Shulou(Shulou.com)11/24 Report--

CTOnews.com December 21 news, in addition to the common TCP / IP protocol, RoCE protocol, Tencent Cloud (Cloud) is developing a brand-new high-performance transmission protocol HARP. With this set of protocols, the data center can easily support large-scale networking of 10000 + nodes, as well as microsecond-level "resurrection" failure networks, realizing zero broken chain.

HARP is a completely self-developed data center high performance transport protocol. HARP can provide high-availability, high-scalability and high-performance network transmission services for upper-layer applications while ensuring reliable end-to-end data packet transmission.

There are already various network transmission protocols on the market. Why does Tencent Cloud start to develop new transmission protocols from 0 to 1? The main reason is that with the continuous development of application scale and demand, the reliability and performance of data center networks are facing great challenges, and the previous protocols are gradually unable to meet the development of existing services.

First of all, in terms of reliability, the switches in the data center will inevitably have sub-health conditions, which will increase the response delay of the service or cause the service timeout to break the chain. Secondly, in terms of performance, bandwidth-sensitive applications and delay-sensitive applications are increasing day by day, and the deployment scale of data center applications is getting larger and larger, and a large number of concurrent data transmissions are often generated in aggregate communication scenarios, which puts greater requirements on the scalability of existing network transmission protocols.

Therefore, Tencent needs to redesign a new transmission protocol, which can not only meet the requirements of upper layer applications for high availability network transmission, but also provide high bandwidth and low latency data transmission services, and maintain network performance under large-scale deployment.

HARP has three main characteristics:

"High availability": Through multi-path parallel transmission and real-time link detection, the fault link is switched at the microsecond level, so that the upper layer service is completely unaware when the network fails, and there is no obvious performance jitter.

"High scalability": Through the design of shared connection and the optimization of the protocol itself, the hardware resource consumption is still maintained at a very low level under large-scale networking conditions, and the communication performance between nodes under 1w + node networking is easily supported without obvious degradation, which is greatly superior to the networking capability of traditional ROCE networks.

"High performance": Self-developed congestion control algorithm, which can achieve extremely high transmission bandwidth and stable delay (including average delay and long tail delay) under various loads without obvious dependence on network capacity.

How is it technically possible?

1. Granularity configurable shared connection, maximum support for 10000 + nodes networking

In order to meet the requirements of various service environments (bare metal, VM, CBS, etc.) and large-scale networking requirements, HARP supports connection modes with granularity such as bare connection, VM level sharing, and host level sharing. By sharing connections, HARP can significantly reduce the number of connections, reduce the resource consumption of hardware implementation, achieve a high degree of scalability, and easily support large-scale networking of 10000 + nodes.

2. Software and hardware transaction separation, providing 200Gbps maximum performance output

HARP uses a layered transaction layer and reliable transport layer design, with a clear division of labor: hardware is responsible for packet-level transactions that require efficient and reliable transmission; software provides highly flexible message processing that meets the needs of service characteristics without occupying limited hardware resources.

Finally, HARP can support the highest performance output of 200Gbps in the network scale of 10000 + nodes, which has unique value for AI training, key value storage, distributed big data applications and other scenarios.

3. Self-developed congestion control algorithm reduces the median message completion time by 35%

HARP adopts PEAD congestion control algorithm, which can accurately sense network congestion and ensure smooth network flow while maintaining high throughput. Compared to TCP, HARP reduces the median completion time by 35%, while ensuring a 90% reduction in network queuing latency for 99% of packets.

4. Deterministic multipath transmission and microsecond path switching

Tencent Cloud HARP protocol absorbs the advantages of TCP and UDP, and tracks the sending and receiving status of each message through its self-developed message numbering scheme. HARP supports out-of-order reception and selective retransmission with extremely low overhead, ensuring reliable transmission of messages.

HARP also provides high availability network services for upper layer applications through deterministic multipath transmission and microsecond path switching.

HARP uses multipath transmission within each connection, with each path having independent congestion detection capabilities. Based on the excellent network control and sensing capabilities of the self-developed congestion control algorithm, HARP's path management module will schedule according to the congestion of the path, quickly and reliably detect path failures, and re-detect a new available path within 100 microseconds to ensure that the probability of link breakage in the event of a single switch failure is 0. HARP reduces 99.9% compared to the typical reconnect recovery time (about 1s) for TCP failures.

CTOnews.com learned that HARP was originally designed for storage and high-performance computing scenarios, and has now been implemented in Tencent's Cloud Block Storage (CBS) and Elastic RDMA (EFI) businesses. As a high-performance communication base, HARP can be applied to data center services with high network performance requirements, such as AI training, key value storage, distributed big data applications, etc.

At present, engineers of Tencent Cloud (Cloud) are constantly improving its high-performance communication library, including using Socket, IB Verbs, libfabric, UCX and other communication interfaces to access HARP transmission protocol, so as to continuously expand HARP ecosystem.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.