[Beowulf] very low performance for very small packets under MPICH (TCP_NODELAY?)

Mark Hahn hahn at physics.mcmaster.ca
Fri Dec 30 11:11:35 PST 2005


> 2.6.15) with MPI 1.2.7 and Gigabit ethernet with a 3COM Switch and
> 3C2000-T NIC cards. We detected a very low efficiency in communication

which appears to use the sk98lin driver.

> packets or so times are in the order of 0.03~secs. This degrades the
> average performance for very small packets by a factor of 100.

is this measurable on real code, or only benchmarks?  the reason I ask
is that real code probably doesn't generate a flood of tinygrams,
but the benchmark might, and if so would trigger interrupt mitigation. 

> Any pointers for understanding why the Nagle algorithm is still active

why have you jumped to the conclusion that Nagle is at fault here?




More information about the Beowulf mailing list