Can't explain this performance drop

Question

Hi everybody,
i am currently working on a data-mining application which uses a hudge amount of system resources, so tha'ts why i want to parallize it.
I've got a dual xeon server with 8 cores, which i want to use.
The good news: My application can use all 8 cores to execute his task in a very fast manner.
The bad news: If a use another dataSet for my analysis it only uses once core. So that's why I write here to hopefully get some help of more experiences coders.
Situation:
1) How do i parallize:
I got a Vector including sequences which i want to analyze. Therefore I iteratare through this vector and for each sequence i assign a new task in my threadpool (like: [http://www.informit.com/articles/article.aspx?p=30483&seqNum=6] ). This threadpool was initially created with 8 threads (4 or 6 have the same result). The results of each task are stored in a collection (synchronized ArrayList).
2) What kind of work has to be done.
Actually a sequence contains elements which contains items. A worker now checks whether two sequences contain the same items.
At the beginning i give the worker the following arguments:
(SequenceToCompare, ListOfSequencesToCompareWith*, SynchronizedListToPutNewSequenceIn)
*same like the list i iterate through while creating these tasks.
So the worker takes the SequenceToCompare (s1) and iterates through the ListOfSequencesToCompareWith(sList) and checks if they contain the same items.
If they do, it takes s1 and adds the last item of the current sequence of sList. After it puts s1 into the SynchronizedListToPutNewSequenceIn which is filled by all the other threads, too.
3) Different DataSets
The "fast running" dataset is much shorter as the slow one. it has less sequences to compare at the beginnig (1000 against 2000+ (up to 500.000)). Structure is the same. All sequences in the fast dataSet are build out of 15 items, while the slow one contains 75 different items. For each check of the two sequences it iterates through all items in the dataset takes the hashCode compares. Then in the case they are same and I can add an item to s1, i take the hashCode and get the item of a CACHE Objekt, which contains the item in a hashMap. (Can this be my bottleneck?)
4) Problem? Solving?
I hope that my explanation was good enough, maybe we can through this in an iterative way, so please tell me which additional information you need and i will post it (+code).
Your support is highly appreciated.
Regards
Andy

Answer 1

1) How do I parallelize:
Use the Java concurrency libraries
e.g.
Executor executor = Executors.newFixedThreadPool(8);
for(final Object toProcess: myVector)
   executor.submit(new Runnable() {
       public void run() {
           // process toProcess;
    }