[Beowulf] interconnect and compiler ?
patrick at myri.com
Fri Jan 30 13:39:38 PST 2009
Greg Lindahl wrote:
> that in practice using multiple VLs will suffer from significant
> negative effects due to implementation details. Does anyone know of a
> proof point of this?
In practice, the per-port amount of buffering in the switch crossbars is
not big enough for multiple VLs (buffer == die space == $$$). So, if you
configure multiple VLs, you divide the buffer and you cannot reach line
rate in any single VL. You can cheat a bit by not supporting long cable,
but it only goes so far. Oh, and buffer size requirement grows with the
link bandwidth, because speed of light is constant.
This is different from per-priority flow-control, because the buffering
can be done at the end, where it's cheaper.
QoS in the switches has always been a giant hoax, starting from the ATM
days. Beyond admission control, the only thing you can reliably do is
decide what to drop first :-)
More information about the Beowulf