[rabbitmq-discuss] High(er) latency with 1.5.1?

[Holger Hoffstätte]

For your Sunday afternoon delight..

Holger Hoffstätte wrote:
> Matthias Radestock wrote:
>> The scalability improvement in 1.5.x affect latency somewhat, but not by
>> a factor of 3. And I cannot think of any other changes off the top of my
>> head that would cause this.
> 
> hmm..okay. I ran the test on two very different systems (dual-core Windows
> laptop and rusty single-core Linux box) and both showed essentially the
> same numbers, which made me curious. I still had the results from the
> previous runs back when we found the Nagle bug, and those were on the
> order of 280,320,500us with a Nagle-disabled 1.4 rabbit.

As it turns out the higher latency was simply caused by my laptop deciding
to power the CPU cores down by a factor of 2.. <:-)
Switching the system to "Maximum Performance" immediately lowered the
numbers, to the order of 350/450/600us. Running the Java client with
-server -XX:CompileThreshold=100 brought it further down to the previous
290/350/550us; anything beyond that is probably thread/OS scheduling
noise, considering that this is Windows and the broker runs on the same
host too. So all is well.

cheerz
Holger