Re: Iperf max b/w
I've done testing with the 64pin sysKonnect Gige card and on completely
. WinXP systems - > 900 mbps TCP iperf transfer
. RedHat 9.0 Linux systems > 940 ,bps UDP iperf transfer
The 64 pin nic allowed alot more "headroom" on the Gige nics I tested.
Matthew J Zekauskas wrote:
> I was recently doing some gigE testing ... 940 Mb/s was a
> number I saw with 1500 byte MTU hosts; you may have to go
> to 9000 byte MTUs to get 990.
> Note that for all of these there must be no loss, with the
> bottleneck being the host interface; if you ever get
> into TCP's "sawtooth" pattern, the maximum you'll achieve
> is ~75% of the raw bitrate and probably much less.
> [FWIW, I've done this with a syskonnect gigE card & 64 bit
> 66 Mhz PCI bus machine; PCI-X should give you more headroom.]
> --On Friday, October 10, 2003 9:50 AM -0500 Tony Rimovsky
> <tony --at-- ncsa.uiuc.edu> wrote:
>> there could be several host bottlenecks:
>> 1. pci bus speed.
>> 2. other host activity.
>> 3. frame size and io/cpu utilization
>> 4. tcp implimentation problems
>> --On Friday, October 10, 2003 8:36 AM -0600 Cameron Harr
>> <charr --at-- lnxi.com> wrote:
>>> Tony Rimovsky wrote:
>>>> It may be a host limitation. We've seen iperf perform faster than
>>>> (990 Mbps) on high latency gigabit paths with larger windows.
>>> Ok. For my aid, when you say host, are you thinking PCI-X bridge to
>>> nic (hardware)?
>>> Cameron Harr
>>> Systems Engineer
>>> Linux Networx Inc
Bob Riddle (bdr --at-- internet2.edu) Technologist,Internet2
3025 Boardwalk, Suite 100 Ann Arbor, Michigan 48108
Business Phone: 734.913.4257 Fax Number: 734.913.4255
"Opportunity is missed by most people because it is dressed in overalls and looks like work" Thomas Edison