Cluster Interconnect via NIC Protocol Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current high-performance computing (HPC) systems lack an efficient method for interconnecting multiple clusters, which hinders the ability to perform high-performance inter-cluster data exchanges and run the same application across multiple clusters effectively.
Innovation Solution
The solution involves configuring each cluster with Ethernet-based high-performance interconnection protocols, such as BXI or Infiniband, and implementing specific routing tables and address resolution protocols in the network cards to enable direct communication between clusters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If protocol translation gateways are used to interconnect multiple clusters, then inter-cluster communication is enabled, but device complexity and performance are worsened due to data copying and protocol translation overhead
Solution Approach 1:
The patent extracts the protocol translation function from the gateway device and relocates it to the network interface cards (NICs) of the computing nodes. Each NIC is configured with routing tables that enable direct protocol translation at the network card level, eliminating the need for complex gateway devices that copy and translate data between clusters. This extraction principle resolves the contradiction by maintaining inter-cluster communication capability while removing the gateway complexity overhead.
Solution Approach 2:
The network interface cards are configured to perform protocol translation and routing functions independently without requiring external gateway devices. The NICs use locally stored routing tables to directly translate between intra-cluster protocols (like BXI) and inter-cluster protocols (like IP), enabling self-service protocol conversion that eliminates the performance penalty and complexity of traditional gateway-based approaches.
2Adaptability or versatility
If protocol translation gateways are used to interconnect multiple clusters, then inter-cluster communication is enabled, but productivity is worsened due to unacceptable performance for running applications across clusters
Solution Approach 1:
The patent extracts the protocol translation function from the gateway device and relocates it to the network interface cards (NICs) of the computing nodes. Each NIC is configured with routing tables that enable direct protocol translation at the network card level, eliminating the need for complex gateway devices that copy and translate data between clusters. This extraction principle resolves the contradiction by maintaining inter-cluster communication capability while removing the gateway complexity overhead.
Solution Approach 2:
The patent implements direct protocol translation at the network card level, allowing data to skip the intermediate copying and translation steps that occur at gateway devices. The NICs perform inline protocol conversion between intra-cluster protocols (BXI) and inter-cluster protocols (IP), enabling data to rush through the translation process without the performance penalty of gateway-mediated communication, thus maintaining acceptable application performance across clusters.
3Speed
If specialized BXI networks are used within clusters, then high-performance intra-cluster communication is achieved, but adaptability is worsened by incompatibility with general-purpose switches and routers
Solution Approach 1:
The patent introduces protocol translation capability at the network interface card level as an intermediary mechanism. The NICs translate between the specialized BXI protocol used within clusters and standard IP protocols used on general-purpose networks. This intermediary translation at the NIC level allows high-speed BXI communication within clusters while enabling compatibility with standard IP networks and general-purpose switches, resolving the adaptability contradiction.
Solution Approach 2:
The patent changes the protocol parameter at the network interface card level, allowing the same physical network infrastructure to operate with different protocols depending on the communication context. Within clusters, the NICs use BXI protocol for high-speed communication, while when communicating with external networks, the NICs translate to IP protocol. This parameter change capability enables both high intra-cluster speed and external adaptability without requiring specialized hardware throughout the entire network.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
One aspect of the invention relates to a high-performance computer comprising a plurality of clusters (21,22) interconnected by an IP network (2), each cluster (21,22) comprising: - At least one Ethernet gateway (215,225) configured to transmit data between the cluster (21,22) and the IP network (2) storing at least one first routing table comprising at least, for each other cluster of the plurality of clusters (21,22), an association of an address of a gateway (215,225) with a destination IP address included in the cluster (21,22) comprising the gateway (215,225), - A plurality of computing and/or storage nodes (N,211,212,221,222), each node (N,211,212,221,222): o being configured to execute at least one instance of a high-performance computing and/or storage application,o comprising at least one network card (NIC1) implementing a high-performance interconnection protocol based on Ethernet and being configured to implement an address resolution protocol, the network card (NIC1) storing at least one second routing table comprising at least, for each other cluster of the plurality of clusters (21,22), an association of an address of a gateway (215) of the first cluster (21) with the identifier of the other cluster (22) accessible from the gateway (215) of the first cluster (21), o storing at least one third routing table comprising at least, for each other cluster of the plurality of clusters (21,22), an association of an identifier of an instance of the application with an identifier of the cluster comprising the node (N,211,212,221,222) executing the instance of the application, with a unique network identifier of the instance of the application and with an IP address of the network card (NIC1) of the node (N,211,212,221,222) running the instance of the application, At least one intra-cluster interconnect switch (214,224) configured to connect each node (N,211,212,221,222) and the gateway (215,225).,