hahn at physics.mcmaster.ca
Tue Nov 11 16:47:50 EST 2003
> > 1. suppose you could attach a QOS/TOS tag to small packets, and
> > have the switch give them preferential treatment. for instance,
> > if there's a congested port with a backlog, let small packets
> > "cut" the queue.
> That means that you actually have a backlog.
if you ever have two nodes sending to one node (eg gather), you will.
> Many switches, especially the
> cheap ones, have small buffers that can be filled fast, so further (small)
right, but irrelevant. the topic is "is there any point to managable
or otherwise fancy switches?"
> Furthermore, by letting packets go out-of-order, you make life harder for
> the receiver...
TCP is good at dealing with out-of-order. practically by by definition!
> > 2. the vendor claims that multicast is reliable.
> See Donald's answers to this very question in this thread:
I remember. the point is that the switch vendor claimed that full multicast
was not lossy, contradicting Becker's claim. this vendor specializes in
large, big-backplane chassis switches, so they might be right. it may be
that Don was thinking of a cluster with multiple switches.
I think latency is the real appeal of hw-supported multicast - if you
want to do a barrier across 256 nodes, do you want a ~8-deep tree of
user-level processes farming out your tinygrams (say, 8x50=400 us),
or do you want a single 30 us multicast?
> > 3. it would be neat to be able to query performance/load/queueing stats
> > from the switch on a per-port basis.
> That is actually one of the 2 things that I use out of a managed switch...
> Althought I think the information should be available as SNMP, I use it
> only when I try to find out if there is some networking problem, which is
> rarely enough, so I always used the switch CLI or web interface to do it.
which misses the point. I'd like to be able to let a user find out that
his node 12 is always bottlenecking the sim because of a problem with his
domain decomp, for instance. summary stats are not useful, only per-port
(preferably per-flow, but...)
regards, mark hahn.
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