[jbosscache-dev] VAM in JDBCCacheLoader : performance numbers and analysis

[Manik Surtani]

On 28 Dec 2006, at 00:05, Galder Zamarreno wrote:

> Hi,
>
> I've been doing some performance tests for JDBCCacheLoader with  
> both normal Java Serialization and VersionAwareMarshaller, using  
> the existing JDBCCacheLoaderPerfTest.
>
> The results I got back were quite suprising. Executing a test  
> consisting of 10 threads (each thread interacts with a single node)  
> and each thread adds 100 key/value pairs to their node. Tests  
> showed that normal serialization was something like 15 times faster  
> than VAM when it came to writes. For reads, normal serialization  
> was still 4 times faster. We're talking about String/String key  
> value pairs here.
>
> Looking a bit deeper with the profiler, I have seen that  
> TreeCacheMarshaller200 takes ages marshalling a map containing  
> string key/value pairs. Is this expected? We're adding quite a few  
> things to a simple String/String map, such as versioning and  
> reference information. Compared to standard writeObject in HashMap,  
> TreeCacheMarshaller200 spends quite a bit of time mantaining/ 
> updating a reference map.

It is possible that such a reference map is meaningless in the case  
of writing to a cache loader.  It makes sense streaming on the wire  
for an RPC call though.  So maybe there needs to be an overloaded  
form of objectTo|FromByteBuffer that could disable the ref map?

>
> In this case, for a single put, each key and value values are  
> different (key0/value0 key1/value1 key2/value2...), maintaining  
> references becomes an extra overhead as the reference maps are  
> created for each request to marshall the original map. However, I  
> can see the point why it was done (avoid extra serialization and  
> reduce the amount of data in the wire).
>
> I had a quick thought about caching reference data so that it  
> survived several requests to marshal data, but it's probably a bad  
> idea as this cache would become a contention for concurrent  
> marshalling and increase the memory requirements. Has anyone been  
> this route before? I think it might only produce positive results  
> in such specific test as this.
>
> I tried JBoss serialization and the results were even worse. What  
> type of Objects is JB serialization recommended for? Is it user  
> defined ones?
>

User defined ones, yes.

> I have attached JProfiler files from all three tests and also a CVS  
> diff in case anyone wants to have a quick glance. Note that even  
> though there's 1000 calls to marshallMap, as threads add new key/ 
> value pairs, the data read-to-be-updated increases and hence the  
> number of items to iterate through in the map.

In any case, as Bela mentioned, a unit test that benchmarks the VAM  
on it's own (against Java serialization) makes sense.

>
> FYI, useRegionBasedMarshalling is set to false.
>
> Galder Zamarreño
> Sr. Software Maintenance Engineer
> JBoss, a division of Red Hat
>
> IT executives: Red Hat still #1 for value http://www.redhat.com/ 
> promo/vendor/
> <JBCACHE-897-serialization.java-10-100.jps>
> <JBCACHE-897-serialization.jboss-10-100.jps>
> <Head-serialization.java-10-100.jps>
> <diff.27.12.2006.txt>
> _______________________________________________
> jbosscache-dev mailing list
> jbosscache-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/jbosscache-dev