RP8240 2 Cell setup as 1 partition - SPOFs

[MIDAS]

We have a RP8240 2 Cell system with the cells setup as 1 partition.

We have been asked to identify single points of failure on the system.

It is attached to an EVA4000 - but we have that one covered (securepath, dual controllers, redundant paths etc).

It is also connected to a UPS - although there is only 1 of them... so 1 SPOF IDed here already.

Question is, that if a processor/memory module was to fail on one of the cells, would the system continue to run (without a reboot / re-partitioning of cells)...

We know the cells are connected to one system backplane - so potentially a SPOF here...

Rgds
-Webteam

[Torsten.]

several failures (e.g. double bit memory errors, failed system/pci backplane) can cause a system crash.

Thats why there is the service guard product for clustering applications.

see also

http://docs.hp.com/en/ha.html#Serviceguard

http://h71028.www7.hp.com/enterprise/cache/4174-0-0-0-121.html

[MIDAS]

Thanks Torsten. Appreciate the use of Serviceguard but the sytem is running as one physical partition so we can't cluster our apps. So...... are you saying that if one of the eight processors in the two boards (four in each) was to fail or one memory module in either cell the whole system would crash or is the rp8420 clever enough to isolate the damaged cell and limp along?

[Joshua Scott]

In the case of an unrecoverable memory fault due to the failure of multiple bits on the memory cards themselves, there would of course be no way to recover, since the data would be corrupted.

Same with the processor, most processor failures are not the processor itself, but the cache on the die. When you have a multi-bit fault in the cache, whatever data or instructions are stored there become corrupted, and the system cannot continue.

In both these cases the system does detect when this occurs, and issues a reset to protect your data from corruption.

Fortunately, these conditions are *extremely* rare on this system, and typically, the system will detect the first bit failure and deallocate that block of memory. This is what the Page Deallocation Table was designed to do. Also, when a single-bit error is detected in the Cache on a processor, the system will attempt to deactivate the processor and continue. If the processor cannot be deactivated, the system will continue to send alerts until the processor is replaced.

Josh

[MIDAS]

Thanks for the replies so far. Any ideas what would happen should an entire cell (i.e. the board) fail?

[Joshua Scott]

The entire cell board can't fail, but components on it could. The most common failures on the cell are:

damaged memory board slot
Voltage regulator module (VRM) failure
damaged cpu slot
damaged connectors

If you are not making any changes, the only failures I've seen are the VRMs. VRMs are redundant, so if one fails, the cell can continue to operate, but if a failure is detected on boot, then the cell will halt.

Josh

[Jeff Schussele]

If I recall correctly I believe the PCI power supply is an SPOF because just one unit powers both busses on an rp7410/7420 as well 2 busses each on an rp8410/8420.

Rgds,
Jeff