PetPS: Supporting Huge Embedding Models with Persistent Memory
Summary: PetPS is the first production PM parameter server for huge embedding models, using a workload-tailored hash index to reduce PM reads and NIC-offloaded gathering to cut CPU stalls. Deployed at Kuaishou, it lowers TCO 30% while improving throughput 1.3–1.7×. (summarized by gpt-5.6-luna on Jul 24 2026)
Incoming Non-self Citations Over Time
Authors
- 1. Minhui Xie (Tsinghua University)
- 2. Youyou Lu (Tsinghua University)
- 3. Qing Wang (Tsinghua University)
- 4. Yangyang Feng (Tsinghua University)
- 5. Jiaqiang Liu (Kuaishou)
- 6. Kai Ren (Kuaishou)
- 7. Jiwu Shu (Tsinghua University)
BibTeX Citation
@article{xie_vldb23,
title = {{PetPS: Supporting Huge Embedding Models with Persistent Memory}},
author = {Xie, Minhui and Lu, Youyou and Wang, Qing and Feng, Yangyang and Liu, Jiaqiang and Ren, Kai and Shu, Jiwu},
journal = {PVLDB},
series = {{VLDB} '23},
volume = {16},
number = {5},
pages = {1013--1022},
doi = {10.14778/3579075.3579077},
url = {https://doi.org/10.14778/3579075.3579077},
year = {2023}
}Incoming Citations (Sorted by Pagerank)
Showing 2 of 2 citing papers.
| Rank | Citing Paper | Year | Venue | Pagerank |
|---|---|---|---|---|
| 10,752 | CARINA: An Efficient CXL-Oriented Embedding Serving System for Recommendation Models | 2025 | SIGMOD | 5.093636e-05 |
| 11,220 | Sorting on Byte-Addressable Storage: The Resurgence of Tree Structure | 2024 | VLDB | 5.093636e-05 |
Previous
Page 1 / 1
Next
Outgoing Citations (Sorted by Pagerank)
Showing 6 of 6 cited papers.
Citations counted here include only citations to other VLDB/SIGMOD/CIDR/PODS papers in this database.
| Rank | Cited Paper | Year | Venue | Pagerank |
|---|---|---|---|---|
| 491 | FPTree: A Hybrid SCM-DRAM Persistent and Concurrent B-Tree for Storage Class Memory | 2016 | SIGMOD | 0.00017575163 |
| 1,368 | Dash: Scalable Hashing on Persistent Memory | 2020 | VLDB | 0.00011003184 |
| 2,471 | Viper: An Efficient Hybrid PMem-DRAM Key-Value Store | 2021 | VLDB | 8.5329393e-05 |
| 2,485 | HET: Scaling out Huge Embedding Model Training via Cache-enabled Distributed Framework | 2022 | VLDB | 8.5145736e-05 |
| 4,912 | HET-GMP: A Graph-based System Approach to Scaling Large Embedding Model Training | 2022 | SIGMOD | 6.4481656e-05 |
| 6,022 | NuPS: A Parameter Server for Machine Learning with Non-Uniform Parameter Access | 2022 | SIGMOD | 6.0046123e-05 |
Previous
Page 1 / 1
Next
Semantically Similar Papers
| # | Overall Rank | Paper | Year | Venue |
|---|---|---|---|---|
| 1 | 9,376 | H-Rocks: CPU-GPU accelerated Heterogeneous RocksDB on Persistent Memory | 2025 | SIGMOD |
| 2 | 6,519 | Dynamic Parameter Allocation in Parameter Servers | 2020 | VLDB |
| 3 | 4,430 | PQCache: Product Quantization-based KVCache for Long Context LLM Inference | 2025 | SIGMOD |
| 4 | 8,082 | Pea Hash: A Performant Extendible Adaptive Hashing Index | 2023 | SIGMOD |
| 5 | 5,345 | PS2: Parameter Server on Spark | 2019 | SIGMOD |
| 6 | 9,520 | Experimental Analysis of Large-scale Learnable Vector Storage Compression | 2024 | VLDB |
| 7 | 2,485 | HET: Scaling out Huge Embedding Model Training via Cache-enabled Distributed Framework | 2022 | VLDB |
| 8 | 4,912 | HET-GMP: A Graph-based System Approach to Scaling Large Embedding Model Training | 2022 | SIGMOD |
| 9 | 6,086 | Plush: A Write-Optimized Persistent Log-Structured Hash-Table | 2022 | VLDB |
| 10 | 2,693 | FlexPS: Flexible Parallelism Control in Parameter Server Architecture | 2018 | VLDB |