Treno talk
Matt Mathis
9/12/95
(All documents are in postscript.)
Presentation slides
Treno examples
The examples below were not covered in the the formal presentation.
All are plots of bandwidth vs delay for successive hops along the
path. The graphs are not labeled by hop numbers, but you can count data
points from the left edge.
- Example 1: Typical treno output. The smooth
curve for hops 2, 4, 5, 6, 7 suggest common bottleneck "physics". Hops 8, 9
and 10 have an additional bottleneck, presumably between hops 7 and 8. Hop 3
appears to be "noise". Hop 1 has a particularly slow ICMP implementation (it
is visible in several of the other exaples.)
- Example 2: Treno to another supercomputing center.
The early part of the path is very similar to example 1. The last point shows
lower delay than the preceding points because it renegotiated the MTU from
4352 bytes down to 1500 bytes, lowering copy times.
- Example 3: The same as example 2, but at 1500 bytes
for the entire path. All routers are ICMP performance limited in this case:
hops 1, 2 and 9 are of similar vintage, hops 3, 4, 5, 6, 7, and 8 are newer.
- Example 4: Another nearby site with 4352 byte MTU to
all but the last hop. I am at a loss to explain why hop 2 is so much faster
than the others (This artifact also appears in a number of traces).
- Example 5: A long path (at least 3 providers) to a
geographically close site. The song and dance at hops 1 through 6 is probably
due to asymmetric routes to and across the first major interconnect. The
factor of 5 bandwidth drop between hops 8 and 9 reflect a boundary in the
infrastructure.
- Example 6: Typical well connected near site. 1500
byte MTU on the last hop.
- Example 7: Another multi-provider path. Clearly the
bottleneck is at the second to last hop, where the path enters the site.
- Example 8: The path to PREPnet (our local regional).
The first hop is within PSC, the 2nd and 3rd, via a T1 into PREPnet's NIC.
- Example 9: The path to Carnegie Mellon University.
Again the first hop is within PSC. Yes the scale really is 85 Mbit/s with 4k
packets.
- Example 10 The path to Carnegie Mellon University,
using 1500 byte packets. Yes, there are routers out there that can
send ICMP pacets at rates corresponding to 26 Mbit/s.