An official website of the United States government
The imbalanced I/O load on large parallel file systems affects the parallel I/O performance of high-performance computing (HPC) applications. One of the main reasons for I/O imbalances is the lack of a global view of system-wide resource consumption. While approaches to address the problem already exist, the diversity of HPC workloads combined with different file striping patterns prevents widespread adoption of these approaches. In addition, load-balancing techniques should be transparent to client applications. To address these issues, we proposeTarazu, an end-to-end control plane where clients transparently and adaptively write to a set of selected I/O servers to achieve balanced data placement. Our control plane leverages real-time load statistics for global data placement on distributed storage servers, while our design model employs trace-based optimization techniques to minimize latency for I/O load requests between clients and servers and to handle multiple striping patterns in files. We evaluate our proposed system on an experimental cluster for two common use cases: the synthetic I/O benchmark IOR and the scientific application I/O kernel HACC-I/O. We also use a discrete-time simulator with real HPC application traces from emerging workloads running on the Summit supercomputer to validate the effectiveness and scalability ofTarazuin large-scale storage environments. The results show improvements in load balancing and read performance of up to 33% and 43%, respectively, compared to the state-of-the-art.
more »
« less
This content will become publicly available on November 15, 2026
EMLIO: Minimizing I/O Latency and Energy Consumption for Large-Scale AI Training
Large-scale deep learning workloads increasingly suffer from I/O bottlenecks as datasets grow beyond local storage capacities and GPU compute outpaces network and disk latencies. While recent systems optimize data-loading time, they overlook the energy cost of I/O—a critical factor at large scale. We introduce EMLIO, an Efficient Machine Learning I/O service that jointly minimizes end-to-end data-loading latency (𝑇) and I/O energy consumption (𝐸) across variable-latency networked storage. EMLIO deploys a lightweight data-serving daemon on storage nodes that serializes and batches raw samples, streams them over TCP with out-of-order prefetching, and integrates seamlessly with GPU-accelerated (NVIDIA DALI) pre-processing on the client side. In exhaustive evaluations over local disk, LAN (0.05 ms & 10 ms round trip time (RTT)), and WAN (30 ms RTT) environments, EMLIO delivers on average up to 8.6X faster I/O and 10.9X lower energy use compared to state-of-the-art loaders, while maintaining constant performance and energy profiles irrespective of network distance. EMLIO’s service-based architecture offers a scalable blueprint for energy-aware I/O in next-generation AI clouds.
more »
« less
- Award ID(s):
- 2313061
- PAR ID:
- 10658033
- Publisher / Repository:
- ACM
- Date Published:
- Page Range / eLocation ID:
- 2022 to 2031
- Format(s):
- Medium: X
- Location:
- St. Louis, MO
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
We study training of Graph Neural Networks (GNNs) for large-scale graphs. We revisit the premise of using distributed training for billion-scale graphs and show that for graphs that fit in main memory or the SSD of a single machine, out- of-core pipelined training with a single GPU can outperform state-of-the-art (SoTA) multi-GPU solutions. We introduce MariusGNN, the first system that utilizes the entire storage hierarchy—including disk—for GNN training. MariusGNN introduces a series of data organization and algorithmic contributions that 1) minimize the end-to-end time required for training and 2) ensure that models learned with disk-based training exhibit accuracy similar to those fully trained in memory. We evaluate MariusGNN against SoTA systems for learning GNN models and find that single-GPU training in MariusGNN achieves the same level of accuracy up to 8× faster than multi-GPU training in these systems, thus, introducing an order of magnitude monetary cost reduction. MariusGNN is open-sourced at www.marius-project.org.more » « less
-
Graph convolutional networks (GCNs) are fundamental in various scientific applications, ranging from biomedical protein-protein interactions (PPI) to large-scale recommendation systems. An essential component for modeling graph structures in GCNs is sparse general matrix-matrix multiplication (SpGEMM). As the size of graph data continues to scale up, SpGEMMs are often conducted in an out-of-core fashion due to limited GPU memory space in resource-constrained systems. Albeit recent efforts that aim to alleviate the memory constraints of out-of-core SpGEMM through either GPU feature caching, hybrid CPU-GPU memory layout, or performing the computation in sparse format, current systems suffer from both high I/O latency and GPU under-utilization issues. In this paper, we first identify the problems of existing systems, where sparse format data alignment and memory allocation are the main performance bottlenecks, and propose AIRES, a novel algorithm-system co-design solution to accelerate out-of-core SpGEMM computation for GCNs. Specifically, from the algorithm angle, AIRES proposes to alleviate the data alignment issues on the block level for matrices in sparse formats and develops a tiling algorithm to facilitate row block-wise alignment. On the system level, AIRES employs a three-phase dynamic scheduling that features a dual-way data transfer strategy utilizing a tiered memory system: integrating GPU memory, GPU Direct Storage (GDS), and host memory to reduce I/O latency and improve throughput. Evaluations show that AIRES significantly outperforms the state-of-the-art methods, achieving up to 1.8× lower latency in real-world graph processing benchmarks.more » « less
-
Today's data centers often need to run various machine learning (ML) applications with stringent SLO (Service-Level Objective) requirements, such as inference latency. To that end, data centers prefer to 1) over-provision the number of servers used for inference processing and 2) isolate them from other servers that run ML training, despite both use GPUs extensively, to minimize possible competition of computing resources. Those practices result in a low GPU utilization and thus a high capital expense. Hence, if training and inference jobs can be safely co-located on the same GPUs with explicit SLO guarantees, data centers could flexibly run fewer training jobs when an inference burst arrives and run more afterwards to increase GPU utilization, reducing their capital expenses. In this paper, we propose GPUColo, a two-tier co-location solution that provides explicit ML inference SLO guarantees for co-located GPUs. In the outer tier, we exploit GPU spatial sharing to dynamically adjust the percentage of active GPU threads allocated to spatially co-located inference and training processes, so that the inference latency can be guaranteed. Because spatial sharing can introduce considerable overheads and thus cannot be conducted at a fine time granularity, we design an inner tier that puts training jobs into periodic sleep, so that the inference jobs can quickly get more GPU resources for more prompt latency control. Our hardware testbed results show that GPUColo can precisely control the inference latency to the desired SLO, while maximizing the throughput of the training jobs co-located on the same GPUs. Our large-scale simulation with a 57-day real-world data center trace (6500 GPUs) also demonstrates that GPUColo enables latency-guaranteed inference and training co-location. Consequently, it allows 74.9% of GPUs to be saved for a much lower capital expense.more » « less
-
Massive multiuser (MU) multiple-input multiple-output (MIMO) promises significant improvements in spectral efficiency compared to small-scale MIMO. Typical massive MU-MIMO base-station (BS) designs rely on centralized linear data detectors and precoders which entail excessively high complexity, interconnect data rates, and chip input/output (I/O) bandwidth when executed on a single computing fabric. To resolve these complexity and bandwidth bottlenecks, we propose new decentralized algorithms for data detection and precoding that use coordinate descent. Our methods parallelize computations across multiple computing fabrics, while minimizing interconnect and I/O bandwidth. The proposed decentralized algorithms achieve near-optimal error-rate performance and multi-Gbps throughput at sub-1 ms latency when implemented on a multi-GPU cluster with half-precision floating-point arithmetic.more » « less
