Research map / Graph learning / Knowledge and temporal graphs
FastPERT: Towards Fast Microservice Application Latency Prediction via Structural Inductive Bias over PERT Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FastPERT, a non-GNN latency predictor that uses a gradient boosted tree to estimate per-microservice-task delays on Trace-PERT graphs and averages all source-sink path sums via a precomputed matrix-vector product.
Problem
End-to-end latency prediction for microservice applications must be both accurate and fast enough for real-time proactive resource allocation, but GNN/graph-transformer models such as PERT-GNN are computationally expensive to train and run.
Approach
Replace a GNN with a node-wise gradient boosted tree delay estimator and compute the expected end-to-end latency as the average over all source-to-sink paths in a Trace-PERT DAG using path-count matrix powers, exploiting the constant path-duration (computational inductive bias) property to reduce prediction to one vector operation.
Key results
- On Alibaba 2022, FastPERT achieves 11.36% MAPE versus 12.84% for PERT-GNN, with training around 7.4 minutes on CPU versus 696 minutes on GPU for PERT-GNN.
- On Alibaba 2021, FastPERT reaches 11.59% MAPE and 1.61 ms MAE, comparable to PERT-GNN's 11.57% and 1.60 ms.
- FastPERT predicts 1.5 million traces in under 10 seconds on CPU, whereas PERT-GNN takes over 10 minutes with GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FastPERT instead of GBDT | Alibaba 2021 and Alibaba 2022 latency prediction · Ali-2021 test and Ali-2022 test | MAPE: 18.69% ± 0.38% -> 11.59% ± 0.53%; 17.29% ± 0.41% -> 11.36% ± 0.51% | Helped |
| use FastPERT instead of PERT-GNN | Alibaba 2021 latency prediction · Ali-2021 test | MAPE: 11.57% ± 0.82% -> 11.59% ± 0.53% | No clear effect |
| use FastPERT instead of PERT-GNN | Alibaba 2022 latency prediction · Ali-2022 test | MAPE: 12.84% ± 0.89% -> 11.36% ± 0.51% | Helped |
| use FastPERT instead of PERT-GNN | Alibaba 2021 and Alibaba 2022 latency prediction · Ali-2021 and Ali-2022 | Training time (minutes): 631.1 ± 181.7 -> 6.5 ± 3.0; 696.0 ± 170.7 -> 7.4 ± 4.2 | Helped |
| replace all-path averaging with the entering-service path (ablation) | FastPERT · Ali-2022 test | MAPE: 11.36% -> 14.44% | Hurt |
| replace all-path averaging with the entering-service path (ablation) | FastPERT · Ali-2022 test | MAE (ms): 0.6178 -> 0.6579 | Hurt |
| replace all-path averaging with the longest-delay path (ablation) | FastPERT · Ali-2022 test | MAPE: 11.36% -> 16.17% | Hurt |
| replace all-path averaging with the longest-delay path (ablation) | FastPERT · Ali-2022 test | MAE (ms): 0.6178 -> 0.8017 | Hurt |
| increase the node-delay loss weight 𝜆 (ablation) | FastPERT · Ali-2022 test | MAPE | Mixed |
Surprising findings
- The all-paths averaging aggregation outperforms the longest-delay path (ProjectPERT) by a clear margin, which the authors attribute to overestimation accumulating along long paths.
Limitations
- Accuracy is only comparable to PERT-GNN, not better, on the two Alibaba datasets.
- Evaluated only on Alibaba traces; relies on ground-truth task delays available in traces for node-level supervision.
- Tail latency (quantile) prediction is left as future work.
Experimental setup
Datasets: Alibaba 2021 microservice traces Alibaba 2022 microservice traces
Benchmarks: Alibaba 2021 microservice traces Alibaba 2022 microservice traces
Metrics: MAPE MAE training time inference time
Baselines: GBDT PERT-GNN ProjectPERT
Builds on: PERT-GNN Project-PERT graph
Closest papers by meaning
- Spatial-Temporal Heterogenous Graph Contrastive Learning for Microservice Workload Prediction (AAAI 2025 · similarity 0.85)
- MAPS: Memory-Aware Predictive Scheduling Framework for Large Language Model Serving (ICML 2026 · similarity 0.82)
- Root Cause Analysis in Microservice Using Neural Granger Causal Discovery (AAAI 2024 · similarity 0.82)
- Intelligence per Watt: Measuring Intelligence Efficiency of Local AI (NeurIPS 2026 · similarity 0.80)
- SkipPredict: When to Invest in Predictions for Scheduling (NeurIPS 2024 · similarity 0.80)
- DON’T STOP ME NOW: EMBEDDING BASED SCHEDULING FOR LLMS (ICLR 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.