Wednesday November 15, 2006

The sobering graphic below was presented this morning by Lloyd Dickman from QLogic at an InfiniBand session here at Supercomputing '06 in Tampa.

The link-level bit error rate (BER) allowed by the InfiniBand standard results in some alarming overall error rates when the number of links is scaled to the level required to build large clusters using current approaches. Clearly there is value in any efforts to reduce the BER or to reduce the number of links required to build large clusters. Of course, software must ultimately be resilient to these failures, but keeping the underlying hardware error rate at manageable levels will allow applications to make forward progress at reasonable and useful rates.