Article Per Year (5 Year)
p-Index From 2021 - 2026
0.408
P-Index
This Author published in this journals
All Journal Journal of Technology Informatics and Engineering
Jiayi Nie
Operations Research, Columbia University, NY, USA
Computer Science & IT

Published : 2 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 2 Documents
Search

 1 
Power-Aware Inventory Planning for AI Infrastructure Using Job-Level Forecasting and LLM Workload Explanations Shilu He; Jiayi Nie; Chengliang Li
Journal of Technology Informatics and Engineering Vol. 5 No. 1 (2026): APRIL | JTIE : Journal of Technology Informatics and Engineering
Publisher : University of Science and Computer Technology

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.51903/jtie.v5i1.548

Abstract

AI infrastructure planning is commonly expressed as a GPU-count problem, yet operational risk is created by the electric and thermal envelope that accompanies each accelerator. This paper evaluates a power-aware planning method on Dataset A, using the B200 eight-GPU Llama-8B training trace with 45,000 raw 20 ms telemetry rows and 8,940 reproducible supervised decision records after a 100 ms decision stride. The forecasting task predicts total eight-GPU power one second ahead from job-level counters, autoregressive lags, and rolling statistics. The planning task converts forecasts into a peak-aware admission rule and a circuit-inventory simulation for 32 concurrent jobs. XGBoost produced the strongest mean forecast, with MAE 273.26 W, RMSE 636.74 W, and R2 0.923. A calibrated high-quantile forecast produced lower peak-error behavior, reducing the scheduling violation rate from 5.31% under GPU-count-only admission to 0.18% while admitting 61.63% of decision points. In the inventory simulation, XGBoost mean forecasting used 21.00 mean circuits with 1.80% violation risk, whereas the calibrated p95 plan used 22.70 circuits and eliminated observed violations in 1,000 trials. The results show that capacity plans based only on GPU count hide measurable electrical risk. A combined GPU-capacity, power-envelope, and workload-explanation view produces a reproducible basis for AI data center purchasing, placement, and sustainability decisions.
Token-Burst-Aware Capacity Planning for LLM Inference Services: Request Arrival, Token Demand, and Failure Risk Modeling from BurstGPT Traces Jiayi Nie; Yinchen Shi; Lucas Zhao
Journal of Technology Informatics and Engineering Vol. 5 No. 2 (2026): AUGUST | JTIE : Journal of Technology Informatics and Engineering
Publisher : University of Science and Computer Technology

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.51903/jtie.v5i2.565

Abstract

Large language model (LLM) inference services convert request arrivals into coupled input-token prefill and output-token decoding workloads. Capacity planning therefore depends on token volume, temporal bursts, service mix, queueing behavior, and reliability signals rather than request counts alone. This study evaluates an integrated planning pipeline on BurstGPT v1.1, comprising 5,288,173 raw requests over 121 trace days and 5,188,507 completed requests. Strictly chronological experiments aggregate demand, forecast hourly completed tokens, detect minute-level burst pressure, estimate zero-response risk, simulate capacity policies, and replay representative test hours in Vidur. Random forest, selected on the validation interval, achieved 64.73% weighted absolute percentage error (WAPE) on the locked test interval; XGBoost achieved the lowest test WAPE (64.57%), while the last-hour baseline reached 67.42%, indicating limited forecastability under a pronounced level shift. A seasonal-residual burst detector achieved F1 = 0.653, although burst prevalence was sensitive to rolling-horizon and quantile settings. For minute-level zero-response risk, raw XGBoost achieved ROC-AUC = 0.813 and average precision = 0.116; isotonic calibration improved the Brier score (0.0230–0.0204) and 10-bin expected calibration error (0.0262–0.0122), despite low F1. Static P90/P95 capacity eliminated under-provisioned test hours at cost indices of 13.10 and 18.09. More economical dynamic baselines achieved 12.65% under-provisioned hours at a cost index of 2.37 and 13.63% at 2.43. The validation-selected random-forest policy was cheaper but less reliable (34.31% at 1.23). Vidur replay linked normalized demand to A100/H100 GPU counts, latency, batching, and memory pressure. The results support conservative interpretation of point forecasts and validation of reserve rules under distribution shift, rare-event calibration, and serving-stack constraints.
Co-Authors Chengliang Li Lucas Zhao Shilu He Yinchen Shi