Multistage cache distributed key value storage system based on programmable switch

Abstract

The invention provides a multi-level cache distributed key value storage system based on a programmable switch. The multi-level cache distributed key value storage system comprises a client, a server, the programmable switch and a controller. And the client, the server and the controller are all connected with the programmable switch. According to the invention, the programmable switch and the server are combined to form the multi-level cache distributed key value storage system, so that the data storage space is expanded from the on-chip memory of the programmable switch to the memory of the server, the respective advantages of the switch and the server are fully utilized, and the throughput and the availability of the system are greatly improved. Key value items of any size are stored in a data plane of the switch, a data packet is processed, data are stored in different positions of a register in a scattered mode by means of the thought of fragmentation, the data are processed by means of the mirroring and circulating functions of the programmable switch, the effectiveness of caching in a network is improved, and the programmable switch can store more complex objects.

Description

technical field [0001] The invention belongs to the technical field of distributed storage, in particular to a multi-level cache distributed key-value storage system based on a programmable switch. Background technique [0002] The distributed key-value storage (Key-Value Storage KVS) system is an important foundation of modern Internet services, such as social networking, e-commerce, etc. The KVS system caches frequently accessed data in memory to meet the high throughput and low latency requirements brought by billions of requests. A key challenge of distributed KVS systems is balancing skewed and dynamic workloads over time. During a fixed time interval, only a small number of entries called hotspots are frequently accessed. For example, about 50% to 90% of access requests come from 1% of key-value items. Without careful design, this load imbalance can overutilize a small number of servers, with the server's CPU becoming the bottleneck, resulting in decreased throughpu...

AI Technical Summary

Problems solved by technology

While providing effective load balancing, the processing power of programmable switches has not been fully exploited and utilized, the performance of the system will be much lower under highly skewed data, and it is not suitable for production with multiple switches and server racks data center

Images