Method for sharing memory among cluster system nodes on demand based on RDMA (Remote Direct Memory Access)

Abstract

The invention discloses a method for sharing a memory on demand among cluster system nodes based on RDMA (Remote Direct Memory Access). The method comprises the following steps: dividing the nodes in a cluster system into three types, namely a memory supplier, a memory requester and a self-sufficiency; a memory supplier announces memory sharing information in the cluster system, and remote shared memory is provided for other nodes by taking a block as a unit; when the memory in the memory requester is insufficient, the shared memory provided by the memory supplier is used through the RDMA, and the shared memory is mapped to a virtual address space of the memory requester by taking a block as a unit to serve as an extended memory; when the application program in the memory requester accesses the extended memory, a missing page is triggered to be interrupted, the virtual memory management subsystem performs swap-in and swap-out of the memory page, and transparent memory access of the application program is realized. The method has the characteristics of high memory access speed, simple system architecture and low software system overhead, and can improve the resource utilization rate of the cluster system and reduce the cost through memory sharing among the nodes.

Description

technical field [0001] The invention relates to the technical field of distributed storage systems, in particular to a method for sharing memory on demand between cluster system nodes based on RDMA (Remote Direct Memory Access). Background technique [0002] Memory is the most confiscated and least resilient resource in today's data center servers, and applications can only use local memory even if memory is available on another server. Such situations unnecessarily terminate workloads under memory pressure and reduce server utilization. In order to meet diverse application requirements, data centers usually configure sufficient hardware resources for nodes. With the support of virtualization and multi-tenancy technology for resource allocation centered on the CPU, processor (core) resources can usually be fully utilized, while memory allocation is generally left over. The average server memory utilization of Google and Alibaba is only 60%, while the memory utilization of ...

AI Technical Summary

Problems solved by technology

This results in a waste of resources

[0015] (2) Patent [2], a distributed high-performance computing method based on RDMA technology, patent number: CN 111382099A; in order to make memory sharing consistent, this method adopts directory control, and needs to control the free memory of each node state mark, this can achieve the purpose of memory sharing, but it still needs to build a global memory address lookup table; in addition, this method does not divide the nodes according to the actual usage of the nodes, often in a cluster system, the nodes It is mainly divided into calculation-intensive and memory-intensive. If the state directory is set for the memory of all nodes, the overhead here is very large. If it is in a cluster with millions of nodes, the situation is more complicated.

[0017] (4) Patent [4], a distributed memory file system based on RDMA, patent number: CN 108268208 A; from the method image 3 It can be seen that it is obvious that this method constructs a shared memory layout graph through service nodes, and it is necessary to divide and integrate free memory resources in all nodes. Therefore, this method is not equivalent to all nodes in the cluster. And still need to build a global memory address lookup table; in addition, this method does not divide the nodes according to the actual usage of the nodes, often in a cluster system, the nodes are mainly divided into computing-intensive and memory-intensive, if the The memory setting state directory of all nodes, the overhead here is very large, if in a cluster with millions of nodes, the situation is more complicated

[0021] (2) In related research, the server and client are often divided. Such hierarchical division itself will affect the fairness between nodes, resulting in the fact that a node as a client cannot provide free memory to other nodes.

Secondly, the server node needs to maintain a global memory address lookup table to schedule the request and release of all client memory resources, which increases the operating load of the server node, thereby reducing the overall performance and communication efficiency of the cluster system

Images