Version 0.4 of the OCaml Bytream library, which streams bytes with low memory allocation, switches its slice handling to the Slice library from the Bstr project. It also adds a Zstandard compression module based on ozstd and a Bin-based module for declarative byte encoding and decoding.
— via OCaml Discuss, dx3mod
A proposal to merge the two argument-conversion loops in CPython's _PyEval_Vector into one loop over total_args, removing duplicated code and a second kwnames check while keeping conversion order and reference ownership unchanged. The test suite passes, and the author notes gcc and clang do not fuse the loops themselves, hinting at a small performance gain.
— via Python Discuss, Shamil
Researchers present SaMOSA, a sandbox for analyzing Linux malware that targets cyber-physical systems, featuring on-demand C2 response replay, lifecycle-bound orchestration, four synchronized side channels and multi-architecture support. A Kinsing cryptominer case study on ARM64 delivered the full C2 payload chain without live infrastructure and revealed a hidden Tor fallback and 59 new IPs.
— via arXiv Cryptography Security, Meet Udeshi, Venkata Sai Charan Putrevu, Prashanth Krishnamurthy, Ramesh Karri, Farshad Khorrami
Birdseed is presented as the first tensor compiler to handle all loop orderings over any standard sparse format in parallel, including parallel sparse scatters into list, bytemap, hash and dense levels. It claims asymptotic memory and time gains and empirical speedups on histogram construction and SpMSpV.
— via arXiv Distributed Computing, Adrian Gushin, Pau Sum, Willow Ahrens
A study of the LLVM -O3 pipeline split it into 113 cumulative per-pass prefixes and took 84,750 measurements on 30 PolyBench/C kernels. It found 36% of significant runtime-changing transitions were regressions, gains were back-loaded and concentrated in a few passes, and runtime-targeted passes also saved 30-60% energy.
— via arXiv Programming Languages, Federico Bruzzone, Walter Cazzola
The paper formalizes the Causal Observability Problem and proves that an external monitor of shared-memory operations cannot be both sound and complete, with external instrumentation giving soundness and internal giving completeness. Of three timestamping implementations, FAInc and Striped are linearizable while Collect fails all standard consistency models, yet all give identical soundness under external placement.
— via arXiv Distributed Computing, Gilde Valeria Rodr'iguez, Armando Casta~neda, Miguel Pi~na
LatentStore is a storage system that keeps compressed latents of AI-generated images as the durable copy and rebuilds pixels on demand with GPUs. Guided by a 35-month, 2-billion-request production trace, a prototype cut persistent storage by 78.7% with competitive or lower latency than pure image storage.
— via arXiv Distributed Computing, Zirui Wang, Yunjia Zheng, Tingfeng Lan, Zhaoyuan Su, Haoran Ni, Juncheng Yang, Yue Cheng
PDASC is a distributed approximate nearest neighbour search algorithm built on a hierarchical index that supports arbitrary, non-metric dissimilarity functions while keeping per-node memory low. Benchmarks across datasets and distance functions show competitive accuracy-efficiency trade-offs with consistently lower per-node memory than state-of-the-art ANN methods.
— via arXiv Databases, Elena Garcia-Morato, Maria Jesus Algar, Cesar Alfaro, Felipe Ortega, Javier Gomez, Javier M. Moguerza
Ray-by-Ray is a proposed GPU design that splits on-chip buffers into per-ray units and adds lightweight management structures, letting more warps run in parallel without extra buffer capacity. It averages a 1.91x speedup over the baseline at 0.19% area overhead.
— via arXiv Computer Architecture, Junrui Pan, Cesar Avalos Baddouh, Kai Ze Ee, Timothy G. Rogers
A benchmark study of four LLM inference optimization families on four Apple machines, with RTX 4090 comparisons, finds that gains do not transfer reliably. KV-cache pruning reached up to 1.75 times throughput on Apple, while all tested speculative decoding methods reduced it and ported CUDA kernels helped only partially, by up to 1.20 times.
— via arXiv Distributed Computing, Kyrylo Yemets, Andrii Yaroshevych, Victor Muryn, Ostap Khomenko, Pavlo Kryven, Maksym Kmet, Maksym Shamrai, Mariya Hirna
SnakeCharmer is a backend control framework for address-prediction-based GPU prefetchers that decides whether, when and where to retain each prefetch. On an A100 GPU model it improves performance and energy efficiency by 12.8% and 20.8% over Snake, and by 1.7% and 3.2% over APOGEE.
— via arXiv Computer Architecture, Yuhan Tang, Tiejun Li, Jianmin Zhang, An Hu, Guangyu Sun, Sheng Ma, Yiqun Lang, Hanqing Li, Siqing Fu
A comparative analysis of CLI versus MCP for connecting AI agents to external systems, spanning 14 evaluation dimensions plus security and economic assessments. It finds CLI more efficient for single-user local workflows, MCP better for access control and governance, and recommends a hybrid architecture for production.
— via arXiv Distributed Computing, Arun Malik, Nitish Talasu, Bradley Klemick, Vineet Tushar Trivedi, Prachi Shah, Sanay Devi, Junjie Zhang, Kunal Sonawane
FIGFLEX is an algorithm-hardware co-designed accelerator for mixed-precision FP-INT matrix multiplication in weight-quantized LLMs, paired with a new quantizer. On LLaMA-2-7B it reports WikiText-2 perplexity of 5.85 (3-bit) and 8.77 (2-bit), beating GPTQ, AWQ and others, and 1.56x and 1.42x higher throughput per area than AxCore and FIGLUT.
— via arXiv Computer Architecture, Faraz Tahmasebi, Haocheng Xu, Andrew Trauger, Yomal Perera, Zhenyu Tang, Nael Chowdhury, Sitao Huang
Mozilla confirmed that eight eMudhra intermediate certificates (six TLS, two code-signing) had flawed key generation that made the private keys trivially recoverable, using integers below 256. Mozilla distrusted them via OneCRL within hours and eMudhra later revoked them as caCompromise, while the Sectigo cross-signs remained unrevoked.
— via Hacker News (top 30), IcePenguin
A Zig build script generates an ISO with xorriso, then modifies it in place with a Limine tool before QEMU runs. The build skips the mutation step unless a dependency is added manually, because LazyPath does not track mutating steps. One reply suggests a wrapper script that copies the file first.
— via Ziggit, mba6comma1
Researchers present a Linux kernel concurrency fuzzing approach that separates concurrent-input generation from thread-interleaving exploration, using May Race Pairs to pick inputs worth scheduling. On eight modules in Linux 6.17 it found 187 real data races, 59 of them harmful, outperforming SegFuzz and CONZZER at equal budget.
— via arXiv Operating Systems, Cen Zhang, Jia-Ju Bai, Zi-Yu Zhang, Zuming Jiang
uGDS is an open-source user-space library that handles NVMe command management on the host CPU while data moves by PCIe peer-to-peer DMA, supporting both NVIDIA CUDA and AMD HIP. On an A100 with a Samsung 990 PRO it reached up to 12.12x GDS's batch-read bandwidth and cut LMCache median time-to-first-token by up to 62.7%.
— via arXiv Distributed Computing, Jing Wei, Guanyi Chen, Qi Chen, Chuyu Zhang, Zheng Zhou, Chutian Wang, Jian Zhang
A software pipeline is proposed for recovering and validating DRAM address mapping functions on embedded LPDDR5 systems. It sweeps latency thresholds to find same-bank address pairs, solves for bank masks using GF(2) RREF, and checks them with group and single-function verification to support Rowhammer analysis on ARM-based devices.
— via arXiv Computer Architecture, Hewen Zhang, Ian G. Harris
Bootlin and the eBPF Foundation, backed by an Alpha-Omega grant, are adding KASAN memory-safety checks to JIT-compiled eBPF programs. The approach has the x86_64 JIT compiler emit calls to generic outline KASAN helpers before memory-accessing BPF instructions, skipping stack accesses.
— via Bootlin, Alexis Lothoré
GitHub describes moving its busiest Git repositories into Azure by shifting its Spokes quorum from repository-wide to per-reference write serialization, which internal tests showed gave 3–4x higher throughput for writes to different references. It also introduces ephemeral virtual replicas backed by object storage to scale reads independently.
— via GitHub Engineering, Eleanor White
