rbw at ahpcrc.org
Wed Sep 25 12:55:11 PDT 2002
Joachim Worringen wrote:
>Do you have an example for this? AFAIK, the hop latency for each SCI node
>in-between is a few ns, and with a 128 node cluster in 3D topology, you would
>never pass more than about 15 nodes. Numbers from a real, busy cluster would
Yes very much ...
But I do not ... and now think I spoke prematurely based on the logical
idea that hopping costs would be subtantial. As you have seen, others
have data from relatively quiet networks/rings of 8 nodes and have
extrapolations to 8x8x8 systems that show minimal latency growth. So
unlike available bandwidth per node, SCI latency appears relatively
stable as the number of nodes per ring grows under these test loads.
That latency should grow in a high traffic environment is logical,
but as I have proven once already "logic is the art of going wrong
with confidence" ... so I will defer to those with data if there
Glad to see the product experts chiming in with better numbers and
hope my interconnect table decreased rather than increased the amount
of entropy on the subject of interconnects here.
More information about the Beowulf