Reducing the Cost of GPU Cold Starts in Serverless Deep Learning Inference Serving
Bibliographic record
Abstract
The rapid growth of Deep Learning (DL) has led to increasing demand for DL-as-a-Service. In this paradigm, DL inferences are served on-demand through a serverless cloud provider, which manages the scaling of hardware resources to satisfy dynamic workloads. This is enticing to businesses due to lower infrastructure management costs compared to dedicated on-site hosting. However, current serverless systems suffer from long cold starts where requests are queued until a server can be initialized with the DL model, which is especially problematic due to large DL model sizes. In addition, low-latency demands such as in real-time fraud detection and algorithmic trading cause long inferences in CPU-only systems to violate deadlines. To tackle this, current systems rely on over-provisioning expensive GPU resources to meet low-latency requirements, thus increasing the total cost of ownership for cloud service providers. In this work, we characterize the cold start problem in GPU-accelerated serverless systems. We then design and evaluate novel solutions based on two main techniques. Namely, we propose remote memory pooling and hierarchical sourcing with locality-aware autoscaling where we exploit underutilized memory and network resources to store and prioritize sourcing the DL model from existing host machines over remote host memory then cloud storage. We demonstrate through simulations that these techniques can perform up to 19.3× and 1.4× speedup in 99th percentile and median end-to-end latencies respectively compared to a baseline. Such speedups enable serverless systems to meet low-latency requirements despite dynamic workloads.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.008 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.001 | 0.003 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".