Mailing-List: contact geronimo-dev-help@incubator.apache.org; run by ezmlm
Precedence: bulk
Reply-To: geronimo-dev@incubator.apache.org
Date: Sat, 6 Sep 2003 01:26:15 -0500
Subject: Re: [WishList] [Deployment] [EJB]
Content-Type: text/plain; charset=US-ASCII; format=flowed
Mime-Version: 1.0 (Apple Message framework v552)
From: Dain Sundstrom <dain@coredevelopers.net>
To: geronimo-dev@incubator.apache.org
Content-Transfer-Encoding: 7bit
In-Reply-To: <3F593902.8000304@tackline.demon.co.uk>
Message-Id: <04B95802-E033-11D7-B8F0-000393DB559A@coredevelopers.net>

On Friday, September 5, 2003, at 08:31 PM, Thomas Hawtin wrote:

> Dain Sundstrom wrote:
>> On Thursday, September 4, 2003, at 02:01 AM, Sergei Batiuk wrote:
>> That is not true.  A normal Java method invocation on an Object 
>> reference (not an interface reference) takes about 3 nanoseconds on 
>> my 1g mac.  A Java reflection invocation takes about 350 nanoseconds 
>> == 0.000 350 ms.  My guess is you thing that is slow (and so do I).  
>> Instead of using Java reflection we will be using proxies generated 
>> by cglib which supports reflective style calls using a MethodProxy 
>> object.  A reflection invocation using MethodProxy on my 1g mac takes 
>> 7 nanoseconds.
>
> I have some issues with your timings there. Are you using the client 
> HotSpot? I get for invocation only around 40 cycles with -server and 
> 440 with -client. For lookup and invocation I get 3,200 with -server 
> and 13,500 -client. Timings on a PII/266 running Linux and Sun JRE 
> 1.4.2, and obviously take microbenchmarks with a pinch of salt. Loops 
> with conventional virtual method invocations can be completely 
> optimised away.

The idea is to get hot spot to optimize the loop so we only measure the 
cost of an invocation, but if you want to play with it I can send you 
the code.  I really don't understand what you code does, but mine times 
somewhere 10,000 and 1,000,000 iterations.  The number doesn't matter 
as long as it takes a while.  For there I calculate the number of 
nanoseconds per operation.

> Something to be wary of is start up times. IIRC, (server) HotSpot (and 
> I assume equivalent in other modern JREs) compile proxies for Method. 
> Obviously they don't necessarily have to go through byte code, so 
> start up times may be significantly faster.

Startup time is really not a big concern in a long running server 
(within reason).  We just need to be sure that all of the byte code 
generation happens during the deployment phase.

> I'm not saying cglib decisions are wrong, just making sure they're 
> made with appropriate information...

Agree.

-dain

/*************************
  * Dain Sundstrom
  * Partner
  * Core Developers Network
  *************************/