commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gilles Sadowski <gil...@harfang.homelinux.org>
Subject Re: [Math] FastMath preset tables
Date Wed, 14 Sep 2011 21:14:42 GMT
On Wed, Sep 14, 2011 at 08:02:20PM +0000, Ted Dunning wrote:
> On Wed, Sep 14, 2011 at 7:57 PM, Phil Steitz <phil.steitz@gmail.com> wrote:
> 
> > > No.  But it is easy to code up some warmup in the startup part of the
> > life
> > > cycle.  That should be a penalty paid once when the server starts, not
> > later
> > > on the first request.  In the worst case, you can build a simple startup
> > > script that starts the server and then hits a few URL's to get the
> > classes
> > > loaded.  You can also do soft start with your load balancer.
> >
> > Yeah and watch it fall over when you bounce it under load or when
> > you are trying to dynamically manage load by quickly starting
> > instances to respond to load surges.  Startup time can be important
> > in online applications.  Luc mentioned other examples as well.
> >
> 
> I have helped design and build a top 25 web site.  I understand these
> issues.  This isn't a big deal and the original comment about web-sites
> launching a JVM and reloading classes several times a second is complete off
> base.
> 
> 
> > >
> > > Seriously.  This is a total red herring.
> >
> > Not to the users who reported the problem.  We have a fix.

Yes, you have a fix, but not to the issue that was reported (see my original
post).

> I am +1
> > on the code in trunk and ++1 on ending this discussion.

-1
Please prove that 100 ms does matter in a real use-case; without imagining
that you can crash a web server with the initialization of "FastMath"...

Not precluding that for some people it might be "nice" to be as fast as
possible, just for the sake of it; if so, the tables should nevertheless
be separate from the "real" source code (cf. previous post).

> 
> Read what *I* wrote.  I was talking about the supposed multi times per
> second JVM restart "issue".  That isn't a real issue.
> 
> As I pointed out in the next paragraph, a long startup for an interactive
> app is an issue, but there seems to be some confusion about how much
> difference this change will make.

Exactly, I say that 100 ms is not "long". Even when launching a command-line
"application" such as the mini-benchmark used to time a single function
call, I get the output "fast enough" (I hit enter and it's there).


Gilles

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org


Mime
View raw message