational overhead, and heuristic auditability. Across the cloud scenarios and 5G case, it compares TRACE with fixed policies and AHD baselines.

4.1 EVALUATION USE CASES

  • The cloud benchmark uses 10- and 50-server scenarios; the 5G vRAN case uses four heterogeneous processors and measured processing-time and energy behavior.Cloud tasks span low, medium, and high loads; the vRAN arrivals replay operational network traces.Source: §4.1 · p. 5, §4.4 · p. 7

4.2 BASELINES AND COMPARATIVE METHODS

  • Comparisons include fixed policies and four scalar-fitness AHD baselines under a shared 3,000-heuristic budget; TRACE’s gap over EoH isolates the workflow’s contribution.Coder failures to produce the surrogate sharply degrade performance, while replacing the Reasoner leaves performance unchanged.Source: §4.2 · pp. 5–6

4.3 CLOUD COMPUTING USE CASE

  • TRACE’s cloud heuristic uses a learned processing-time surrogate, penalizes fast processors to prevent saturation, and balances load with randomized top-k selection.The agentic workflow supplies processor-speed statistics and dominant task features; baseline heuristics instead converge to deterministic hashing.Source: §4.3 · p. 6
  • 36× lower cost in Scenario 2 (0.44 vs. 15.88, EoH), versus about 2× in Scenario 1 (0.06 vs. 0.13, HSEvo); TRACE also achieves zero loss and lowest delay in both.In Scenario 2, TRACE is the only AHD method with zero loss and the only method overall to achieve it without inflating delay.Source: §4.3 · p. 6, Table 1 · p. 7

4.4 MEASUREMENT-DRIVEN CASE STUDY: ENERGY-EFFICIENT 5G VRAN

  • TRACE matches 0% loss while using 64.9% less energy than HA-Only and 25.4% less than EoH in the measurement-driven 5G vRAN case.Its 0.36 ms delay remains within the 1 ms deadline.Source: Table 2 · p. 7, §4.4 · p. 7

4.6 AUDITABILITY ANALYSIS

  • TRACE scores 3.33 vs. 1.00–1.27 on Semantic Grounding and 4.67 against 1.33–2.00 on Auditability; every judge ranks it first on both.The scores are indicative because the study assesses one heuristic per method.Source: Table 3 · p. 8, §4.6 · p. 8

4.5 COMPUTATIONAL OVERHEAD

  • 1.9% wall-clock overhead: the workflow adds 161 seconds over 20 epochs and stops after five successful executions.The workflow runs a fixed number of times, so its overhead does not grow with processor count.Source: §4.5 · p. 8

4.6 AUDITABILITY ANALYSIS

  • TRACE’s policy cannot be evaluated by hand (Simulatability 1.00), while ReEvo’s modulo rule can (5.00) but encodes no system property.The learned surrogate can be inspected separately, unlike hashing rules, which lack such a decomposition.Source: §4.6 · p. 8

5 RELATED WORK

Earlier scheduling work assumes known processing times and infinite buffers, while AHD work has advanced search strategies without deriving system dynamics from execution logs. TRACE changes what the LLM observes, independently of its evolutionary backbone.

Dynamic resource assignment

  • Dynamic job-shop and related scheduling work typically assumes deterministic, known processing times and infinite buffers, unlike the streaming, finite-buffer systems targeted here.Prior work also includes hand-crafted heuristics, DRL, graph-based policies, and LLM-based approaches to job-shop variants.Source: §5 · p. 8

LLM-based Automatic Heuristic Design

  • To the authors’ knowledge, TRACE is the first AHD framework where an agent chooses what to extract from target-system logs and writes the extraction code, including executable surrogate models.Other AHD approaches use fitness feedback or feedback content fixed by the framework designer; TRACE’s workflow is orthogonal to the choice of evolutionary backbone.Source: §5 · p. 9

LLM-based design for networked and computing systems

  • Prior AHD work on networked and computing systems includes decoding, routing, virtual-network embedding, and edge-computing heuristics, but none derives unknown system dynamics from execution logs.The broader AHD literature has also varied its search procedures, including harmony search, reinforcement learning, tree search, and world-model planning.Source: §5 · p. 9

6 LIMITATIONS

TRACE’s gains depend on capable agents and informative logs, while discovery cost and surrogate-model latency can constrain deployment.

6 LIMITATIONS

  • TRACE depends on agents that can infer latent dynamics and produce valid extraction code, as well as logs containing informative fields; otherwise it falls back to scalar-feedback EoH.The authors report this failure mode in Llama-3.1-70B runs where the Coder failed to parse logs and train a surrogate.Source: §6 · p. 9
  • The roughly 2.5-hour one-time discovery cost is non-negligible when hardware changes frequently, though prior heuristics and new platform logs can reduce rediscovery compute.The workflow can re-derive hardware-dependent quantities, such as per-processor energy profiles and surrogate models, while adapting existing heuristics.Source: §6 · p. 9
  • For highly complex dynamics, the surrogate regressors needed may exceed the latency budget, although low-order regressions approximate the representative benchmark dynamics accurately.The authors identify the latency concern as a boundary for complex dynamics, not as an observed failure in their benchmarks.Source: §6 · p. 9

7 CONCLUSION

TRACE uses execution logs to guide heuristic evolution and outperforms state-of-the-art AHD methods in cloud and 5G vRAN scenarios. Its algorithm combines an agentic extraction workflow with an evolutionary generation loop.

7 CONCLUSION

  • 65% less energy than the industry-standard policy without losing reliability is one reported 5G vRAN result; TRACE also outperforms state-of-the-art AHD methods across cloud and 5G scenarios.The conclusion also reports more auditable heuristics at under 2% overhead.Source: §7 · p. 9
  • TRACE distills system execution logs into textual insights and executable surrogate models that guide heuristic evolution, extending AHD beyond scalar feedback.The paper presents system logs as a valuable, largely untapped source of AHD feedback.Source: §7 · p. 9

A TRACE’S ALGORITHM

  • At each epoch, TRACE alternates insight and surrogate workflows while within its workflow budget, and retries sandboxed Coder artifacts up to Mmax attempts before merging successful outputs.The algorithm initializes from a random-assignment-policy log and an EoH population initialized with system information.Source: A TRACE’S ALGORITHM · p. 13
  • The generation loop proposes and evaluates heuristics using available tools, records fitness and logs for valid candidates, and returns the candidate with minimum J(g).The algorithm bounds the retained log records by discarding the oldest entries beyond capacity.Source: A TRACE’S ALGORITHM · p. 13

B EXPERIMENTAL METHODOLOGY: CLOUD COMPUTING USE CASE

The cloud use case tests heterogeneous processors, finite queues, and varied task loads; TRACE uses a bounded agentic-workflow budget and separates reasoning from code generation.

B EXPERIMENTAL METHODOLOGY: CLOUD COMPUTING USE CASE

  • The cloud simulation uses d = 5 task features and hidden processor capability vectors; processing time is modeled from task–processor alignment, with finite FIFO queues of capacity 5–15.Tasks assigned to a full queue are dropped immediately.Source: B EXPERIMENTAL METHODOLOGY: CLOUD… · p. 13
  • Tasks arrive in 1 ms slots, with 90% specialized features and 10% uniformly random features; the heuristic observes task features but not hidden processor capabilities or instantaneous queue lengths.The simulation produces post-hoc event logs with arrival context, selected processor, latency when admitted, drop status, and arrival queue states.Source: B EXPERIMENTAL METHODOLOGY: CLOUD… · p. 13
  • The protocol evaluates 10- and 50-processor scenarios over 3,000 time steps at loads λ = 2, 5, and 10, measuring loss, delay, and cost.The drop penalty Γ is 135 for 10 processors and 600 for 50 processors.Source: B EXPERIMENTAL METHODOLOGY: CLOUD… · pp. 13–14

C TRACE CONFIGURATION

  • TRACE is configured with I = 150, E = 20, Mmax = 10, and Emax = 5; a frontier reasoning model handles the Reasoner while Qwen3 serves the Coder and heuristic-generation loop.The Reasoner is invoked at most five times per run, while the Coder and evolution loop perform thousands of generations.Source: C TRACE CONFIGURATION · p. 14
  • All methods receive the same total of 3,000 generations: TRACE runs 20 epochs of 150 generations, matching the AHD baselines’ generation count.The baselines named in the protocol are FunSearch, EoH, ReEvo, and HSEvo.Source: C TRACE CONFIGURATION · p. 14

D COMPUTATIONAL COST ANALYSIS

TRACE’s agentic workflow adds little runtime beyond standard AHD, while its external-token use is confined to the Reasoner. The appendix also describes symmetric backbone comparisons for TRACE and EoH.

D COMPUTATIONAL COST ANALYSIS

  • The Reasoner uses fewer than 600 input tokens and 30 output tokens per call, while the Coder, heuristic generation, and evaluation run locally.The local components use a non-reasoning 30B LLM and CPU execution.Source: D COMPUTATIONAL COST ANALYSIS · p. 15, Table 4 · p. 15
  • Across a 20-epoch run, the agentic workflow accounts for 1.8% of TRACE’s wall-clock time, or 1.9% overhead relative to equal-budget EoH.An epoch with the workflow takes about 32 seconds of agentic work; Table 5 profiles the per-epoch breakdown.Source: D COMPUTATIONAL COST ANALYSIS · p. 15, Table 5 · p. 15

E LLM COMPARISON

  • The LLM comparison varies TRACE’s Coder and heuristic-generation backbones, then separately varies EoH’s backbone, including TRACE’s Reasoner model.Both experiments use Scenario 1, 3,000 generations, and 10 independent runs.Source: E LLM COMPARISON · p. 16

E.1 LLM CHOICE FOR TRACE

The appendix compares TRACE and EoH across LLM backbones, audits the resulting heuristics, and examines a 5G vRAN scheduling objective. It reports that TRACE’s performance is tied to its agentic architecture, alongside strong grounding and auditability but low simulatability.

E.1 LLM CHOICE FOR TRACE

  • Qwen3-30B attains TRACE’s lowest median cost among the tested Coder and heuristic-generation backbones, while the other models are only marginally higher.The authors select Qwen3-30B as the default backbone for the main evaluations.Source: §E.1 · p. 16, Figure 5 · p. 16
  • Two TRACE runs with Gemma and Llama backbones suffered sharp efficiency drops after surrogate-model code failed within Mmax trials and the loop fell back to coarse text feedback.The authors state that TRACE’s gains depend on successful extraction and exploitation of system-log knowledge.Source: §E.1 · p. 16
  • Final TRACE performance is essentially unchanged across three Reasoner backbones, whose independently generated hypotheses converge on three ideas used in comparable heuristics.These are predicting processing time with regression, ranking processors by aggregate performance, and estimating queue length from drop events.Source: §E.1 · p. 16

E.2 LLM CHOICE FOR THE EOH BASELINE

  • EoH with gemini-3-pro-preview remains roughly 30× above TRACE’s median cost, indicating that the performance gap is governed by architecture rather than LLM choice.Upgrading EoH from Qwen3-30B reduces its median cost by roughly 30% but produces no qualitative jump.Source: §E.2 · p. 17, Figure 6 · p. 17

F AUDITABILITY STUDY

  • TRACE scores 4–5 on Auditability and is the only method above 1.3 on Semantic Grounding, while baselines score 1–2 on Auditability.The study scores best-found Scenario 1 algorithms using three judges.Source: F AUDITABILITY STUDY · p. 20, Table 6 · p. 20
  • TRACE scores 1/5 on Simulatability because its algorithm calls predict(), whereas ReEvo receives a unanimous 5/5.The authors distinguish mental trace-through from operational auditability: simulatability reflects ease of mental execution, not ease of modifying safety behavior.Source: F AUDITABILITY STUDY · p. 20

G.1 MOTIVATION AND PROBLEM FORMULATION

  • CPU latency rises with workload and may exceed deadlines, while GPU processing remains near-constant but consumes more energy, especially for smaller transport blocks.Lower SNR increases both processing latency and energy through additional decoding iterations.Source: §G.1 · p. 21
  • The 5G vRAN objective minimizes energy subject to a hard processing-delay constraint, treating latency as a binary feasibility condition rather than a continuous cost.The use case targets hard signal-processing deadlines and energy consumption in a production-grade testbed setting.Source: §G.1 · pp. 21–22

G.2 EXPERIMENTAL SETUP

This section describes the trace-driven heterogeneous 5G scheduling testbed and the prompts that combine problem context, extracted system knowledge, parent heuristics, and evolutionary instructions. It also presents prompt log fields and generated processor-selection algorithms, including reliability-aware, baseline, and input-driven stochastic approaches.

G.2 EXPERIMENTAL SETUP

  • The 5G setup models four heterogeneous processing units, empirical vRAN performance, operational traffic traces, and a hard 1 ms deadline with a reliability penalty.The cluster has two CPUs and two accelerators; HA1 is modeled as 20% better than HA2 in throughput and energy efficiency. Four base stations serve four users each at up to 100-MHz bandwidth, while replayed traces supply request timing and properties; missed deadlines discard TBs, with Γ = 10.Source: §G.2 · pp. 22–23

H.1 DEFINITIONS

  • TRACE’s generation prompt combines a formal problem description, dynamic system information, sampled parent heuristics, and crossover or mutation instructions.The system information conveys diagnostic knowledge and inferred constraints, while parent selection favors high-scoring candidates to balance exploration and exploitation.Source: §H.1 · p. 24

H.2 PROMPT EXAMPLES

  • The example prompts define scheduling objectives and provide the Reasoner with a CSV log schema for proposing new, data-grounded diagnostic hypotheses.The illustrated task prompt describes processor heterogeneity, queueing, task loss, and minimizing losses before total time in system; the log fields include time, features, loss, processing time, processor, and queue states.Source: §H.2 · pp. 24–25
  • Coder prompt examples request executable diagnostic artifacts, including code to process logs and generate insights or train and use a surrogate model.The prompts specify loading and processing the log data, printing extracted insights, and— for surrogate models—training, saving, and loading a model for prediction.Source: §H.2 · pp. 25–26
  • The prompt example lists the log fields time, features, lost, proc_time, proc, queue_states, and energy.Source: §H.2 · p. 27

I BEST GENERATED ALGORITHMS

  • The appendix presents the best-performing generated algorithm for each strategy discussed in Sec. 4.3.Source: I BEST GENERATED ALGORITHMS · p. 28
  • TRACE’s algorithm predicts processing times, avoids overloading a much-faster processor, and favors speed balanced with reliability and stability.It estimates processor variance on synthetic features and can switch from a low-reliability choice to a more reliable fast alternative.Source: I BEST GENERATED ALGORITHMS · pp. 28–29
  • Funsearch’s example transforms features nonlinearly and uses a deterministic hash to map them to one of 10 processors.Its design is described as stateless and memoryless under complete uncertainty.Source: I BEST GENERATED ALGORITHMS · pp. 29–30
  • EoH’s example hashes raw features using fixed random coefficients, applies tanh, and maps the output to a processor ID.Source: I BEST GENERATED ALGORITHMS · p. 30
  • ReEvo weights features with a fixed seed, while HSEvo derives a seed from input statistics and uses robust scaling.ReEvo’s listed weights emphasize later features; HSEvo’s fingerprint includes statistics such as variance and skewness.Source: I BEST GENERATED ALGORITHMS · pp. 30–31
  • The algorithm derives base phase, frequency, and modulation from entropy, skewness, and kurtosis.It sets phases across a frequency-scaled range and modulates them with entropy and kurtosis.Source: I BEST GENERATED ALGORITHMS · p. 31
  • It combines global, feature-variance-scaled, and phase-modulated noise as layered structured noise.The fragments describe adaptive noise scaling based on input complexity.Source: I BEST GENERATED ALGORITHMS · p. 31
  • The algorithm raises softmax temperature with feature variance, absolute skewness, and kurtosis, then computes log-sum-exp with centering for numerical stability.The temperature expression is 0.4 + 0.8 * (1 + variance + absolute skewness + 0.5 * kurtosis).Source: I BEST GENERATED ALGORITHMS · p. 31
  • The algorithm selects a processing unit stochastically using input-driven, reproducible randomness.Source: I BEST GENERATED ALGORITHMS · p. 31

J ANALYSIS OF THE AGENTIC WORKFLOW OUTPUTS

The workflow identifies substantial speed and feature-sensitivity differences across processors, and extracts information that can inform scheduling heuristics. Its analysis includes processing-time prediction and queue-length observations.

J ANALYSIS OF THE AGENTIC WORKFLOW OUTPUTS

K ENERGY-EFFICIENT 5G VIRTUALIZATION: RESULT ANALYSIS

TRACE's vRAN result analysis connects agent-derived system insights to a transparent, task-adaptive control policy balancing delay and loss risk, energy, and processor speed.

K ENERGY-EFFICIENT 5G VIRTUALIZATION: RESULT ANALYSIS

  • The diagnostic workflow identified a latency surrogate, hardware energy-throughput trade-offs, and Feature 3 (Transport Block Size) as the CPU bottleneck.The paper says this analysis recovered empirical profiles examined in Appendix G, Figure 7.Source: K ENERGY-EFFICIENT 5G VIRTUALIZATION:… · p. 34
  • TRACE's best-found vRAN policy uses an auditable weighted utility combining estimated processing time, queue-based loss risk, and relative speed penalties, with random jitter.The paper describes this policy as aligning energy-latency trade-offs with reliability targets.Source: K ENERGY-EFFICIENT 5G VIRTUALIZATION:… · p. 34
  • The policy changes weights with task feature magnitude: complex tasks prioritize speed and delay avoidance, while simple tasks prioritize energy and tolerate queues.The code comments specify different priorities for high- and low-feature-norm tasks.Source: K ENERGY-EFFICIENT 5G VIRTUALIZATION:… · p. 35
  • The policy assumes a model that estimates processing time from task features and processor identity, plus known average energy consumption per processor.The evolved system-information prompt provides the prediction model and per-processor energy values to the heuristic.Source: K ENERGY-EFFICIENT 5G VIRTUALIZATION:… · p. 35

Related papers

Loading 2610.01887v1…