>>>>> "tc" == Tim Cook <[email protected]> writes:
tc> Channeling Ethernet will not make it any faster. Each
tc> individual connection will be limited to 1gbit. iSCSI with
tc> mpxio may work, nfs will not.
well...probably you will run into this problem, but it's not
necessarily totally unsolved.
I am just regurgitating this list again, but:
need to include L4 port number in the hash:
http://www.cisco.com/en/US/products/ps9336/products_tech_note09186a0080a963a9.shtml#eclb
port-channel load-balance mixed -- for L2 etherchannels
mls ip cef load-sharing full -- for L3 routing (OSPF ECMP)
nexus makes all this more complicated. there are a few ways that
seem they'd be able to accomplish ECMP:
FTag flow markers in ``FabricPath'' L2 forwarding
LISP
MPLS
the basic scheme is that the L4 hash is performed only by the edge
router and used to calculate a label. The routing protocol will
either do per-hop ECMP (FabricPath / IS-IS) or possibly some kind of
per-entire-path ECMP for LISP and MPLS. unfortunately I don't
understand these tools well enoguh to lead you further, but if
you're not using infiniband and want to do >10way ECMP this is
probably where you need to look.
http://bugs.opensolaris.org/bugdatabase/view_bug.do?bug_id=6817942
feature added in snv_117, NFS client connections can be spread over multiple
TCP connections
When rpcmod:clnt_max_conns is set to a value > 1
however Even though the server is free to return data on different
connections, [it does not seem to choose to actually do so] --
6696163 fixed snv_117
nfs:nfs3_max_threads=32
in /etc/system, which changes the default 8 async threads per mount to
32. This is especially helpful for NFS over 10Gb and sun4v
this stuff gets your NFS traffic onto multiple TCP circuits, which
is the same thing iSCSI multipath would accomplish. From there, you
still need to do the cisco/??? stuff above to get TCP circuits
spread across physical paths.
http://virtualgeek.typepad.com/virtual_geek/2009/06/a-multivendor-post-to-help-our-mutual-nfs-customers-using-vmware.html
-- suspect. it advises ``just buy 10gig'' but many other places
say 10G NIC's don't perform well in real multi-core machines
unless you have at least as many TCP streams as cores, which is
honestly kind of obvious. lego-netadmin bias.
pgputFUSXDRds.pgp
Description: PGP signature
_______________________________________________ zfs-discuss mailing list [email protected] http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
