[netperf-talk] Add an option to pause before starting tests

[Rick Jones]

Andrew Gallatin wrote:
> On 09/14/10 18:01, Rick Jones wrote:
> 
>> Folks -
>>
>> I have adapted another patch from Google. This time, a global -s
>> <seconds> option is added which will cause an omni test to pause for
>> <seconds> after setting-up the test but before actually starting it.
>> This is used for situations where one might be starting many, Many, MANY
>> concurrent netperfs and is an attempt to avoid them utterly obliterating
>> one another.
> 
> 
> Cool.  But with a delay, won't they sill be a bit out of
> sync due to fork/exec/connect latencies? I almost wonder
> if it might be almost as useful to  accept an absolute time,
> rather than a delay, so that they all actually start at the
> same time.

Agreed, it is far from a complete solution.  IMO, one still has to (ab)use the 
confidence intervals support to minimize skew error (which itself is not a 
complete solution).  It is my understanding that this feature is helpful in even 
just getting many, Many, MANY netperfs started in the first place.

> When I was writing a script to use multiple netperfs, I was
> thinking about hacking up some barrier (like waiting on a
> named Posix semaphore, listening on a multicast socket, etc).
> For my purposes, the startup delay didn't really matter, so
> I never got around to it.  I applaud Google for doing something,
> which is more than I did ;)

I've wondered about using a multicast socket as well, and worried about packet 
losses etc.  I suspect the one true answer is to get a broadly usable netperf 
with explicit test synchronization going a la netperf4.

happy benchmarking,

rick