Providing Open Architecture High Availability Solutions

For fault tolerant systems, communications among fault domains will vary depending on the

specific characteristics of each fault domain. For example, current sharing power supplies

communicate via monitoring the voltages and currents being produced or via a separate current

share signal. Even more extreme, fans may communicate by jointly pressurizing a common air

plenum that is engineered in such a way that airflow continues across all system components even

if one of the fans fails.

Other fault domains, such as processing subsystems and I/O controllers, communicate through

various data paths within the system. These may be I/O busses or point-to-point data paths.

Because they typically are non-redundant, bussed interconnects define fault domains consisting of

everything connected to a single bus, but bus failures themselves are generally very low

probability. As a result, a common compromise in the design of hardware for high availability

systems is to define a hierarchy of fault domains, as shown in Figure14.

Because of the compromises required for this sort of configuration, next-generation system

interconnects such as InfiniBand, RapidIO, or other switched fabric technologies provide higher

levels of availability with the same amount of hardware by eliminating these larger fault domains,

as shown in Figure15.

Finally, communication paths to and from the high availability system itself may need to be

redundant. In a sense, this is simply a redefinition of the system to the next higher level – i.e., to

incorporate the devices connected to the system in question into the system itself. Of more

relevance, the termination of a single data path will be either:

•to a single point, which will necessarily be contained within a single fault domain

•to multiple points within multiple fault domains

Figure 15. Redundant Switched Interconnects

I/O Card

I/O Card I/O Card

I/O Card

Host CPUHost CPU

Each box and each of the

switched networks are fault

domains in this system.

Switch