Standard Metrics for network performance

[Steve Kadish]

Hi,

I am currently engaged in some network performance testing to attempt to demonstrate to management that problems we are seeing with an ASP application are not WAN "latency."

I know that no matter what numbers I present as far as our latency, % of dropped packets, etc., I will be asked, "well, what's normal?" I am trying to find some resources that will attempt to let me intelligently answer that question. I took a look at the SANS Reading Room and found another site on which all the papers were 5-10 years old. I'm hoping that someone on this list will have some suggestions.

Thanks,
- Steve

[Jonathan Axford]

Hmmmm, i know the feeling.

Is quite a difficult one to answer specifically though as it really depends on what else the link is being used for etc. and what is termed as acceptible performance for your network.

Is there anything else running over the link that is NOT experiencing any problems that you could use as a comparison?

I will see if i can find anything out for you as the managers here sound pretty similar!

[Olaf Borowski]

Steve,

If you don't think it's latency, it must be something else. Have you looked at the WAN utilisation? Most of the time you can poll SNMP counters to get a good idea of what is happening. If the link is peeked, you might have to implement some QoS to ensure that ASP traffic has higher priority. I personally use a network analyzer right before the traffic hits the WAN and monitor what is going on. Typically the analyzer will tell me if there is a large amount of retransmissions etc.
Regarding normal: It all depends on your application. TCP behaves different than UDP and the application itself has to account for certain things. A badly written application can have lots of problems if you experience any kind of packet loss, latency number bigger than LAN etc.
Try using Ethereal (Wireshark) and capture some traffic. It might tell you right away what's wrong. If it's a "homegrown" application, try to find out what packet loss and latency is tolerable.

[rick jones]

How long is a typical piece of string?-)

At the risk of demonstrating ignroance - "ASP application?"

If investigating "latency" I would probably start with some packet traces taken at _both_ ends. Or at least a packet trace taken at the client side, combined with a system call trace taken at the server side. The idea is to see when the client sent a request, when that request hit the server (one way network latency), when the server responded (service time) and when the response hit the client (one way network latency).

It would help if the two systems had reasonably well synchronized clocks. Syncing both to a good time source via NTP would be helpful here.

If the application is noticably faster on a LAN than on a WAN, chances are it is indeed sensitive to network latency. Either in terms of that application only having one or a very small number of "transactions" in flight at one time, or in the case of pushing lots of bits (bulk transfer) not having a large enough TCP window to meet the bandwidthXdelay product of the WAN.