jkubicky@nntp-server.caltech.edu (Joseph J. Kubicky) (06/04/91)
I posted an article about the Touchstone Delta the other day and I have received several letters requesting additional information and pointing out some potential problems. I talked to one of the people in the router group here and he gave me some better information about the inter-node latency. Basically, the 75-125ns figure is wrong. The node-to-node jump latency is around 50ns. So, for the entire mesh, the maximum transport latency is around 2us. However, it turns out that the OS that each node runs accounts for considerably more delay than this. I get the impression that, as the software is now written, we're talking at least 10us or so to actually get a message somewhere (although, as I also pointed out, each node can multi-task, so this delay at least need not contribute to wasted processing time). However, the theoretical lower limit (that is, that imposed by the harware), is something like 2us (this also ignores delays in message delivery due to messages already being delivered, but the OS automatically breaks big messages up into little ones so this is minimized). More information is availble in the 6th Annual MIT Conference on Advanced Research in VLSI proceedings (maybe best to key on Advanced Research in VLSI). Look for and article by C. Seitz called something like "Let's Route Packets Instead of Wires". Jay Kubicky jkubicky@cobalt.cco.caltech.edu
tve@sprite.berkeley.edu (Thorsten von Eicken) (06/04/91)
In article <1991Jun3.230539.7162@nntp-server.caltech.edu> jkubicky@nntp-server.caltech.edu (Joseph J. Kubicky) writes: > I get the impression that, as the software is now written, we're talking > at least 10us or so to actually get a message somewhere (although, as I > also pointed out, each node can multi-task, so this delay at least need > not contribute to wasted processing time). So, $$$ question: how long does it *really* take to ping-pong a message back and forth between two processors? I.e. A sends to B, B receives and sends back, A receives and sends back, etc.. Measuring one round-trip. I bet it's >100us. By the way, the OS delay costs you: the processor *is* busy. No multi-tasking changes anything to that. Thorsten von Eicken (tve@sprite.berkeley.edu)