Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

From: Brian Barrett (brbarret_at_[hidden])
Date: 2006-02-25 00:41:02


On Feb 24, 2006, at 10:11 PM, Allan Menezes wrote:

> I have a 16 node AMD/P4 machine cluster running Oscar 4.2.1 Beta and
> FC4. Each machine has two Gigabit network cards. One being realtek8169
> all connected to a netgear GS116 gigabit switch with max MTU =1500 and
> the other NIC being aDlink Syskonnect chipset gigabit card
> connected to
> a managed NetgearGS724T Gigabit switch with Jumbo MTU enabled on the
> switch and each Dlink card's MTU set at 9000 from the ifcfg-eth1
> file. I
> want to know how I can use open mpi (any version >=1.01) to use
> ethernet
> bonding to get 2 gig eth cards to boost performance. I have 512MB
> memory
> on each machine for a total of 8 GigBytes Memory and a hard disk on
> each
> node.. I use HPL/Linpack to run benchmarks and get 28.36GFlops with
> open
> mpi 1.1a1br9098 and with mpich2 1.03 i get 28.7Gflops using only one
> ethernet (GIg Dlink) with jumbo frames on both benchmarks at
> MTU=9000. I
> use -mca btl tcp on openmpi. The N = 26760 and NB= 120 for HPL.dat at
> P=4 Q=4 for 16 processors for both benchmarks using openmpi and
> mpich2.
> Can any one tell me if I will get a performance increasse > 29GFlops
> using the two Switches and 2 GigEth Nic cards per node if I use
> Ethernet
> Channel Boning in FC4?

It is highly unlikely that channel bonding will increase your
performance. Given the unbalanced nature of your TCP stack, it is
likely to actually hurt performance more than it helps. Channel
bonding is really only a win for bandwidth. In most cases, it does
not improve (and sometimes even hurts) latency. HPL doesn't really
strain a network, and it's doubtful that the extra 10% (best case)
bandwith is going to help that much (especially considering the fact
that it can adversely effect latency).

Open MPI will try to use all available TCP links between hosts, so it
was already striping messages across both your NICs (assuming they
were configured at the time) round-robin short messages where
appropriate. The algorithm is fairly simplistic at this point and
assumes that all network links are created equal.

Brian

-- 
   Brian Barrett
   Open MPI developer
   http://www.open-mpi.org/