Enhancing reliability and robustness of a cluster

• US 6,766,470 B1
• Filed: 03/29/2000
• Issued: 07/20/2004
• Est. Priority Date: 03/29/2000
• Status: Expired due to Fees

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Shah, Rajesh R.
• Primary Examiner(s): Beausoliel, Robert
• Assistant Examiner(s): MASKULINSKI, MICHAEL C
• Application Number:
  US09/537,394
• Time in Patent Office: 1,574 Days
• Field of Search: 714/41, 714/707, 714/4, 714/9, 714/13, 714/43, 714/56, 700/28, 375/296, 382/156, 709/324, 719/324
• US Class Current: 714/4.2
• CPC Class Codes: G06F 11/0709 in a distributed system consisting of a plurality of standalone computer nodes, e.g. clusters, client-server systems
• CPC Class Codes: G06F 11/0793 Remedial or corrective actions recovery from an exception in an instruction pipeline G06F9/3861; by retry G06F11/1402; for recovering from a failure of a protocol instance or entity H04L69/40
• CPC Class Codes: G06F 11/1443 Transmit or communication errors
• CPC Class Codes: G06F 11/2005 using redundant communication controllers

Abstract

Reliability and robustness of a cluster having a host connected thereto via a cluster interconnection fabric may be enhanced by determining if an error condition exists in an I/O controller connected to the host via the cluster interconnection fabric by attempting to communicate with it a first predetermined time period after an inquiry by an operating system as to whether or not an I/O controller driver stack should be unloaded and commanding the operating system to unload the I/O controller driver stack upon a determination that the error condition still exists. The determination as to whether the error condition still exists may be repeated a predetermined number of times prior to commanding the unloading of the I/O controller driver stack upon a determination that the error condition still exists. Furthermore, a determination may be made to determine if the error condition still exists in the I/O controller by attempting to communicate with it a predetermined period of time after the I/O controller driver stack has been unloaded by the operating system in response to the command to unload and commanding the operating system to reload the I/O controller driver stack upon a determination that the error condition no longer exists. The above-noted additional determination may also be repeated a predetermined number of times upon the determination that the error condition still exists.

Claims

1. A host in a cluster comprising a cluster fabric and one or more I/O controllers attached to the cluster fabric, the host comprising:

an operating system (OS); and

a fabric control driver included in the operating system (OS) and configured to;

to create, when each cluster adapter is installed to access one or more fabric-attached I/O controllers, a virtual fabric control device object representing connectivity to the cluster fabric, and one or more I/O controller device objects for the corresponding fabric-attached I/O controllers that constitute “

child”

device objects of the virtual fabric control device object;

to destroy affected I/O controller device objects, while maintaining the virtual fabric control device object, when one cluster adapter malfunctions but another cluster adapter remains available; and

alternatively, to destroy the affected I/O control device objects and the virtual fabric control device object, when all connectivity to the cluster fabric is lost.

2. The host as claimed claim 1, wherein the “

child”

device objects each has an associated I/O controller device driver that is loaded into host memory to access resources available on the corresponding fabric-attached I/O controller.

3. The host as claimed in claim 1, further comprising at least two cluster adapter hardware, each being associated with a respective one of channels to connect the host to a cluster fabric.

4. The host as claimed in claim 2, further comprising at least two cluster adapter hardware, each being associated with a respective one of channels to connect the host to a cluster fabric.

5. A method of operating a host in a cluster comprising a cluster fabric and one or more I/O controllers attached to the cluster fabric to eliminate a cluster adapter as a single point of failure in the cluster host, the method comprising:

creating, when each cluster adapter is installed to access one or more fabric-attached I/O controllers, a virtual fabric control device object representing connectivity to the cluster fabric, and one or more I/O controller device objects for the corresponding fabric-attached I/O controllers that constitute “

child”

device objects of the virtual fabric control device object;

destroying affected I/O controller device objects, while maintaining the virtual fabric control device object, when one cluster adapter malfunctions but another cluster adapter remains available; and

alternatively, destroying the affected I/O control device objects and the virtual fabric control device object, when all connectivity to the cluster fabric is lost.

6. The method as claimed in claim 5, wherein the virtual fabric control device object is destroyed by the fabric control driver when all connectivity between the host to the cluster fabric is lost.