For small packets, it's around 200 usecs, or .2 milliseconds. He's only
10x off. As the packet size goes up, he gets closer and closer to be
correct.
: Consider, the Avalon which is interconnected with switched 100Mb/s
: achieved enough performance (after a memory upgrade) to reach the 50th
: fastest computer in the world at a total price of around what? $280K?
That's great. There are a ton of examples like this but they don't really
detract from his point. If you have an embaressingly parallel app, then
you could also make the claim that 10mbit ethernet or even parallel ports
are fast enough. That's nice. But it isn't in the space where people who
actually have apps want to run.
I'm sitting here working on a contract for LLNL right now, involving
clusters. Want to know what they want? 10,000 CPUs, with node to node
round trip latency of 10 usecs. Think about that for a while. We can't
do that with point to point ethernet today. Now add to the mix that you
will have to go through a stack of switches and back again, each of which
will add at least 1usec of latency per hop (in fact, I'd love to find
switches that good).
The difference between the hand wavers and the people who actually spend
money on the problem is substantial. I think Richard was trying to point
out that the people who are a little more serious about this stuff are also
quite sensitive to latencies.
That doesn't mean forget about it, it just means be realistic about what
the goals are. DIPC doesn't come close to being useful for the traditional
clustered market. Nowhere near.
Now ask yourself this: if DIPC is for cluster users and cluster users don't
like it, why are we arguing about it?
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/