perfsonar development work

[Joe Metzger <>]

On Jul 21, 2007, at 9:43 AM, Wenji Wu wrote:

> Hi, Joe,
>
> Just based on the one-way delay, it might be hard to differentiate  
> the load
> balancing and route flapping if those paths have similar one-way  
> delays. And
> the queuing delay might easily hide the differences. We might need to
> combine with other measurement.

Some OWD tools report the TTL. Is this sufficent?


> Route flap might not happen so frequent because most routers would  
> limit the
> frequency to update the routing tables or limit the rate to  
> broadcast the
> route updates. For example, BGP router has the MRAI timer to limit  
> the route
> updates.

There are several known causes for persistent route flapping on short
timescales. There are techniques to dampen announcements caused by these
flaps but they are not all universally deployed.

> However, load balancing could happen at the packet level.
>
> It sees to me that the best way to differentiate the route flap and  
> load
> balancing might be traceroute.
>
> With traceroute, we could generate the sequence of router along the  
> path
> from the src to dest. For Route flaps, the path might not change so
> frequently. Most traceroute will generate the paths within the  
> intervals of
> route flaps.
>
> For load balancing, the paths might change at the packet levels. And
> traceroute will generate some weird results. For example, if we  
> have the
> topology like:
>     c-d-e
>    /       \            
> a-b       I
>    \     /
>     f-g-h
>
> node b will do the load balancing. We run traceroute from a to I. Then
> traceroute might generate no-existent links. Let's say, the third  
> ICMP goes
> to C, and the fourth ICMP goes to F. then non-existent link C-F  
> might be
> generated!
>
> So if it is the load balancing case, you run traceroute at  
> different times,
> various results will be generated. It depends on the crossing  
> traffic at the
> load-balancing nodes.

It is not just dependent on cross traffic. For example the Juniper  
load balancing
algorithm can be configured to balance on various parts of the packet  
header including
IP address and TCP/UDP port numbers.  But in reality it doesn't look  
at the IP addresses
and port numbers, it looks at specific offsets in the header. The TCP  
destination port
and the ICMP checksum bytes are at the same offset. So ICMP based  
traceroutes generate
confusing results.


> wenji
>
> -----Original Message-----
> From: Joe Metzger 
> [mailto:]
> Sent: Friday, July 20, 2007 4:23 PM
> To: 
>
> Cc: 
>
> Subject: Question about interpreting One Way Delay data.
>
> Hi,
> An interesting question came up in the course of debugging a
> performance problem today.
> It is not real important from an operational perspective, but it is
> intriguing from
> an academic point of view.
>
> Is there a way to analyze one way latency measurements between
> endpoints over long
> slow congested paths and differentiate between route-flapping and
> load balancing?
>
> If so, do you need access to raw data, or can you do it from the  
> type of
> summarization that the DFN and I2 tools are typically doing today?
>
> Thoughts?
>
> --Joe