Daniel Waddington

Shaobo Li, Yirui Zhou, Yuan Xu et al.

We present the design and implementation of a RAG-based AI system benchmarking (RAGPerf) framework for characterizing the system behaviors of RAG pipelines. To facilitate detailed profiling and fine-grained performance analysis, RAGPerf decouples the RAG workflow into several modular components - embedding, indexing, retrieval, reranking, and generation. RAGPerf offers the flexibility for users to configure the core parameters of each component and examine their impact on the end-to-end query performance and quality. RAGPerf has a workload generator to model real-world scenarios by supporting diverse datasets (e.g., text, pdf, code, and audio), different retrieval and update ratios, and query distributions. RAGPerf also supports different embedding models, major vector databases such as LanceDB, Milvus, Qdrant, Chroma, and Elasticsearch, as well as different LLMs for content generation. It automates the collection of performance metrics (i.e., end-to-end query throughput, host/GPU memory footprint, and CPU/GPU utilization) and accuracy metrics (i.e., context recall, query accuracy, and factual consistency). We demonstrate the capabilities of RAGPerf through a comprehensive set of experiments and open source its codebase at GitHub. Our evaluation shows that RAGPerf incurs negligible performance overhead.

Daniel G. Waddington

Building storage systems has remained the domain of systems experts for many years. They are complex and difficult to implement. Extreme care is needed to ensure necessary guarantees of performance and operational correctness. Furthermore, because of restrictions imposed by kernel-based designs, many legacy implementations have traded software flexibility for performance. Their implementation is restricted to compiled languages such as C and assembler, and reuse tends to be difficult or constrained. Nevertheless, storage systems are implicitly well-suited to software reuse and compositional software construction. There are many logical functions, such as block allocation, caching, partitioning, metadata management and so forth, that are common across most variants of storage. In this paper, we present Comanche, an open-source project that considers, as first-class concerns, both compositional design and reuse, and the need for high-performance.

Daniel Waddington, Cornel Constantinescu

During the training of Large Language Models (LLMs), tensor data is periodically "checkpointed" to persistent storage to allow recovery of work done in the event of failure. The volume of data that must be copied during each checkpoint, even when using reduced-precision representations such as bfloat16, often reaches hundreds of gigabytes. Furthermore, the data must be moved across a network and written to a storage system before the next epoch occurs. With a view to ultimately building an optimized checkpointing solution, this paper presents experimental analysis of checkpoint data used to derive a design that maximizes the use of lossless compression to reduce the volume of data. We examine how tensor data and its compressibility evolve during model training and evaluate the efficacy of existing common off-the-shelf general purpose compression engines combined with known data optimization techniques such as byte-grouping and incremental delta compression. Leveraging our analysis we have built an effective compression solution, known as Language Model Compressor (LMC), which is based on byte-grouping and Huffman encoding. LMC offers more compression performance than the best alternative (BZ2) but with an order-of-magnitude reduction in the time needed to perform the compression. We show that a 16-core parallel implementation of LMC can attain compression and decompression throughput of 2.78 GiB/s and 3.76 GiB/s respectively. This increase in performance ultimately reduces the CPU resources needed and provides more time to copy the data to the storage system before the next epoch thus allowing for higher-frequency checkpoints.

Francesco Quinzan, Rajiv Khanna, Moshik Hershcovitch et al.

We study the fundamental problem of selecting optimal features for model construction. This problem is computationally challenging on large datasets, even with the use of greedy algorithm variants. To address this challenge, we extend the adaptive query model, recently proposed for the greedy forward selection for submodular functions, to the faster paradigm of Orthogonal Matching Pursuit for non-submodular functions. The proposed algorithm achieves exponentially fast parallel run time in the adaptive query model, scaling much better than prior work. Furthermore, our extension allows the use of downward-closed constraints, which can be used to encode certain fairness criteria into the feature selection process. We prove strong approximation guarantees for the algorithm based on standard assumptions. These guarantees are applicable to many parametric models, including Generalized Linear Models. Finally, we demonstrate empirically that the proposed algorithm competes favorably with state-of-the-art techniques for feature selection, on real-world and synthetic datasets.

Andrew Wood, Moshik Hershcovitch, Daniel Waddington et al.

Bayesian inference allows machine learning models to express uncertainty. Current machine learning models use only a single learnable parameter combination when making predictions, and as a result are highly overconfident when their predictions are wrong. To use more learnable parameter combinations efficiently, these samples must be drawn from the posterior distribution. Unfortunately computing the posterior directly is infeasible, so often researchers approximate it with a well known distribution such as a Gaussian. In this paper, we show that through the use of high-capacity persistent storage, models whose posterior distribution was too big to approximate are now feasible, leading to improved predictions in downstream tasks.

Andrew Wood, Moshik Hershcovitch, Daniel Waddington et al.

Dimensionality reduction algorithms are standard tools in a researcher's toolbox. Dimensionality reduction algorithms are frequently used to augment downstream tasks such as machine learning, data science, and also are exploratory methods for understanding complex phenomena. For instance, dimensionality reduction is commonly used in Biology as well as Neuroscience to understand data collected from biological subjects. However, dimensionality reduction techniques are limited by the von-Neumann architectures that they execute on. Specifically, data intensive algorithms such as dimensionality reduction techniques often require fast, high capacity, persistent memory which historically hardware has been unable to provide at the same time. In this paper, we present a re-implementation of an existing dimensionality reduction technique called Geometric Multi-Scale Resolution Analysis (GMRA) which has been accelerated via novel persistent memory technology called Memory Centric Active Storage (MCAS). Our implementation uses a specialized version of MCAS called PyMM that provides native support for Python datatypes including NumPy arrays and PyTorch tensors. We compare our PyMM implementation against a DRAM implementation, and show that when data fits in DRAM, PyMM offers competitive runtimes. When data does not fit in DRAM, our PyMM implementation is still able to process the data.