geronimo-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "John C. Dale" <>
Subject Re: Reflection Bad, OO and direct Method invocation Good...
Date Sat, 09 Aug 2003 13:39:54 GMT
Microbenchmarks can in fact be misleading when we aren't talking about 
as much as a factor of 10!  Reflection, in the best case scenario below, 
was a factor of 10 slower!  The bottom line is that reflection takes a 
lot more electricity than does non-reflection.

You are absolutely right in that the round trip to the database is going 
to consume more time for the life of the user request than is a sequence 
of reflection-based calls.  That will always be the bottleneck in 
systems that don't have any kind of intelligent middle-tier cache - but 
wait, there's more!

Entity EJBS are just that (a cache) aren't they?  Through intelligent 
caching and good design of our applications, we're able to avoid the 
extra knocks on the database if the entity of our type with our PK has 
already been loaded.  This, in my mind, is one of the chief advantages 
of using Entity Beans; throughput.  This throughput will definitely be 
impacted negatively if there are several reflection calls for each user 
request as the system comes under (and maintains) load.  Furthermore, 
although the bottleneck is clearly the round trip to the database for 
the hydration/storage of data, the time it takes to get to the point of 
JDBC query invocation will be effected negatively if reflection is used 
- this effect becomes more pronounced as the system comes under heavy 
load.  The sooner the JDBC query starts, the sooner it will finish.

I would also note that 'optimizing away' performance issues due to 
reflection (assuming that the core of the system is based on reflection) 
is MUCH easier to type than it is to implement.  If we see the train 
coming now, why don't we get out of the


John C. Dale
Professional Services

Aaron Mulder wrote:

>	First of all, can we agree that microbenchmarks are dangerous?
>	But all that aside, try 1 million HTTP requests, or 1 million
>iterations of a JDBC query, and then provide a ratio between JDBC/HTTP
>time and reflection time.  If you can perform a million requests of either
>type in 30 seconds I'd be shocked, and that's two orders of magnitude.
>	IMHO, reflection is something that can be optimized away later,
>once it becomes the bottleneck on performance.  I don't think we need to 
>avoid it from the outset, because I don't think it will be the bottleneck 
>at the outset.
>On Sat, 9 Aug 2003, John C. Dale wrote:
>>I'm concerned about your mention of reflection.  I really like 
>>reflection, and think it has many useful applications in the application 
>>development space, but should be kept out of any runtime application 
>>that has stringent scalability and performance requirements.  Although 
>>reflection allows an incredibly generic way to solve many difficult 
>>problems, and really opens the door to completely grandios perfectly 
>>idealistic genericititousness ;)  , a straight-forward OO approach to 
>>solving the problems would 1) be faster and 2) be easier to read and 
>>maintain.  I've seen reflection based approaches work, and I've seen 
>>them bomb.
>>One might argue that in a container based system that will eventually 
>>support clustering and scalability doesn't have to worry about savoring 
>>individual CPU cycles that might be expended needlessly with a 
>>reflection-based design.  Just add more hardware and the user experience 
>>will be perpetually preserved.  I would argue, however, that the 
>>development community will be able to make better use of the product if 
>>the rate at which additional hardware purchases becomes necessary 
>>decreases.  The philosophy, IMHO, should be to solve the problems at 
>>hand with straight-forward OO solutions, and not to focus so much on 
>>being generic and supporting n * infinity solutions up-front.  Using 
>>things like the Policy/Strategy pattern at well-selected locations will 
>>afford the opportunity for pluggability without compromising 
>>maintainability and performance.
>>Here are some dramatic results:
>>public void init()
>>   customer = new Customer();
>>public void run()
>>*Interface (Polymorphism):*
>>public void init()
>>  customer = new Customer();
>>public void run()
>>  Person person = (Person) customer;
>>public void init()
>>  customer = new Customer();
>>  try 
>>  {
>>      method = customer.getClass().getMethod("work", new Class[0]);
>>  } 
>>  catch (Exception e) 
>>  {
>>  }
>>   params = new Object[0];
>>public void run() 
>>  try 
>>  {
>>    method.invoke(customer, params);
>>  } 
>>  catch (Exception e) 
>>  {
>>  }
>>With 1000000 of the above code, here were the results:
>>JDK DirectTest InterfaceTest ReflectionTest
>>*Sun 1.4* 52 ms 54 ms 543 ms
>>*Sun 1.4 -server* 26 ms 56 ms 279 ms
>>*Sun 1.3* 124 ms 128 ms 2168 ms
>>*Sun 1.3 -server* 41 ms 58 ms 2012 ms
>>*IBM 1.3* 75 ms 78 ms 2134 ms
>>Reflection will significantly effect the performance of the system and 
>>should be avoided for any runtime operations.  Where it should and could 
>>definitely be applied is with the dynamic generation and compilation of 
>>code, or generation of metadata.  Otherwise, even at the recommendation 
>>of Sun, it should be avoided for runtime operations.
>>John C. Dale
>>Leo Simons wrote:
>>>Jason Dillon wrote:
>>>>PS. Can someone write up something about the current state of the 
>>>>major component containers out there with a feature blurb... no soap 
>>>>boxes, just the facts jack.
>>>What, no soap boxes? How on earth can anyone comply with that? ;) No 
>>>you weren't asking me anyway, were you. Ah, e-mail already typed. Bummer.
>>>= My Opinion =
>>>You should *not* be evaluating component containers. You should save 
>>>this discussion
>>>for a later date and just code your way to 1.0. The basic design idea 
>>>below should show that most of the architectural concepts behind all 
>>>these containers are
>>>very similar. I've been experimenting (no, I will not plug it, you 
>>>will just get confused) with
>>>some reflection that will allow any component written for any of the 
>>>below to run in any
>>>of the other containers, and that is feasible, straightforward and 
>>>So write your components to plug in whatever you have, do a nice IoC, 
>>>SoC, SAI, AOP
>>>setup, and you will be able to defer refactoring around an external 
>>>container until much later.
>>>But that's my opinion, and I have now said it three times, and your a 
>>>responsible adult
>>>(yep, its a guess, you could also be 11 years old :D). Switching 
>>>soapbox mode off.
>>>= Disclaimer =
>>>Comparing component containers is comparing apples with pears. Avalon 
>>>is by far the
>>>biggest 'generic' project at the moment, for example, but recent 
>>>developments utilize AOP
>>>and interceptor architecture to support a much 'lighter' 
>>>container-component api and
>>>contract. Indeed, picocontainer was started by an avalon elder from 
>>>the firm belief that
>>>things should be simpler and smaller. So to actually evaluate all this 
>>>stuff, you really should
>>>spend a day or so delving into the websites and the code of all these 
>>>projects, and backing
>>>tech like nanning ( and aspectj 
>>>( I would start
>>>by looking at pico/nano and xwork, then take a look at the tutorials 
>>>for avalon-merlin.
>>>The other projects have a smaller community atm, and I tend to value 
>>>community size
>>>and vibe.
>>>Furthermore, I have strong opinions about stuff, and allegations to 
>>>various projects, hence
>>>this is not an objective overview, even though I tried to make it 
>>>somewhat objective.
>>>= Features/ design idea shorthands =
>>>IoC = Inversion of Control, the idea that an application is controlled 
>>>from the top down
>>>SoC = Seperation of Coccerns, the idea that a class (aspect) should do 
>>>one job and do it well
>>>SAI = Seperation of API from Implementation, the idea that you define 
>>>and code to work
>>>   interfaces
>>>AOP = Aspect Oriented Programming, mostly lightweight nowadays where 
>>>you add a chain
>>>   of interceptors around a method call that can handle orthogonal 
>>>DecP = Declarative Programming, where you use a declarative-style 
>>>language (usually xml) to
>>>   determine things like component wiring (ie your average tomcat 
>>>config file, generalized)
>>>EBP = Event Based Programming, basically making the inter-object 
>>>method call asynchronous
>>>   and encapsulating such a call into some kind of event object that 
>>>can be queued, modfied,
>>>   etc
>>>= No particular order, incomplete list =
>>> - IoC, SoC, SAI, AOP, DecP, 
>>>EBP. Nearing
>>>1.0 release. Used in webwork2 (a competitor to struts). EBP very basic 
>>>only. Lean and mean,
>>>but not mature and some client-server web-layer specific assumptions. 
>>>Don't like the XXXAware
>>>interfaces. Very vibrant and active community and many famous peeps 
>>>with J2EE experience
>>>around at opensymphony.
>>> and -
>>>IoC, SoC, SAI, AOP, DecP. 1.0 beta releases. Lean and mean and very 
>>>extensible and
>>>embeddable, developed by smart XP peeps, some stuff already in use in 
>>>some apps, but
>>>otherwise pretty much alpha. I love picocontainer and the way they're 
>>>doing the project.
>>>The dev community is intentionally kept small atm, but many peeps are 
>>>watching this one.
>>> - IoC, SoC, SAI, DecP. Container supporting
>>>avalon-framework components. Used to be 
>>>yet-another-novel-avalon-container, but I
>>>think they're growing to be container-component-contract-agnostic. 
>>>Corporate backed
>>>development. Smart guys, not so much focussed on releases as on 
>>>getting all the
>>>functionality they need (which is a lot) in place. Very much a 
>>>pragmatic project.
>>> - Not yet public container development
>>>(dubbed 'loom' IIRC) by a smart ex-avaloner 'n others. Haven't seen 
>>>any code yet
>>>but my guess is it'll be good. The website says "move along" so you 
>>>prolly should.
>>> - haven't looked at in too much depth. 
>>>Seems similar to
>>>nanocontainer and xwork. Think it has one active developer and a beta 
>>>release. Some
>>>smart points made, but too much xml for my taste. Hoping to see some 
>>>of this rolled into
>>>Xwork and/or pico.
>>> - IoC, SoC, SAI, DecP, EBP (EBP for fortress 
>>>only). By far
>>>the oldest 'generic container' project. Big committer base, mature 
>>>codebase, mature
>>>ASF project (which can be a good thing and a bad thing ;) rather 
>>>'avalon-framework' (comparatively heavy compared to more recent 
>>>that defines the contracts between a component and a container. Has 3
>>>container projects to consider: avalon-phoenix, a mature microkernel 
>>>avalon-fortress, similar in weight and featureset to something like 
>>>nanocontainer with a
>>>1.0 release (successor to avalon-ecm, the container used in (among 
>>>other projects)
>>>apache-cocoon), and avalon-merlin, a more recent development which we 
>>>to be converging on as the successor to all other previously produced 
>>>containers. Merlin
>>>is further described in the email by Stephen Mcconnell. Arguably the 
>>>most extensive and most dynamic IoC container implementation around 
>>>(and hence
>>>also the most complex). Re: my earlier blurb on 'geronimo and avalon' 
>>>for more
>>> - haven't looked at 
>>>in much detail, but seems
>>>very similar in scope and setup to avalon at first glance. Framework 
>>>being refactored out
>>>of hibernate, one developer, still in alpha with no releases I think. 
>>>No offense to Howard
>>>intended, but I think he's cut himself a rather big piece of the 
>>>puzzle to recode from scratch
>>>at once. But I am an uninformed whiner, so I'm not going to comment 
>>>further in the hope that
>>>Howard will just eventually see the light and direct his energy 
>>>towards collaboration with the
>>>avalon peeps :P
>>>I am spending too much time on writing messages to this mailing list 
>>>(it is nearly 3am over here).
>>>I promise this is the last message from me for two weeks :D
>>>- Leo

View raw message