An official website of the United States government
Deep neural networks (DNNs) are increasingly popular owing to their ability to solve complex problems such as image recognition, autonomous driving, and natural language processing. Their growing complexity coupled with the use of larger volumes of training data (to achieve acceptable accuracy) has warranted the use of GPUs and other accelerators. Such accelerators are typically expensive, with users having to pay a high upfront cost to acquire them. For infrequent use, users can, instead, leverage the public cloud to mitigate the high acquisition cost. However, with the wide diversity of hardware instances (particularly GPU instances) available in public cloud, it becomes challenging for a user to make an appropriate choice from a cost/performance standpoint. In this work, we try to address this problem by (i) introducing a comprehensive distributed deep learning (DDL) profiler Stash, which determines the various execution stalls that DDL suffers from, and (ii) using Stash to extensively characterize various public cloud GPU instances by running popular DNN models on them. Specifically, it estimates two types of communication stalls, namely, interconnect and network stalls, that play a dominant role in DDL execution time. Stash is implemented on top of prior work, DS-analyzer, that computes only the CPU and disk stalls. Using our detailed stall characterization, we list the advantages and shortcomings of public cloud GPU instances for users to help them make an informed decision(s). Our characterization results indicate that the more expensive GPU instances may not be the most performant for all DNN models and that AWS can sometimes sub-optimally allocate hardware interconnect resources. Specifically, the intra-machine interconnect can introduce communication overheads of up to 90% of DNN training time and the network-connected instances can suffer from up to 5× slowdown compared to training on a single instance. Furthermore, (iii) we also model the impact of DNN macroscopic features such as the number of layers and the number of gradients on communication stalls, and finally, (iv) we briefly discuss a cost comparison with existing work.
more »
« less
This content will become publicly available on June 13, 2028
ZenFlow: Enabling Stall-Free Offloading for LLM Training
Fine-tuning large language models (LLMs) often exceeds GPU memory limits, prompting systems to offload model states to CPU memory. However, existing offloaded training frameworks like ZeRO-Offload treat all parameters equally and update the full model on the CPU, causing severe GPU stalls, where fast, expensive GPUs sit idle, waiting for slow CPU updates and limited-bandwidth PCIe transfers. We present ZenFlow, a new offloaded training framework that decouples updates between GPU and CPU by prioritizing important gradients. ZenFlow performs in-place updates of important gradients on GPU while offloading and accumulating less important ones on CPU without blocking training iterations, fully overlapping CPU work with GPU computation. To make this importance-aware design scalable, ZenFlow leverages a novel spatio-temporal locality of important gradients—concentrated in specific positions that remain stable across iterations—to make important gradient identification efficient and lightweight while preserving fidelity. Evaluation shows that ZenFlow achieves up to 5×end-to-end speedup, 2×lower PCIe traffic, 85% less GPU stalls, and up to 72% lower cloud GPU cost, all while preserving accuracy.
more »
« less
- PAR ID:
- 10708186
- Publisher / Repository:
- ACM SIGMOD/PODS International Conference on Management of Data
- Date Published:
- Volume:
- 4
- Issue:
- 4
- Format(s):
- Medium: X
- Location:
- Huntington Beach, CA
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Deep neural networks (DNNs) are increasingly popular owing to their ability to solve complex problems such as image recognition, autonomous driving, and natural language processing. Their growing complexity coupled with the use of larger volumes of training data (to achieve acceptable accuracy) has warranted the use of GPUs and other accelerators. Such accelerators are typically expensive, with users having to pay a high upfront cost to acquire them. For infrequent use, users can, instead, leverage the public cloud to mitigate the high acquisition cost. However, with the wide diversity of hardware instances (particularly GPU instances) available in public cloud, it becomes challenging for a user to make an appropriate choice from a cost/performance standpoint. In this work, we try to address this problem by (i) introducing a comprehensive distributed deep learning (DDL) profiler Stash, which determines the various execution stalls that DDL suffers from, and (ii) using Stash to extensively characterize various public cloud GPU instances by running popular DNN models on them. Specifically, it estimates two types of communication stalls, namely, interconnect and network stalls, that play a dominant role in DDL execution time. Stash is implemented on top of prior work, DS-analyzer, that computes only the CPU and disk stalls. Using our detailed stall characterization, we list the advantages and shortcomings of public cloud GPU instances for users to help them make an informed decision(s). Our characterization results indicate that the more expensive GPU instances may not be the most performant for all DNN models and that AWS can sometimes sub-optimally allocate hardware interconnect resources. Specifically, the intra-machine interconnect can introduce communication overheads of up to 90% of DNN training time and the network-connected instances can suffer from up to 5× slowdown compared to training on a single instance. Furthermore, (iii) we also model the impact of DNN macroscopic features such as the number of layers and the number of gradients on communication stalls, and finally, (iv) we briefly discuss a cost comparison with existing work.more » « less
-
There has been a growing interest in using GPU to accelerate data analytics due to its massive parallelism and high memory bandwidth. The main constraint of using GPU for data analytics is the limited capacity of GPU memory. Heterogeneous CPU-GPU query execution is a compelling approach to mitigate the limited GPU memory capacity and PCIe bandwidth. However, the design space of heterogeneous CPU-GPU query execution has not been fully explored. We aim to improve state-of-the-art CPU-GPU data analytics engine by optimizing data placement and heterogeneous query execution. First, we introduce a semantic-aware fine-grained caching policy which takes into account various aspects of the workload such as query semantics, data correlation, and query frequency when determining data placement between CPU and GPU. Second, we introduce a heterogeneous query executor which can fully exploit data in both CPU and GPU and coordinate query execution at a fine granularity. We integrate both solutions in Mordred, our novel hybrid CPU-GPU data analytics engine. Evaluation on the Star Schema Benchmark shows that the semantic-aware caching policy can outperform the best traditional caching policy by up to 3x. Compared to existing GPU DBMSs, Mordred can outperform by an order of magnitude.more » « less
-
Cloud applications are increasingly relying on hundreds of loosely-coupled microservices to complete user requests that meetan application’s end-to-end QoS requirements. Communication time between services accounts for a large fraction of the end-to-endlatency and can introduce performance unpredictability and QoS violations. This work presents our early work onDagger, a hardwareacceleration platform for networking, designed specifically with the unique qualities of microservices in mind. The Dagger architecturerelies on an FPGA-based NIC, closely coupled with the processor over a configurable memory interconnect, designed to offload andaccelerate RPC stacks. Unlike the traditional cloud systems that use PCIe links as the NIC I/O interface, we leverage memory-interconnectedFPGAs as networking devices to provide the efficiency, transparency, and programmability needed for fine-grained microservices. We showthat this considerably improves CPU utilization and performance for cloud RPCs.more » « less
-
This paper presents a comparative performance analysis of three popular Python data manipulation libraries—Pandas, Polars, and Dask—within the context of deep learning training pipelines. The existing studies in this area do not embed the libraries inside a full deep-learning training pipeline where data loading, preprocessing, and batch feeding interact tightly with GPU workloads. To bridge this gap, we integrate Pandas, Polars, and Dask into representative deep learning training and inference pipelines and conduct experi- ments across a wide range of various machine learning models and datasets, measuring key performance indicators such as runtime, memory usage, disk usage, and energy consumption (CPU and GPU). Our comprehensive analysis reveals that Polars consistently minimizes CPU energy consumption on larger workloads, while Pandas remains competitive for moderate sizes. Dask’s overhead can lead to higher energy usage on small to moderate datasets. All three libraries achieve similar runtimes for heavy GPU workloads (ResNet, Mask R-CNN). Polars and Pandas maintain lower CPU memory footprints than Dask, but Dask offers easier scalability if data truly exceeds available RAM. Polars shows marginal energy savings on the CPU during preprocessing.more » « less
