Interactions with mxge, pf, nfsd, and the kernel
Hans Petter Selasky
hps at selasky.org
Wed Jul 9 14:53:54 UTC 2014
On 07/09/14 16:47, Bob Healey wrote:
> These machines primary purpose is to provide nfs services to an HPC
> cluster (12 hosts, 384 cores) on a gigabit interconnection. mxge0 is a
> 10G link to an HP Procurve 2910al with the hpc nodes connected to it,
> existing in RFC 1918 space. bce1 is connected to the public network.
> bce0 is only used for accessing the IPMI chipset. 9K packets are
> enabled in what appears to be a futile attempt to improve nfs and mpi
> performance within the cluster. I can revert back to 1500 byte packets,
> and just chalk this up to being things mere mortals are not to meddle with.
>
> Bob Healey
> Systems Administrator
> Biocomputation and Bioinformatics Constellation
> and Molecularium
> healer at rpi.edu
> (518) 276-4407
>
Hi,
Have you tried 4K packets instead of 9K ones? I understand the mbufs are
special and must be contiguous in physical memory, so memory allocation
fragmentation can lead up to starvation of mbufs.
--HPS
More information about the freebsd-stable
mailing list