struts-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chris Brock <bro...@gmail.com>
Subject Re: MVEL?
Date Mon, 13 Oct 2008 03:23:04 GMT

Well, that's one way of implementing a cache.  Even if you did do that, the
performance gain from compiling the expression is on the order of a
magnitude (or greater), so any contention in the hash lookup would be far
worth it.  However, there are other more efficient strategies that have been
employed through the use of external code generation, and tying compiled
expressions directly to the instance of a JSP tag.  

But we're talking about web-based stuff here.  MVEL is used for a broad
range of stuff, like actual scripting in Smooks and JBoss Drools.  It's used
for straight-up data binding in jBPM and JBoss ESB.  It's used for some
custom UI stuff in Mule Galaxy, etc.  

To you P.S.: MVEL works directly on the expression by using a sliding window
algorithm. Think of it like this.  

0 { f }
1 { o }
2 { o }
3 { . }
4 { b }
5 { a }
6 { r }
7 { [ }
8 { 0 }
9 { ] }
MVEL takes the string as an array.  It holds a start position, and a cursor
position.

When it starts parsing, it starts scanning until it finds the first
non-identifier character, which is '.'.  At this point it does a capture. At
this point: start=0; cursor=3;  This represents the boundary of the first
token.  We then process the token by converting it to a String, marking the
start position as cursor+1 and repeat. When we hit the next non-identifier
which is '[' MVEL knows this is a index accessor and acts appropriately,
following the same principle.

MVEL does this inline with actual evaluation.  It works out to be
significantly more efficient than StringTokenizer and also allows you to
incrementally add complexity over time.

Unfortunately, it becomes increasing more difficult to maintain a design
like this, especially as it evolves into more sophisticated constructs.

You're welcome to contact me directly in e-mail to discuss this, as I don't
want to pollute the Struts mailing list with non-Struts related talk.


Brian Pontarelli wrote:
> 
> Not sure I follow. If I compile an expression, how can I reuse the  
> compiled version? I would assume it would need to be in a cache where  
> the key is the expression String and the value is the compiled  
> version. Correct me if I'm wrong.
> 
> -bp
> 
> P.S. based on your knowledge, I'm wondering if JCatapult would perform  
> better if the expression wasn't divided into its pieces but rather  
> evaluated character by character until a boundary is hit, at which  
> time that portion is evaluated? I wouldn't think this would increase  
> performance drastically, but you would know much better than I.
> 
> 
> 
> On Oct 12, 2008, at 8:38 PM, Chris Brock wrote:
> 
>>
>> " I'd also be interested to hear a
>> good discussion about caching compiled MVEL expressions and whether or
>> not thread contention for the cache is an issue at all"
>>
>> There is no contention in the cache.  MVEL returns self-contained,
>> stateless, evaluation trees (or bytecode via the JIT) that do not  
>> require
>> synchronization or contention in multi-threaded scenarios.  The  
>> payload
>> returned by the compiler is essentially stateless code, and there is  
>> no
>> "cache" that is used such as reflection cache or otherwise as there  
>> is in
>> things like Commons EL, or JEXL.  This is actually, from an  
>> architectural
>> perspective what makes MVEL stand apart from these technologies.
>>
>>
>> Brian Pontarelli wrote:
>>>
>>>> Sure.  But OGNL will return similar results with 50 tests.  Yet
>>>> people have
>>>> run into performance problems.  The issue is that you're not looking
>>>> at
>>>> performance in terms of resource contention, and in terms of  
>>>> aggregate
>>>> resource usage.
>>>
>>> I'd say that for web application expressions OGNL and MVEL are about
>>> equal then. In fact, I've never wanted to replace OGNL for  
>>> performance
>>> reasons. It was for primarily other reasons.
>>>
>>>>
>>>> Say you have a page which contains 20 expressions.  And your pages  
>>>> are
>>>> getting hit 15 times a second (a reality in some high traffic  
>>>> sites).
>>>> That's 300 expressions running every second.  Now, in insolation
>>>> that's
>>>> probably chump change.  But as resource contention rises in these
>>>> situation,
>>>> the overall efficiency drops and resource usage is exaggerated as a
>>>> result.
>>>
>>> I've worked with this level of traffic and higher and it is still not
>>> an issue to be setting 20 values for 1ms per request.
>>>
>>>
>>>>
>>>> You might in term start to find that what is only 0ms in an isolate
>>>> closed-loop test (which is not a very good way to benchmark in Java,
>>>> by the
>>>> way) could very well be something that contributes to a significant
>>>> amount
>>>> of CPU time in systems with high load.
>>>
>>> Probably not in this case though and the scale between 1 iteration  
>>> and
>>> 50 is decent testament to that. It the CPU was pinned it would be  
>>> more
>>> linear.
>>>
>>>
>>>>
>>>> Take these real benchmarks (from MVEL 1.2--which is old):
>>>>
>>>> Test Name            : Deep Property
>>>> Expression           : foo.bar.name
>>>> Iterations           : 50000
>>>> Interpreted Results  :
>>>> (OGNL)               : 1955.20ms avg.  (mem delta: -790kb)
>>>> [1936,1949,1943,1994,1954]
>>>> (MVEL)               : 114.80ms avg.  (mem delta: -112kb)
>>>> [119,113,110,117,115]
>>>> Compiled Results     :
>>>> (OGNL Compiled)      : 92.80ms avg.  (mem delta: -580kb)
>>>> [92,92,92,92,96]
>>>> (MVEL Compiled)      : 1.80ms avg.  (mem delta: -18kb) [1,2,2,2,2]
>>>
>>> Here's what I got for 50K on my box using MVEL and JCatapult side by
>>> side:
>>>
>>> MVEL 808ms
>>> JCatapult 1200ms
>>>
>>> MVEL had a hit for the first method call, but it was only 40ms.
>>> Otherwise, they performed exactly the same for anything up to 50
>>> iterations. MVEL often poked above 1ms for single iterations, while
>>> JCatapult never did, but that's negligible for both. JCatapult is
>>> definitely slower as the iterations go up.
>>>
>>> I tossed in a thread test with 50 threads each running 50K iterations
>>> and the averages were:
>>>
>>> MVEL 8000ms
>>> JCatapult 23000ms
>>>
>>> However, under one test condition, MVEL never returned and caused a
>>> load of 50 on my box. It was quite distressing, but it looked like
>>> MVEL got into a bunch of infinite loops or something. I let it run at
>>> a load of 50 for a while and then I had to kill it, but none of the
>>> threads had finished yet.
>>>
>>> I also did a 50 thread and 50 iteration test and the averages were
>>> roughly:
>>>
>>> MVEL 30ms
>>> JCatapult 120ms
>>>
>>> Except for the case above, MVEL definitely out-performs JCatapult.
>>>
>>>
>>>>
>>>> ... 50,000 iterations on MVEL interpreted in 114.80ms.  This is a
>>>> 1000x more
>>>> iterations than your benchmark.  If I divide 114.8ms / 1000 ... I
>>>> get 0.1ms
>>>> (or what would otherwise be rounded down to 0ms). In OGNL's case, it
>>>> did 50
>>>> iterations in 1.95ms (or what would be measured as 1ms -- as these
>>>> time
>>>> measurements always round down because of the fact  
>>>> currentTimeMillis()
>>>> returns the result in MS).
>>>
>>> Although JCatapult is slower, I'd be careful with such math because  
>>> it
>>> isn't always as linear as this.
>>>
>>>
>>>
>>>>
>>>>
>>>> You can talk about "good enough" all you want, but faster is always
>>>> better
>>>> when it comes to scale. :)
>>>
>>> I know a lot about scale and this is not the only truth. In fact, for
>>> what we are talking about, good enough should be just fine. Most  
>>> scale
>>> problems occur because of bottlenecks and I doubt that our case of  
>>> web
>>> applications and setting parameters is a bottleneck.
>>>
>>> However, I'm definitely welcome to suggestions on improvements for my
>>> quite simple expression evaluator. I'd also be interested to hear a
>>> good discussion about caching compiled MVEL expressions and whether  
>>> or
>>> not thread contention for the cache is an issue at all.  
>>> Unfortunately,
>>> because JCatapult uses my concept of dynamic attributes quite  
>>> heavily,
>>> it might be difficult to swap in MVEL without some tweaks to the type
>>> conversion API. But I could look into it.
>>>
>>>
>>> -bp
>>>
>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@struts.apache.org
>>> For additional commands, e-mail: dev-help@struts.apache.org
>>>
>>>
>>>
>> -- 
>> View this message in context:
>> http://www.nabble.com/MVEL--tp19867360p19948098.html
>> Sent from the Struts - Dev mailing list archive at Nabble.com.
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: dev-unsubscribe@struts.apache.org
>> For additional commands, e-mail: dev-help@struts.apache.org
>>
> 
> 
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@struts.apache.org
> For additional commands, e-mail: dev-help@struts.apache.org
> 
> 
> 
-- 
View this message in context: http://www.nabble.com/MVEL--tp19867360p19948361.html
Sent from the Struts - Dev mailing list archive at Nabble.com.


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


Mime
View raw message