2024-W22

MinFlow: This paper presents MinFlow, a novel framework designed to enhance data passing in I/O-intensive serverless analytics applications. The framework aims to address the performance bottlenecks and high storage costs associated with the shuffle operation in serverless computing environments. MinFlow achieves this by optimizing data passing topologies and leveraging an interleaved partitioning strategy to reduce the number of PUT/GET operations and improve function scheduling.

DEX: Scalable Range Indexing on Disaggregated Memory explores a novel approach to designing scalable B+-trees for memory disaggregation. The paper introduces DEX, a scalable B+-tree that leverages logical partitioning, lightweight caching, and cost-aware offloading to enhance performance and scalability in disaggregated memory environments. The study demonstrates that DEX can significantly outperform state-of-the-art solutions under various configurations.

DLB-MPK: Sparse matrix-vector products (SpMVs) are a bottleneck in many scientific codes due to their low arithmetic intensity and heavy strain on the main memory interface. This paper introduces a novel method for enabling cache blocking in distributed-memory parallelization of Matrix Power Kernels (MPKs). By interleaving cache-blocking capabilities with an MPI communication scheme, the proposed Distributed Level-Blocked MPK (DLB-MPK) yields substantial speed-ups on modern Intel and AMD architectures across a wide range of sparse matrices. The method is demonstrated to achieve up to 4× speed-up on 832 cores of an Intel Sapphire Rapids cluster, showcasing its applicability to contemporary quantum physics problems.

PriCE: This paper introduces PriCE, a novel method aimed at optimizing the scheduling of large medical image processing tasks over hybrid clouds. The method focuses on privacy preservation and cost-effectiveness, addressing the challenges of running deep neural networks on large medical images. PriCE employs a multi-objective optimization approach to minimize privacy risks, execution time, and monetary costs under user budget constraints.

CuckooGraph: This paper introduces CuckooGraph, a novel data structure designed to efficiently handle large-scale dynamic graphs. The structure is capable of adapting to various data scales without prior knowledge of the graph size, ensuring both memory efficiency and fast query performance. The key innovations in CuckooGraph are the TRANSFORMATION and DENYLIST techniques. These techniques allow CuckooGraph to dynamically adjust its memory usage and handle insertion failures effectively, respectively. Extensive experiments demonstrate that CuckooGraph significantly outperforms the state-of-the-art in terms of query speed and memory usage.

TURNIP: This paper introduces TURNIP (short for "nondeTerministic gpU RuNtime wIth cPu offload"), a system designed to optimize AI computations on GPU servers by leveraging CPU RAM to handle memory constraints. The key innovation of TURNIP is its ability to manage the non-determinacy introduced by asynchronous memory transfers between CPU and GPU RAM, ensuring efficient computation without blocking operations. The system compiles AI computations into a dependency graph called a MEMGRAPH, allowing TURNIP to dynamically choose the best execution order at runtime based on real-time events.

Three Laws of Software Complexity: The blog post "Three Laws of Software Complexity (or: why software engineers are always grumpy)" posits three fundamental laws that lead to unnecessary complexity in software systems. The First Law states that well-designed systems inevitably degrade into badly designed ones over time, as every change either maintains or worsens the design. The Second Law suggests that complexity serves as a moat, filled by leaky abstractions, making popular systems inherently complex and hard to replace. The Third Law asserts that there is no upper limit to software complexity, as it is only constrained by human creativity and the diverse influences of multiple developers. [link]

Don't DRY Your Code Prematurely: The article "Don't DRY Your Code Prematurely" emphasizes that while the "Don't Repeat Yourself" (DRY) principle is valuable, applying it too early can lead to premature abstractions that complicate future changes. It argues that superficially similar code may serve different purposes and evolve independently, thus maintaining some duplication can be more beneficial in the early stages. The article suggests that developers should wait for clear patterns to emerge before creating abstractions, aligning with the "You Aren't Gonna Need It" (YAGNI) principle. [link]

成为数字游民，年轻人换了种“活法”: The article "Becoming Digital Nomads: Young People Are Changing Their Way of Life" discusses the trend of young people embracing flexible work arrangements such as freelancing, remote work, and becoming digital nomads. It highlights the increasing number of digital nomads who leverage the internet to earn a living while traveling the world, a lifestyle that frees them from traditional 9-to-5 office constraints. The article features personal stories of individuals like Wang Huanling and Sherry, who have chosen this lifestyle for its flexibility and opportunities for new experiences, despite the challenges it may present. [link]

Spellar AI: A useful Mac app for recording meetings and offering improvement suggestions. However, it occasionally fails to capture audio and is still in a beta phase.

PrettyProgress: This app helps me track significant dates and deadlines, displaying them as widgets on MacOS, iOS, and iPadOS.