CN-121984907-A - End network cooperative reliable multicast method, system and computer equipment

CN121984907ACN 121984907 ACN121984907 ACN 121984907ACN-121984907-A

Abstract

The invention discloses a method and a system for multicasting by end network coordination and reliability, and relates to the technical field of artificial intelligent distributed computing and high-performance network communication. The method specifically comprises the steps that a sender analyzes a target expert set of the gate-controlled output of a hybrid expert model, constructs a layered multicast bitmap comprising a network node bitmap and a local expert bitmap, and encapsulates the layered multicast bitmap in the target expert set The switching equipment analyzes the network node bitmap in the message expansion head, and performs parallel bitwise AND operation with a locally preset port bitmap mask, determines a target port set to be forwarded, and performs message copying and path pruning. The invention aims to solve the problem of highly dynamic target set distribution in a mixed expert model scene through a source bitmap coding and a switch hardware pruning mechanism, and obviously reduce system overhead and long tail delay while ensuring high-reliability transmission.

Inventors

TAN XIAOBIN
ZENG YINGYING
XU NING
YUAN SHENZHI
YANG FENG
ZHENG QUAN

Assignees

中国科学技术大学

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. The method is applied to a logic multicast tree formed by a transmitting end, a switching device and a receiving end, and comprises the following steps: the method comprises the steps that a sending end analyzes a target expert set output by a mixed expert model gate to construct a layered multicast bitmap, wherein the layered multicast bitmap comprises a network node bitmap for identifying the node position of a target physical machine and a local expert bitmap for identifying the position of the target expert in a single physical machine node; the transmitting end encapsulates the layered multicast bitmap in The message expansion header is sent to a physical network; The switching equipment analyzes the network node bitmap in the message expansion head, carries out parallel bitwise AND operation on the network node bitmap and a local preset port-bit mapping mask, determines a target port set to be forwarded, carries out message copying aiming at each selected target port, and rewrites the bitmap of the copy message by utilizing the bitwise AND operation result so as to carry out hop-by-hop path pruning.
2. The method for coordinated reliable multicast of an end network according to claim 1, further comprising the step of path optimization: the sending end obtains network topology information and pre-constructs a plurality of candidate multicast paths covering expert nodes, and analyzes the remote measurement including in-band network returned from each candidate multicast path in real time Extracting link index of each hop exchanger, comparing each branch under candidate multicast path Index, regard the most serious branch state of congestion degree as the bottleneck congestion grade of this route, and choose the optimum route label to pack in the said current transmission task according to this The message extension header.
3. The method for coordinated reliable multicast of an end network according to claim 2, wherein the first step The first candidate multicast path Link quality index of a branch link Calculating by adopting a normalization function: Wherein, the Is the first The number of hops traversed by a branch link, Is the first The depth of the queues of the hop switches, For the maximum cache capacity of the switch, Is a preset congestion sensitivity factor.
4. The method of claim 2 wherein the sender maintains a candidate path state table, the state table records path identifiers, real-time congestion scores, and last updated time stamps for each candidate multicast path, and calculates the effective scores for each candidate multicast path before selecting the optimal path identifiers : Wherein, the For the minimum value of all the branch link quality indicators extracted under the corresponding path, For the predetermined coefficient of restitution, For the current time period of time, The last updated timestamp for the corresponding path identification.
5. The method for coordinated reliable multicast of an end network according to claim 1, further comprising the step of directional retransmission: the receiving end monitors the packet loss or disorder state and feeds back the range negative confirmation message containing the node identification of the receiving end ; The transmitting end receives the information And then, initiating directional data restoration to the receiving end by rewriting bitmap information.
6. The method for coordinated reliable multicast of end networks according to claim 5, wherein said sender assigns a triplet identifier to each physical packet : Wherein, the For the identification of micro-batches of data, For the identification of the logical data block(s), Is the packet sequence number of the physical data packet within the corresponding logical data block.
7. The method of claim 6, wherein the receiving end maintains a receiving sliding window for each logical data block and monitors a packet loss state by using a packet loss timer, wherein the packet loss timer is started if a packet sequence number received by the receiving end is not equal to an expected sequence number, and the corresponding sequence number set is determined to be a determined lost set if the packet loss timer expires and the missing sequence number is not full.
8. The method for end-network cooperative reliable multicast according to claim 1, wherein the step of constructing the hierarchical multicast bitmap comprises: the transmitting end maps each global expert mark in the target expert set into a physical addressing binary group through an expert catalog table Wherein For the identification of the physical machine node, Numbering specific experts; The sending end constructs the network node bitmap according to the set of the physical machine node identifiers, constructs the local expert bitmap according to the specific expert numbers in each physical machine node, and encapsulates the local expert bitmap in the network node bitmap The message extension header.
9. An end network cooperative reliable multicast system, comprising: a transmitting end for constructing a layered multicast bitmap comprising a network node bitmap and a local expert bitmap, and encapsulating the layered multicast bitmap in Transmitting in a message extension head; the switching equipment is used for analyzing the network node bitmap in the message expansion header, executing parallel bitwise AND operation with a locally preset port-bit mapping mask, determining a target port and executing message copying and path pruning; A receiving end for receiving the message and feeding back a negative acknowledgement message containing a node identifier to the transmitting end when packet loss is detected 。
10. A computing device comprising a processor and a memory storing a computer program that when executed by the processor performs the steps of the method of any of claims 1 to 8.

Description

End network cooperative reliable multicast method, system and computer equipment Technical Field The invention relates to the technical field of artificial intelligent distributed computing and high-performance network communication, in particular to a method, a system and computer equipment for end network cooperative and reliable multicasting. Background With the continuous increase of the large language model scale, a single computing device is difficult to bear massive parameters, so that the mixed expert model architecture is widely adopted. In the distributed training and reasoning scene of the hybrid expert model, the gate control network dynamically distributes target experts according to token characteristics, so that data transmission presents high randomness and sparsity, and communication traffic severely fluctuates in destination node combination and burst strength. The existing intelligent computing center mostly adopts a remote direct memory access protocol based on Ethernet, namely a remote direct memory access protocol (RDMA over Converged Ethernet, roCE) based on enhanced Ethernet, but the reliable connection service of the protocol is established on the basis of point-to-point semantics, and a serial number management and retransmission mechanism of the protocol cannot be directly applied to one-to-many multicast scenes. Existing native multicast schemes face severe switch state maintenance pressures when handling dynamically changing target node sets and lack fine-grained congestion awareness. In the traditional scheme, the memory bandwidth load of a transmitting end is overhigh by multi-time unicast analog multicasting, and the application layer multicasting adds extra kernel context switching overhead and increases the hop count of a transmission path, so that microsecond low-delay requirements cannot be met. In addition, the existing network congestion control algorithm generally depends on the feedback of the receiving end to adjust the sending rate, so that the congestion degree of different branch links in the packet broadcast tree cannot be distinguished, and the bandwidth utilization rate of non-congested branches is easily reduced. Therefore, how to realize efficient and reliable multicast transmission and congestion control facing to the hybrid expert model becomes a technical problem to be solved urgently. Disclosure of Invention The invention mainly aims to provide a method, a system and computer equipment for end network cooperative reliable multicast, and aims to realize efficient and reliable multicast transmission and congestion control for a hybrid expert model. In order to achieve the above objective, the present invention provides a method for coordinated reliable multicast of an end network, where the method is applied to a logical multicast tree formed by a sending end, a switching device and a receiving end, and includes: the method comprises the steps that a sending end analyzes a target expert set output by a mixed expert model gate to construct a layered multicast bitmap, wherein the layered multicast bitmap comprises a network node bitmap for identifying the node position of a target physical machine and a local expert bitmap for identifying the position of the target expert in a single physical machine node; the transmitting end encapsulates the layered multicast bitmap in The message expansion header is sent to a physical network; The switching equipment analyzes the network node bitmap in the message expansion head, carries out parallel bitwise AND operation on the network node bitmap and a local preset port-bit mapping mask, determines a target port set to be forwarded, carries out message copying aiming at each selected target port, and rewrites the bitmap of the copy message by utilizing the bitwise AND operation result so as to carry out hop-by-hop path pruning. Preferably, the method further comprises the step of path optimization: the sending end obtains network topology information and pre-constructs a plurality of candidate multicast paths covering expert nodes, and analyzes the remote measurement including in-band network returned from each candidate multicast path in real time Extracting link index of each hop exchanger, comparing each branch under candidate multicast pathIndex, regard the most serious branch state of congestion degree as the bottleneck congestion grade of this route, and choose the optimum route label to pack in the said current transmission task according to thisThe message extension header. Preferably, the firstThe first candidate multicast pathLink quality index of a branch linkCalculating by adopting a normalization function: Wherein, the Is the firstThe number of hops traversed by a branch link,Is the firstThe depth of the queues of the hop switches,For the maximum cache capacity of the switch,Is a preset congestion sensitivity factor. Preferably, the sender maintains a candidate path state table, which records path identifie