Storage glitches fell Australian supercomputers
DDN and SGI find fix after faulty firmware fingered for Fornax fail
Supercomputers at two Australian research organisations have experienced substantial downtime after glitches hit their storage area networks.
iVEC’s incident report for the outage says the machine was shut down on December 17th to allow the update of firmware in a “SGI DDN IS16000 storage unit.” The update produced symptoms including disks going missing from storage pools and loss of availability of some data. Instructions from the vendors concerned saw the arrays go through a lengthy “force verify process” followed by a RAID rebuild in some cases.
The rebuild started on December 18th and finished on the 20th, when it became possible to restart Fornax.
eResearch South Australia’s Tizard  and Corvus  supers, which offer 40 TFlops and 6 TFlops apiece, went down on December 27th and were coaxed back to life on January 3rd, a period covering three business days.
The organisation described the outage as follows on its (since updated) system maintenance page:
“During the holiday break our primary storage system locked up causing the Tizard and Corvus HPC systems to become unresponsive. At the moment there are a large number of orphaned jobs on the compute nodes of both HPC systems due to the outage.”
The culprit for the outage was not the Dell Compellent array that serves as shared NAS for the two supers, but a Dell server that sits at a critical juncture in the InfiniBand network linking the supers to their storage. Simon Brennan, a computer systems specialist at eResearch SA, told The Reg the server was overwhelmed by the quantity of data coursing through it, crashed and was restored once staff resumed from their Christmas break.
Brennan said the server at the heart of the issue was a beta, and will shortly be completely rebuilt. ®