Solved: Load and response time

yves.haemmerli · ‎02-07-2005

This question is related to my previous append. The CSS documentation says that the CSS calculates the average response time for a service, based on the measured lifetime of flows to that service. What does this really mean ?

For instance, in a persistent http connection, does the CSS measure individual response time between HTTP GET requests and their corresponding answer ? Same question for an FTP flow, does the CSS measure respnse time between data packets and ACK ?

Thank you

Yves

Gilles Dufour · ‎02-09-2005

it is the flow duration and nothing else.

And it does not need to be smarter.

The reason is that if you use the same method for all servers, then make the average response time for all connections, and then compare the services to the best response time [and not a predefine value] we can get a good approximation of server load.

Because the distribution of short and long connections should be the same on all servers, if you take the average response time per server and compare it with other servers, you define which one is the fastest and which is the lowest.

Computing a response time for every GET, would require the CSS to always inspect all traffic.

This would have a negative impact on performances.

We inspect the data only for L5 rule with persistent reblance configured.

Also, the response time is computed by the CSS in order to then define the "load" on each server.

Only the load value should be used.

You should not try to make statistics on the response time.

Regards,

Gilles.

View solution in original post

Gilles Dufour · ‎02-08-2005

Yves,

it's good to ask yourself questions about how things work.

But the load is really useless unless you use ACA loadbalancing [and you should not] or GSLB.

So, we can dig into the code to find out what is done exactly, but how useful would it be ?

What is your concern with the load ?

Gilles.

yves.haemmerli · ‎02-08-2005

Hi Gilles,

Thank you for your answer.

Actually, I am working in a very large multinational customer project and I am studying solutions, based on the GSS/CSS aready in place in the network, to provide Global Load Balancing for SAP portals among four regional Data Centers around the world (Intranet only). The SAP team in the project wrote a comprehensive set of requirements. On of these is to provide a good and realistic image of each portal load status, in order to be able to direct user requests to the best portal, regardless of its geographic position. Therefore, I am trying to understand the details of what the CSS does.

In addtion, I read the DFP chapter, but here too, there is few or no information on the way load measurement is performed. Does Cisco recommend a third party DFP agent software ?

As I also designed and tested Data Center proximity in the Internet using the GSS, I want to design the global SAP portal for my customer with the same technology...else, the customer will evaluate a non-Cisco solution...

Can you give me some informations ?

Thank you for your support

Yves (IBM Switzerland)

Gilles Dufour · ‎02-09-2005

Yves,

the internal specification indicates the following

ResponseTime – response time in milliseconds indicating response time measured from flow set-up to flow teardown.

So it's not per request but from SYN to FIN/RESET.

Also the value shown [refering to your previous question] should indicate a value in msec even if it was normalized [I'm not sure what the normalize function does].

Regards,

Gilles.

yves.haemmerli · ‎02-09-2005

Hi Gilles,

Thanks again for your answer. However it does not make sense for me. The measured value in this case has nothing to do with a "response time", but with the duration of a flow !

Imagine, we have http 1.1 TCP sessions with multiple GET requests within the session. If the CSS measures the time between the SYN and the FIN, he actually measures the duration of the http session, not the response time the user is experiencing. Hence, the CSS does not measure the server load in this case!

In fact, I suspect that the CSS is doing something more sophisticated, but what ? In the documentation, one can read that "the CSS is creating a teardown report when flows are being torn down". OK, but what information are present in the report ? for sure not only the flow duration.

I am sorry to insist, but our project depends on this information...

Thank you for your appreciated support

Yves

Gilles Dufour · ‎02-09-2005

it is the flow duration and nothing else.

And it does not need to be smarter.

The reason is that if you use the same method for all servers, then make the average response time for all connections, and then compare the services to the best response time [and not a predefine value] we can get a good approximation of server load.

Because the distribution of short and long connections should be the same on all servers, if you take the average response time per server and compare it with other servers, you define which one is the fastest and which is the lowest.

Computing a response time for every GET, would require the CSS to always inspect all traffic.

This would have a negative impact on performances.

We inspect the data only for L5 rule with persistent reblance configured.

Also, the response time is computed by the CSS in order to then define the "load" on each server.

Only the load value should be used.

You should not try to make statistics on the response time.

Regards,

Gilles.

yves.haemmerli · ‎02-09-2005

Gilles,

OK, it makes more sense now. Effectively, if we assume that the distribution of length of requested objects is equal on all servers in the server farm, than the average session duration should be the same for equally loaded servers.

The term "response time" confused me actually. And the two last colomns in the show load command could be eliminated from the screen output as it is also confusing and useless...

I thank you very much for your well apreciated support

Yves