Question about netperf terminology in the RTL8192SE and 802.11n problem thread

Question about netperf terminology in the RTL8192SE and 802.11n problem thread

[Rick Jones]

Hi -

 From time to time I web search for instances of netperf usage, and came 
across http://www.spinics.net/lists/linux-wireless/msg76387.html  which 
has verbiage like:

> TCP_MAERTS TX Test:  47.33 53.94 55.19 44.24 57.74 55.44 53.74 54.63 47.87 57.82
> TCP_MAERTS RX Test:  66.02 69.79 67.70 52.15 82.56 80.30 79.43 80.98 76.26 71.34
> Results: TX: max 57.82, min 44.24. Mean 52.79(4.42)
>          RX: max 82.56, min 52.15. Mean 72.65(8.85)
>
> TCP_STREAM TX Test:  71.83 80.44 72.88 26.11 40.85 58.70 58.49 58.96 59.52 59.35
> TCP_STREAM RX Test:  46.41 52.64 43.85 48.44 52.15 49.66 52.81 50.61 43.18 52.93
> Results: TX: max 80.44, min 26.11. Mean 58.71(14.93)
>          RX: max 52.93, min 43.18. Mean 49.27(3.51)
>
> TCP_SENDFILE TX Test: 57.86 55.94 55.21 56.13 56.70 61.71 56.85 54.68 55.04 51.30
> TCP_SENDFILE RX Test: 37.82 47.51 41.61 42.88 45.37 35.11 45.09 40.11 46.48 22.86
> Results: TX: max 61.71, min 51.30. Mean 56.14(2.50)
>          RX: max 47.51, min 22.86. Mean 40.48(6.96)


in it. Seeing separate TX and RX lines for netperf TCP tests is 
unfamiliar to me and I was wondering if someone (Stephan?) could explain 
the split?  Netperf itself tends to emit only the one figure for a 
transfer rate (measured up at the socket level). (Modulo some of the 
recentish omni output selectors anyway, though for a TCP transfer test 
they would/should be very very similar...)

thanks, and happy benchmarking,

rick jones

Re: Question about netperf terminology in the RTL8192SE and 802.11n problem thread

[Larry Finger]

On 09/08/2011 06:54 PM, Rick Jones wrote:
> Hi -
>
>  From time to time I web search for instances of netperf usage, and came across
> http://www.spinics.net/lists/linux-wireless/msg76387.html which has verbiage like:
>
>> TCP_MAERTS TX Test: 47.33 53.94 55.19 44.24 57.74 55.44 53.74 54.63 47.87 57.82
>> TCP_MAERTS RX Test: 66.02 69.79 67.70 52.15 82.56 80.30 79.43 80.98 76.26 71.34
>> Results: TX: max 57.82, min 44.24. Mean 52.79(4.42)
>> RX: max 82.56, min 52.15. Mean 72.65(8.85)
>>
>> TCP_STREAM TX Test: 71.83 80.44 72.88 26.11 40.85 58.70 58.49 58.96 59.52 59.35
>> TCP_STREAM RX Test: 46.41 52.64 43.85 48.44 52.15 49.66 52.81 50.61 43.18 52.93
>> Results: TX: max 80.44, min 26.11. Mean 58.71(14.93)
>> RX: max 52.93, min 43.18. Mean 49.27(3.51)
>>
>> TCP_SENDFILE TX Test: 57.86 55.94 55.21 56.13 56.70 61.71 56.85 54.68 55.04 51.30
>> TCP_SENDFILE RX Test: 37.82 47.51 41.61 42.88 45.37 35.11 45.09 40.11 46.48 22.86
>> Results: TX: max 61.71, min 51.30. Mean 56.14(2.50)
>> RX: max 47.51, min 22.86. Mean 40.48(6.96)
>
>
> in it. Seeing separate TX and RX lines for netperf TCP tests is unfamiliar to me
> and I was wondering if someone (Stephan?) could explain the split? Netperf
> itself tends to emit only the one figure for a transfer rate (measured up at the
> socket level). (Modulo some of the recentish omni output selectors anyway,
> though for a TCP transfer test they would/should be very very similar...)

Those numbers are from my tests. The TX numbers are the standard netperf output 
and give the rate from my test laptop to a server that is wired to the 
AP/router. Those are the numbers that you are used to.

The RX numbers are obtained by starting a server on my laptop and ssh'ing a 
netperf command to the machine that was the server in the TX tests, i.e. I am 
measuring the TX rate from the former server, or the RX rate for the laptop. My 
script does 10 samples of each and calculates the mean and standard deviation.

Larry

Re: Question about netperf terminology in the RTL8192SE and 802.11n problem thread

[Rick Jones]

> Those numbers are from my tests. The TX numbers are the standard netperf
> output and give the rate from my test laptop to a server that is wired
> to the AP/router. Those are the numbers that you are used to.
>
> The RX numbers are obtained by starting a server on my laptop and
> ssh'ing a netperf command to the machine that was the server in the TX
> tests, i.e. I am measuring the TX rate from the former server, or the RX
> rate for the laptop. My script does 10 samples of each and calculates
> the mean and standard deviation.

Thanks.  If I have interpreted your answer correctly what you call 
TCP_STREAM RX should be the same as TCP_MAERTS TX and vice versa 
(modulo having the same -s, -S, -m and -M values or system defaults 
anyway)  Ie

system-A> netperf -H system-B -t TCP_MAERTS ...

should be the same as

system-A> ssh system-B netperf -H system-A -t TCP_STREAM

I put the TCP_MAERTS test into netperf specifically to help people avoid 
having to ssh :)

happy benchmarking,

rick jones
PS, if I or anyone else ever gets around to implementing the sendfile 
functionality in the "omni" tests, then they should also provide what we 
might call a TCP_ELIFDNES test - the netserver side calling sendfile() 
to send data to the netperf side.