We present a formalized, fully decentralized runtime semantics for a core subset of ABS, a language and framework for modelling distributed object-oriented systems. The semantics incorporates an abstract graph representation of a network infrastructure, with network endpoints represented as graph nodes, and links as arcs with buffers, corresponding to OSI layer 2 interconnects. The key problem we wish to address is how to allocate computational tasks to nodes so that certain performance objectives are met. To this end, we use the semantics as a foundation for performing network-adaptive task execution via object migration between nodes. Adaptability is analyzed in terms of three Quality of Service objectives: node load, arc load and message latency. We have implemented the key parts of our semantics in a simulator and evaluated how well objectives are achieved for some application-relevant choices of network topology, migration procedure and ABS program. The evaluation suggests that it is feasible in a decentralized setting to continually meet both the objective of a node-balanced task allocation and make headway towards minimizing communication, and thus arc load and message latency.
BlueFog: Make Decentralized Algorithms Practical for Optimization and Deep Learning
University of California, Los Angeles(加利福尼亚大学洛杉矶分校);Google Inc.(谷歌公司);DAMO Academy, Alibaba Group(阿里巴巴集团达摩院);University of California, Santa Barbara(加利福尼亚大学圣巴巴拉分校)
FusionLLM: A Decentralized LLM Training System on Geo-distributed GPUs with Adaptive Compression
The Hong Kong University of Science and Technology(香港科技大学);The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));Hong Kong Baptist University(香港浸会大学);A*STAR(新加坡科技研究局);Northeastern University(东北大学);Harbin Institute of Technology, Shenzhen(哈尔滨工业大学(深圳));XGRIDS;National University of Singapore(新加坡国立大学)
DMAPF: A Decentralized and Distributed Solver for Multi-Agent Path Finding Problem with Obstacles
多智能体路径寻找(MAPF)是一个问题,找到一个移动的序列,以达到他们的分配位置没有冲突的代理。集中式算法通常会给出最优解,但在不采用各种技术的情况下很难扩展-通常会牺牲最优性;但解决代理数量超过一千的MAPF问题仍然是一个挑战。为了解决可扩展性问题,我们提出了DMAPF -一个分散和分布式的MAPF求解器,这是我们最近发表的工作,ros-dmapf的延续。我们解决了ros-dmapf的问题,它(i)只适用于没有障碍的地图;和(ii)有一个低的成功率与密集的地图。给定一个MAPF问题,ros-dmapf和DMAPF都在空间上将地图划分为子问题,但后者进一步将每个子问题划分为称为区域的不连通区域。每个子问题被分配给一个分布式求解器,然后单独创建一个抽象的计划-一个代理需要访问的区域序列-为每个代理,并交错代理迁移与移动规划。Answer Set Programming以其在小而复杂的问题中的性能而闻名,用于许多部分,包括问题划分,抽象规划,迁移的边界分配和移动规划。机器人操作系统用于促进求解器之间的通信,并使其有机会与机器人系统集成。DMAPF在求解器之间引入了一种新的交互协议,以及一些机制,这些机制共同导致了更高的成功率和更好的解决方案质量,而不会牺牲太多的性能。我们实现了DMAPF,并通过与其他国家的最先进的MAPF求解器进行比较,实验验证,结果表明,我们的系统实现了更好的可扩展…