gig card v/s netperf

[Deepak Seth_1]

We have recently added GIg cards in all our servers. We used netperf tool to find the increase in our throughput after configuring the gig cards. No doubt , there is increase in speed to about 3-4 times as comapare to our older 100 Base cards. But few questions :

1. Why there is speed variation when using a netperf tool between the same class machine (L class) . One show slower speed as compare to other. Does it related to load on the machine .
2. Anyother test tools ?
3. Why it is just 3-4 times faster? I should ideally be atleast 8-9 times faster than 100 Base Card.
4. Do i need to do some tunning - some gig card paramters ...
5. Do i need to something at the switch level to set my interfaces to JUMBO frames. I tried setting the one using lanadmin -M 9000 1 command . But didn't find any differnce in results ?

Pls. suggest.

thanx in advance.

[Paula J Frazer-Campbell]

Hi

Ten times bigger does not mean ten times faster.

Your data at the cards id controled and built into ip packets by processes on the server which have a speed limitation.

External devices which also carry the data will affect the speed seen at the card, an example is Gig cards in servers with a 10 Meg hub in the middle - Extreame but you get the idea.

In perfect conditions with perfect cables, hubs, routers and data the card manufacturer can say that the card will carry data at GByte speed, but when does perfect happen.

So as always "It Depends"

Paula

[rick jones]

in and of itself, gigabit as a technology does _nothing_ to make life easier on host CPUs. it still uses the puny 1500 byte MTU.

now, if your switches and all the interfaces in the network support it, you can use jumboframes, but it has to be everywhere. the tcp connection establishment mechanisms will have 1460 used as the Maximum Segement Size if one side still has a 1500 byte MTU.

indeed system load can be a factor. if your system was say 30% CPU bound prior to switching to GbE, it is not unreasonable to see only a 3X increase in perf.

also, window size is a factor. you might try increasing the -s and -S settings in netperf and see what that does.