Enabling GOMAXPROCS for jujud

[John Arbash Meinel]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 2013-10-29 18:58, Andrew Deane wrote:
> 
> I try never to avoid important features because they might be
> scary.
> 
> 
> Oh, I am saying use it, but just test it (like you said with the
> race detector enabled) and know the profile, before rolling it
> out.
> 
> 
> 
> Also, just for the record, buffered channels will block just like 
> unbuffered channels once they're full, so you can't count on that 
> saving you from blocking.  Also why I almost never use buffered 
> channels. They're actually much more difficult to reason about
> than unbuffered channels that just always block.
> 
> 
> As the docs say, unbuffered for semaphore wait, and buffered for 
> producer/consumer. What I have seen in the past is that maxing the 
> number of CPUs will effect the way in which an application
> containing unbuffered channels behaves. Just pointing it out, thats
> all.
> 

In scale testing up to an m3.2xlarge (8-cores) and using the 'use max
cpus' I was only ever able to get up to about 600% CPU. And that was
with some caching put in place for Addresses, before that I peaked at
around 400%. But I hit 400% w/ m3.xlarge (4-cores).

Anyway, I think if we can find a reasonable way to set up state
servers to allow for GOMAXPROCS>1 we'll see good results for the
scaling story.

I did end up seeing the system go into failure mode when adding the
nrpe-external-master charm (2 a system with 5000 units of Ubuntu
running). Unfortunately, I haven't been able to figure out why *all*
the agents (even machine agents) go into the 'down' state.

So it might be premature to do today, but I definitely think it is
something we'll want soon.

John
=:->

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.13 (Cygwin)
Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/

iEYEARECAAYFAlJwhNQACgkQJdeBCYSNAANEggCgonTAAOvNSPPyafN+5dqOH2go
1OsAoNLWjVuDPY2rq38pAFSa5X7xy2w7
=z4JK
-----END PGP SIGNATURE-----