Hostname-to-Node ID Mapping for HPC Fabric Scalability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing method for identifying Node IDs of high-bandwidth network cards in distributed-memory high-performance computers is inefficient, requiring multiple translation steps that lead to scalability issues, inconsistency, and increased computation time, especially in clusters with a large number of nodes.
Innovation Solution
Directly obtaining Node IDs from hostnames using a mapping table maintained by the fabric manager, which associates hostnames with unique logical addresses of physical cards, reducing the need for intermediate translations and enhancing scalability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple translation steps (hostname to IP address to Node ID) are used to identify network cards, then compatibility with existing systems is maintained, but the process becomes complex and time-consuming
Solution Approach 1:
The patent combines multiple translation steps (hostname to IP address, then IP address to Node ID) into a single direct mapping from hostname to Node ID. This merging eliminates the intermediate IP address translation step, reducing the number of API calls and translations required while maintaining system compatibility.
Solution Approach 2:
The patent introduces a new intermediary component (a service or library) that provides direct hostname-to-Node ID mapping. This intermediary abstracts the complex multi-step translation process, presenting a simplified interface to users while handling the underlying complexity internally.
2Reliability
If multiple translation steps are performed to obtain Node ID, then existing identification methods are preserved, but computation time increases
Solution Approach 1:
The patent performs preliminary action by pre-establishing the hostname-to-Node ID mapping relationship and making it available through a service or cached data structure. This allows direct lookup without performing multiple translation steps at runtime, significantly reducing computation time while preserving compatibility with existing identification methods.
3Ease of operation
If DNS service or local mapping files are used for hostname translation, then standard identification protocols are followed, but scalability is limited in large clusters
Solution Approach 1:
The patent adds another dimension to the identification system by introducing a specialized hostname-to-Node ID mapping service that operates alongside or replaces traditional DNS and hosts file approaches. This new dimension provides direct mapping capability that scales better with cluster size, while still allowing standard protocols to function where needed.
4Reliability
If intermediate translations through IP address are performed, then network card identification is achieved, but the number of exchanges over the interconnected network increases
Solution Approach 1:
The patent extracts and eliminates the unnecessary intermediate IP address translation step from the identification process. By directly mapping hostname to Node ID, it removes the redundant exchange that occurs when translating through IP address, improving network exchange efficiency while maintaining identification accuracy.
Data Source
AI summary
Processes pending execution in respective compute nodes (N1-Nn) of a cluster of a distributed HPC computer can communicate with one another by message exchange through an interconnected fabric. To exchange messages between the processes a method is proposed to identify the physical cards associated with the compute nodes directly from the hostname of the compute nodes as they are used in the user program. This direct identification is made from at least one mapping table (MT1) associating bijectively the hostname of each compute node of the cluster with the unique logical address of the associated physical card. This mapping table is kept in a component of the computer responsible for managing the fabric, i.e. the fabric manager (FM1). Various implementations enable scaling when accomplishing the method of message exchange between processes.