[net-perf] Performance test results

[Michael Sinatra]

Mark Boolootian wrote:
>>FWIW, these are typical maximum results I've seen with 9000 and
>>1500 byte MTU's, respectively, when there is no loss and no
>>other bottleneck.  (I suspect that with faster CPUs (& buses?)
>>the 1500 byte results will get better...)
> 
> 
> I thought someone told me (Mike Van Norman?) they had run full rate
> gigE with 1500 byte packets.  I haven't seen the CPU running out of
> gas with small packets and have puzzled over where the problem is.  You
> mention the bus as a possible bottleneck.  32 bit/33MHz PCI can run a 
> sustained rate of 115 MBytes/sec, enough to nearly saturate a gigE 
> connection.  Where would the bus limitation come from?  
> 
> SysKonnect has a white paper at 
> 
>   http://www.syskonnect.com/technology/64bit66MHz_Computing.PDF
> 
> with a table showing theoretical data rate vs burst length.  Small
> burst length can produce much lower throughput, but I don't know
> how to interpret this in terms of packet size.  It doesn't look
> relevant to me. 

Yeah, I can't really make out what they're trying to say, at least not 
at first glance.

One observation is that at 1500-byte MTU, saturation level appears to be 
941 mb/s of data on a gig link and 94.1 mb/s of data on a fastE link. 
In other words, when iperf is reporting 941 mb/s of data throughput (not 
counting protocol overhead), the router/switch interface is reporting 
1000 mb/s utilization.  (Strangely, in the iperf 2.02, I am also seeing 
up to 943 mb/s on the same link, so perhaps it's doing slightly 
different math in the later version.)

michael