AW: mulitcast copy or snowball copy
becker at scyld.com
Mon Aug 18 16:00:04 EDT 2003
On 18 Aug 2003, Mitchell Skinner wrote:
> On Mon, 2003-08-18 at 09:50, Donald Becker wrote:
> > This is costly. "Open loop" multicast protocols work by having the
> > receiver track the missing blocks, and requesting (or interpolating)
> > them later. Here you are discarding that information and doing much
> > extra work on both the sending and receiving side by later locating the
> > missing blocks.
> There's an ietf working group on reliable multicast that wasn't making a
> whole lot of progress the last time I checked.
It's a hard problem, and when they agree on a protocol it likely won't
apply to clusters.
The packet loss characteristic and cost trade-off is much different on a
WAN than with a local Ethernet switch on a cluster. On a WAN every
packet is costly to transport, so it's worth having both end stations
doing extensive computations.
On a cluster we might talk about doing more computation to avoid
communication, but that's only for a few applications. In reality we
prefer to do minimal work. Thus we prefer OS-bypass for application
communication, and kernel-only for file system I/O. Notice the
attention given to zero copy, TCP offload, TOE/TSO and sendfile().
Multicast and packet FEC add exactly what people are trying to avoid,
extra copying, complexity and work.
Donald Becker becker at scyld.com
Scyld Computing Corporation http://www.scyld.com
914 Bay Ridge Road, Suite 220 Scyld Beowulf cluster system
Annapolis MD 21403 410-990-9993
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf