China Mobile and Partners Release Full Scheduling Ethernet Technology and Commercial Products
TapTechNews September 30th news, during the China Computing Power Conference last week, China Mobile, in collaboration with over 50 global partners, released the Full Scheduling Ethernet (GSE) complete technical standards, as well as the world's first set of GSE commercial products.
Ethernet has become the development direction of the new generation of intelligent computing center network technology. In May 2023, China Mobile, in conjunction with more than 10 Chinese enterprises, took the lead in publishing the Full Scheduling Ethernet Technology Architecture (GSE) white paper, and in August at the China Computing Power Conference, it launched the GSE Promotion Plan. The members include over 50 industry-university-research institutions such as domestic and foreign mainstream cloud service providers, equipment manufacturers, chip manufacturers, and universities, and it has become one of the two influential technical systems globally along with the Ultra Ethernet Consortium (UEC) led by American companies.
In order to take into account the rhythm of intelligent computing center construction and the needs of technological development, the Full Scheduling Ethernet technology is divided into two commercial stages, GSE1.0 and GSE2.0, and TapTechNews provides the following introduction:
GSE1.0 maximally supports new GSE technologies based on existing chips to optimize network performance. It has currently achieved the first commercial use in the intelligent computing center (Harbin) of China Mobile with over ten thousand cards cluster, reducing the communication time ratio in the training process by more than 20%, reaching the international advanced level.
GSE2.0 completely renovates the underlying forwarding mechanism and upper protocol stack of Ethernet. Through original technological innovations such as multi-path spraying based on PKTC, congestion control mechanism based on DGSQ, and fault detection and notification based on 66B atomic code blocks, it fundamentally solves the problems of traditional lossless Ethernet performance and reliability.
In addition, in order to meet the networking needs of both GPU servers with integrated network cards represented by Huawei Ascend 910 series and GPU servers with independent network cards represented by Nvidia H800 series, GSE constructed two major technical scenarios, GSE-N2N and GSE-E2E. The former realizes all GSE functions in the network to achieve natural decoupling of computing and network, and the latter extends some GSE capabilities to the server network cards to realize cluster interconnection through end-network coordination.
The achievements released this time mainly include the complete technical standards of GSE, including the
The GSE1.0 technical standard mainly covers core technologies such as port-level + computing-network coordination load balancing and end-network coordination congestion-aware authorization, and has been commercially used in a large scale;
The GSE-N2N technical standard mainly defines core technologies such as multi-path spraying based on PKTC, congestion control based on DGSQ, control plane design, network robustness, and network visualization. The commercial switch products released this time have been fully realized;
The GSE-E2E technical standard mainly invo lves core contents such as end-network coordination mechanism, multi-path spraying mechanism, congestion control based on DGSQ, out-of-order reordering on the end side, and rapid packet loss retransmission. The commercial network card chip has been completed.
The GSE commercial switch released this time supports a kilocard-level networking scale, and the networking performance is claimed to be more than 50% higher than that of the traditional RoCEv2 switch.
