Data-centric enterprises are migrating their analytic workloads from on-premises clusters to cloud object stores. While cloud storage offers elasticity and durability, it also introduces new overhead expenses (e.g., egress fees) and performance penalties (e.g., longer request latencies). To address these issues, we present Nexuio, a distributed cache system that bypasses conventional filesystem-style metadata operations to reduce both latencies and costs. It consists of two key techniques: (1) a page-based cache organization that avoids file operations (e.g., stat calls) by accessing data directly at the object level; and (2) a crossplatform runtime with zero-copy data transfer based on Arrow Flight. Nexuio can be integrated with various data analytics engines such as Presto and Velox to provide functionalities like failure handling, small-read optimizations, and skew mitigation. Deployed in a production system at Uber, Nexuio saves millions of dollars in GCS API cost, cuts egress traffic by 84.6%, and improves throughput by $4.4 \times \sim 8 \times$ compared to GCS gRPC clients. Lessons learned from empirical evaluation of Nexuio shed light on the deployment of petascale analytics in the cloud.
Data processing units (DPUs)-network interface controllers (NICs) with onboard general-purpose processors-have been explored as communication accelerators, but their utility as asymmetric coprocessors for memory-bandwidth-bound scientific HPC workloads remains an open question. We investigate this question in the context of the MiniEM proxy application. Specifically, we accelerate an algebraic multigrid (AMG)-based Maxwell solver, representing a class of solvers whose work decomposes into concurrent, imbalanced subtasks. Our implementation offloads one subtask to NVIDIA BlueField-3 (BF3) DPUs using a heterogeneous SPMD MPI model, requiring no DPU-specific offload framework. We then characterize when this strategy is profitable as a function of the CPU/DPU rank ratio. On a Xeon Broadwell + BF3 cluster and an EPYC Zen2 + BF3 cluster, we measure end-to-end speedups of up to 1.3× and 1.7×, respectively. We also show that as few as 4 of a BF3's 16 cores suffice to outperform the best CPU-only configuration.
High-Performance Linpack (HPL) remains the dominant metric for ranking supercomputers, yet its dense linear algebra workload often overstates the performance attainable under memory- and communication-intensive applications. This study investigates the structural determinants of the gap between HPL and the High Performance Conjugate Gradients (HPCG) benchmark. Using matched public metadata from the TOP500 and HPCG lists, we define an HPCG realism gap as the ratio between LINPACK-based achieved performance and HPCG performance, and estimate a log-linear model on 129 supercomputing systems. The results show that core density per socket has a consistently positive and statistically significant association with the realism gap, while system scale, strong scaling pressure, LINPACK efficiency, and GPU accelerator adoption do not exhibit independent explanatory power. These findings suggest that realistic performance degradation is driven less by headline scale or accelerator presence than by node-level resource concentration and memory-access contention. The results support the need to complement FLOPS-centered rankings with structurally informed evaluation of realistic workload performance.
Retrieval-augmented generation (RAG) answers questions by retrieving relevant passages from a document collection and conditioning an LLM on them [4]. As AI assistants increasingly process user-provided content [5], [7], including local files, notes, and application state, browser-local RAG is attractive because these data already enter through the browser, require no separate installation, and can remain on the user's device. Recent WebGPU inference systems and JavaScript ML runtimes make it feasible to run LLMs and embedding models inside a web page [3], [9], [10].
Accurate estimation of real-time power consumption is essential for energy-efficient computing systems and data centers. Existing Linux interfaces expose power information for only a subset of hardware components and therefore do not provide a unified view of total system power consumption. We present SysPower, a user-space framework that estimates system-wide power consumption by combining measurements from hardware telemetry sources with lightweight models for uninstrumented components. Unlike infrastructure-level monitoring solutions that rely on dedicated hardware and primarily target infrastructure management, SysPower provides a realtime estimate of system-wide power consumption together with subsystem-level power contributions using interfaces commonly available on existing systems. The resulting estimates are directly accessible to application and system software at a latency suitable for runtime energy-aware decision making. Evaluation against wall meter measurements shows a mean estimation error of 7.5% across diverse workloads while improving estimation accuracy relative to representative existing tools. SysPower is publicly available at https://github.com/UOFL-CSL/syspower.
