On 02:28 PM 9/3/2004, Neil wrote:
>Hey guys,
>how would you capture traces of a distributed application? Where are the best points of putting the sniffer at? And assuming that we have the traces, how would you correlate each other so we can say the WAN is the contributing factor to the latency of the app? Is there tools to do this?
>Any help/link will surely be appreciated.
This is a difficult task at best. The obvious points are at the PC, at the server (ingress/egress) and at the backend app server. But if your web server opens one or two connections to the backend apps server and streamlines all the user requests through it, you're out of luck. One solution is to use a secondary IP address on the web/app servers so that only your client will use it.
Also, keep in mind that depending on your point of capture, one side will report retransmissions while the other side will report out of sequence packets.
This can be made more difficult because you need two sides of the capture (client-wan-server) to see what's really going on. This is required because you have to know when it left the client's PC and when the packet train arrived at the server.
It's tedious w/o tools to figure out what the congestion delay is as opposed to protocol delays. The former is experience in flight while the latter happens before the packets even leave the PC.
hsb