[netperf-talk] concurrent sessions with Netperf 2.4.2

[Andrew Gallatin]

anil mishra writes:
 > Hi all,
 > I was trying to do some benchmarking to run the concurrent sessions. It does  not seem to run the seesions concurrently. Am I missing something?
 > 
 > I tried to run TCP_RR using Netperf and got the result as around ~18000 TPS for 1 byte request/response size which gives me latency as around ~50 usec. 
 > 
 > Then I tried to run the latency test using Iperf 2.0.2 for 1 byte data transfer and that gives me the number something like 124 KBytes/s which gives me the latency around 7.8 usec. I am not sure why so  much of difference there in Iperf and Netperf numbers.
 > Can someone help me please?

I think you are comparing apples and oranges.  How do you "run the
latency test using iperf"?  AFAIK, iperf measures streaming bandwidth
only.

Netperf -tTCP_RR is a request/response latency test.  The client sends
a request to the server, which then replies.  The client sends one
request at a time, and waits for the server to reply before it can
send the next request.  This in no way equates to running a streaming
bandwidth test with a data size of 1 byte, where the client sends 1
byte as fast as it can.  

In my experience TCP_RR yields nearly the same numbers as other
latency tests (PMB over openmpi, netpipe), so even if I'm somehow
wrong, and iperf has a ping-ping latency mode, I'd still beleive the
netperf results over the iperf results.

Speaking of TCP/UDP_RR, it would be really nice if netperf did the
math, and also output the 1/2rtt latency in microseconds in addition
to the transfers per second..

Drew