Search for: All records

Modern model hubs, such as Hugging Face, store tens of petabytes of LLMs, with fine-tuned variants vastly outnumbering base models and dominating storage consumption. Existing storage reduction techniques---such as deduplication and compression---are either LLM-oblivious or not compatible with each other, limiting data reduction effectiveness. Our large-scale characterization study across all publicly available Hugging Face LLM repositories reveals several key insights: (1) fine-tuned models within the same family exhibit highly structured, sparse parameter differences suitable for delta compression; (2) bitwise similarity enables LLM family clustering; and (3) tensor-level deduplication is better aligned with model storage workloads, achieving high data reduction with low metadata overhead. Building on these insights, we design BitX, an effective, fast, lossless delta compression algorithm that compresses XORed difference between fine-tuned and base LLMs. We build ZipLLM, a model storage reduction pipeline that unifies tensor-level deduplication and lossless BitX compression. By synergizing deduplication and compression around LLM family clustering, ZipLLM reduces model storage consumption by 54%, over 20% higher than state-of-the-art deduplication and compression approaches. 

Interactive notebook programming is universal in modern ML and AI workflows, with interactive deep learning training (IDLT) emerging as a dominant use case. To ensure responsiveness, platforms like Jupyter and Colab reserve GPUs for long-running notebook sessions, despite their intermittent and sporadic GPU usage, leading to extremely low GPU utilization and prohibitively high costs. In this paper, we introduce NotebookOS, a GPU-efficient notebook platform tailored for the unique requirements of IDLT. NotebookOS employs replicated notebook kernels with Raft-synchronized replicas distributed across GPU servers. To optimize GPU utilization, NotebookOS oversubscribes server resources, leveraging high inter-arrival times in IDLT workloads, and allocates GPUs only during active cell execution. It also supports replica migration and automatic cluster scaling under high load. Altogether, this design enables interactive training with minimal delay. In evaluation on production workloads, NotebookOS saved over 1,187 GPU hours in 17.5 hours of real-world IDLT, while significantly improving interactivity. 

The InterPlanetary File System (IPFS) is a pioneering effort for Web 3.0, well-known for its decentralized infrastructure. However, some recent studies have shown that IPFS exhibits a high degree of centralization and has integrated centralized components for improved performance. While this change contradicts the core decentralized ethos of IPFS and introduces risks of hurting the data replication level and thus availability, it also opens some opportunities for better data management and cost savings through deduplication. To explore these challenges and opportunities, we start by collecting an extensive dataset of IPFS internal traffic spanning the last three years with 20+ billion messages. By analyzing this long- term trace, we obtain a more complete and accurate view of how the status of centralization evolves over an extended period. In particular, our study reveals that (1) IPFS shows a low replication level, with only 2.71% of data files replicated more than 5 times. While increasing replication enhances lookup performance and data availability, it adversely affects downloading throughput due to the overhead involved in managing peer connections, (2) there is a clear growing trend in centralization within IPFS in the last 3 years, with just 5% of peers now hosting over 80% of the content, significantly decreasing from 21.38% 3 years ago, which is largely driven by the increase of cloud nodes, (3) the default deduplication strategy of IPFS using Fixed-Size Chunking (FSC) is largely inefficient, especially with the default 256KB chunk size, showing near-zero duplication being detected. Although Content-Defined Chunking (CDC) with smaller chunks could save ∼1.8 petabytes (PB) storage space, it could impact user performance negatively. We thus design and evaluate a new metadata format that optimizes deduplication without compromising performance. 

Cold start delays are a main pain point for today’s FaaS (Function-as-a-Service) platforms. A widely used mitigation strategy is keeping recently invoked function containers alive in memory to enable warm starts with minimal overhead. This paper identifies new challenges that state-of-the-art FaaS keep-alive policies neglect. These challenges are caused by concurrent function invocations, a common FaaS workload behavior. First, concurrent requests present a trade-off between reusing busy containers (delayed warm starts) versus cold-starting containers. Second, concurrent requests cause imbalanced evictions of containers that will be reused shortly thereafter. To tackle the challenges, we propose a novel serverless function container orchestration algorithm called CIDRE. CIDRE makes informed decisions to speculatively choose between a delayed warm start and a cold start under concurrency-driven function scaling. CIDRE uses both fine-grained container-level and coarse-grained concurrency information to make balanced eviction decisions. We evaluate CIDRE extensively using two production FaaS workloads. Results show that CIDRE reduces the cold start ratio and the average invocation overhead by up to 75.1% and 39.3% compared to state-of-the-art function keep-alive policies. 

Federated learning (FL) has emerged as a new paradigm of machine learning (ML) with the goal of collaborative learning on the vast pool of private data available across distributed edge devices. The focus of most existing works in FL systems has been on addressing the challenges of computation and communication heterogeneity inherent in training with edge devices. However, the crucial impact of I/O and the role of limited on-device storage has not been explored fully in FL context. Without policies to exploit the on-device storage for placement of client data samples, and schedule clients based on I/O benefits, FL training can lead to inefficiencies, such as increased training time and impacted accuracy convergence. In this paper, we propose FedCaSe, a framework for efficiently caching client samples in-situ on limited on-device storage and scheduling client participation. FedCaSe boosts the I/O performance by exploiting a unique characteristic--- the experience, i.e., relative impact on overall performance, of data samples and clients. FedCaSe utilizes this information in adaptive caching policies for sample placement inside the limited memory of edge clients. The framework also exploits the experience information to orchestrate the future selection of clients. Our experiments with representative workloads and policies show that compared to the state of the art, FedCaSe improves the training time by 2.06× for accuracy convergence at the scale of thousands of clients. 

FaaS (Function-as-a-Service) workloads feature unique patterns. Serverless functions are ephemeral, highly concurrent, and bursty, with an execution duration ranging from a few milliseconds to a few seconds. The workload behaviors pose new challenges to kernel scheduling. Linux CFS (Completely Fair Scheduler) is workload-oblivious and optimizes long-term fairness via proportional sharing. CFS neglects the short-term demands of CPU time from short-lived serverless functions, severely impacting the performance of short functions. Preemptive shortest job first—shortest remaining process time (SRPT)—prioritizes shorter functions in order to satisfy their short-term demands of CPU time and, therefore, serves as a best-case baseline for optimizing the turnaround time of short functions. A significant downside of approximating SRPT, however, is that longer functions might be starved. In this paper, we propose a novel application-aware kernel scheduler, ALPS (Adaptive Learning, Priority Scheduler), based on two key insights. First, approximating SRPT can largely benefit short functions but may inevitably penalize long functions. Second, CFS provides necessary infrastructure support to implement user-defined priority scheduling. To this end, we design ALPS to have a novel, decoupled scheduler frontend and backend architecture, which unifies approximate SRPT and proportional-share scheduling. ALPS’ frontend sits in the user space and approximates SRPT-inspired priority scheduling by adaptively learning from an SRPT simulation on a recent past workload. ALPS’ backend uses eBPF functions hooked to CFS to carry out the continuously learned policies sent from the frontend to inform scheduling decisions in the kernel. This design adds workload intelligence to workload-oblivious OS scheduling while retaining the desirable properties of OS schedulers. We evaluate ALPS extensively using two production FaaS workloads (Huawei and Azure), and results show that ALPS achieves a reduction of 57.2% in average function execution duration compared to CFS. 

The InterPlanetary File System (IPFS) has recently gained considerable attention. While prior research has focused on understanding its performance characterization and application support, it remains unclear: (1) what kind of files/content are stored in IPFS, (2) who are providing these files, (3) are these files always accessible, and (4) what affects the file access performance. To answer these questions, in this paper, we perform measurement and analysis on over 4 million files associated with CIDs (content IDs) that appeared in publicly available IPFS datasets. Our results reveal the following key findings: (1) Mixed file accessibility: while IPFS is not designed for a permanent storage, accessing a non-trivial portion of files, such as those of NFTs and video streams, often requires multiple retrieval attempts, potentially blocking NFT transactions and negatively affecting the user experience. (2) Dominance of NFT (non-fungible token) and video files: about 50% of stored files are NFT-related, followed by a large portion of video files, among which about half are pirated movies and adult content. (3) Centralization of content providers: a small number of peers (top-50), mostly cloud nodes hosted by tech companies, serve a large portion (95%) of files, deviating from IPFS's intended design goal. (4) High variation of downloading throughput and lookup time: large file retrievals experience lower average throughput due to more overhead for resolving file chunk CIDs, and looking up files hosted by non-cloud nodes takes longer. We hope that our findings can offer valuable insights for (1) IPFS application developers to take into consideration these characteristics when building applications on top of IPFS, and (2) IPFS system developers to improve IPFS and similar systems to be developed for Web3. 

As the number of pre-trained machine learning (ML) models is growing exponentially, data reduction tools are not catching up. Existing data reduction techniques are not specifically designed for pre-trained model (PTM) dataset files. This is largely due to a lack of understanding of the patterns and characteristics of these datasets, especially those relevant to data reduction and compressibility. This paper presents the first, exhaustive analysis to date of PTM datasets on storage compressibility. Our analysis spans different types of data reduction and compression techniques, from hash-based data deduplication, data similarity detection, to dictionary-coding compression. Our analysis explores these techniques at three data granularity levels, from model layers, model chunks, to model parameters. We draw new observations that indicate that modern data reduction tools are not effective when handling PTM datasets. There is a pressing need for new compression methods that take into account PTMs' data characteristics for effective storage reduction. Motivated by our findings, we design Elf, a simple yet effective, error-bounded, lossy floating-point compression method. Elf transforms floating-point parameters in such a way that the common exponent field of the transformed parameters can be completely eliminated to save storage space. We develop Elves, a compression framework that integrates Elf along with several other data reduction methods. Elves uses the most effective method to compress PTMs that exhibit different patterns. Evaluation shows that Elves achieves an overall compression ratio of 1.52×, which is 1.31×, 1.32× and 1.29× higher than a general-purpose compressor (zstd), an error-bounded lossy compressor (SZ3), and the uniform model quantization, respectively, with negligible model accuracy loss. 

Cloud object storage such as AWS S3 is cost-effective and highly elastic but relatively slow, while high-performance cloud storage such as AWS ElastiCache is expensive and provides limited elasticity. We present a new cloud storage service called ServerlessMemory, which stores data using the memory of serverless functions. ServerlessMemory employs a sliding-window-based memory management strategy inspired by the garbage collection mechanisms used in the programming language to effectively segregate hot/cold data and provides fine-grained elasticity, good performance, and a pay-per-access cost model with extremely low cost. We then design and implement InfiniStore, a persistent and elastic cloud storage system, which seamlessly couples the function-based ServerlessMemory layer with a persistent, inexpensive cloud object store layer. InfiniStore enables durability despite function failures using a fast parallel recovery scheme built on the auto-scaling functionality of a FaaS (Function-as-a-Service) platform. We evaluate InfiniStore extensively using both microbenchmarking and two real-world applications. Results show that InfiniStore has more performance benefits for objects larger than 10 MB compared to AWS ElastiCache and Anna, and InfiniStore achieves 26.25% and 97.24% tenant-side cost reduction compared to InfiniCache and ElastiCache, respectively.