Pager 3 - Cuterwrite's Blog

👋 Welcome to Cuterwrite 's Blog

High Performance Computing

RDMA: Address Handle

This article is reprinted from Zhihu column: 8. RDMA Address Handle, author: Savir. AH stands for Address Handle. In service types such as UD, the local end uses AH to specify the remote node.

High Performance Computing

RDMA: Protection Domain

This article is reprinted from Zhihu Column: 7. RDMA Protection Domain, author: Savir. To isolate various RDMA resources, the protocol introduces the concept of PD. PD stands for Protection Domain, which "groups" related software resources together, and resources between different protection domains cannot access each other.

High Performance Computing

RDMA: Memory Region

This article is reprinted from Zhihu Column: 6. RDMA Memory Region, Author: Savir. The essence of network communication is the migration of data in storage media, and RDMA manages memory through MR. MR is a special memory region, and this article introduces its concept and function.

Comprehensive Tech Stack

Record: Install Intel® OneAPI-2024.0

OneAPI is an open and standards-based specification that supports multiple architecture types, including but not limited to GPU, CPU, and FPGA. The specification has direct programming and API-based programming paradigms. Intel® OneAPI is a software suite that includes various tools and libraries to support the OneAPI specification. This article documents the process of installing Intel® OneAPI-2024.0 on CentOS 7.6.

Research Related

Notes: Pure - Improving Message Passing to Better Utilize Intra-Node Shared Memory

Pure is a new programming model and runtime system designed to fully leverage shared memory within nodes in environments based on the Message Passing Interface (enhancing the utilization of idle core capabilities through tasks). Pure utilizes shared memory in two ways: (1) allowing ranks to steal work from each other while waiting for messages to arrive; (2) enabling high-performance message passing and collective operations between processes within a node using efficient lock-free data structures. Researchers evaluated the key message passing and collective features of Pure through micro benchmark tests and demonstrated that in CoMD molecular dynamics and miniAMR adaptive mesh refinement applications, Pure can achieve up to 2.1x application acceleration when scaling to 4096 ranks.