Coordination Point Devices for Cluster Split Brain Arbitration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for handling network partitions in computer clusters, such as I/O fencing, are inefficient and limited, particularly with SCSI-3 Persistent Reservation, which requires expensive SAN infrastructure and fails to properly handle coordination point device failures or inaccessible scenarios, leading to data corruption and split brain issues.
Innovation Solution
The use of coordination point devices to process arbitration requests and select client computers for continued membership in the cluster based on established connections and sub-cluster membership, employing an arbitration technique to determine which nodes to fence during I/O fencing, ensuring that the larger sub-cluster survives by prioritizing the client with the earliest request or largest sub-cluster size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If SCSI-3 Persistent Reservation is used for I/O fencing, then arbitration capability is improved, but device complexity and infrastructure cost increase due to requiring expensive SAN infrastructure and third coordinator disk
Solution Approach 1:
The patent extracts the arbitration function from the complex SCSI-3 PR infrastructure and relocates it to existing coordination point devices that are already part of the cluster. This eliminates the need for dedicated third coordinator disks and expensive SAN infrastructure while maintaining the essential arbitration capability for split-brain prevention.
Solution Approach 2:
The coordination point devices, which originally served other cluster coordination functions, are made multi-functional by adding I/O fencing arbitration capabilities. This allows existing infrastructure to handle multiple roles including arbitration, reducing the need for specialized dedicated components.
2Ease of operation
If coordination point devices are used for I/O fencing, then ease of operation is improved, but reliability deteriorates when coordination point devices fail or become inaccessible
Solution Approach 1:
The patent implements preliminary action by having coordination point devices record arbitration state information in persistent storage before potential failures occur. This pre-recorded state information enables the system to recover and continue arbitration operations even when coordination point devices become inaccessible or fail.
Solution Approach 2:
The system implements feedback mechanisms where client computers report their arbitration requests and coordination point devices track the state information. This continuous feedback loop allows the system to detect failures and trigger appropriate recovery actions, maintaining reliability despite coordination point device issues.
3Speed
If arbitration race is used to determine cluster partition, then speed is improved, but fairness deteriorates as smaller sub-clusters may win over larger sub-clusters
Solution Approach 1:
The patent applies preliminary action by having client computers record their arbitration request timestamps before the arbitration race concludes. The coordination point device uses these pre-recorded timestamps to determine the winner, ensuring that the first requester wins while maintaining fast arbitration speed. This prevents smaller sub-clusters from unfairly winning due to timing advantages.
Data Source
AI summary
A method and apparatus for partitioning a computer cluster through coordination point devices that provide enhanced split brain arbitration during I/O fencing. In one embodiment, a method for using coordination point devices to support Input/Output fencing in a computer cluster comprises processing at least one arbitration request from at least one client computer upon an occurrence of a split brain within a computer cluster, wherein each arbitration request indicates at least one established connection between at least one coordination point device and each client computer of the at least one client computer and each client sub-cluster membership, wherein the at least one coordination point device form a coordination point cluster membership and selecting a client computer of the at least one client computer to continue to be a portion of the computer cluster based on the arbitration request.