2008/2/11, Simon Laws <[EMAIL PROTECTED]>: > On Feb 10, 2008 12:50 PM, Giorgio Zoppi <[EMAIL PROTECTED]> wrote: > > > I've tried my app a while ago in a 16 node-cluster and I'm publishing > > my result at > > http://www.cli.di.unipi.it/~zoppi/out/ch06.html<http://www.cli.di.unipi.it/%7Ezoppi/out/ch06.html> > > They are in italian but you can find many images. I suppose to cache > > component uri > > as you can see in the code: > > http://www.cli.di.unipi.it/~zoppi/out/apas06.html#d4e1968<http://www.cli.di.unipi.it/%7Ezoppi/out/apas06.html#d4e1968> > > in order to speed up applications. > > At the end of this work there's all code that i crafted in order to > > have a behavioural skeleton > > task farm: a SCA distributed task executor. > > Next I have to port it to the new Tuscany 1.1. > > Cheers, > > Giorgio > > > > --------------------------------------------------------------------- > > To unsubscribe, e-mail: [EMAIL PROTECTED] > > For additional commands, e-mail: [EMAIL PROTECTED] > > > > > Hi Giorgio, thanks for posting the results. Looking at the graphs you can > see the measured completion time diverging from the ideal completion time as > the task size shrinks and the number of nodes increases. This is to be > expected as the infrastructure becomes more noticeable in these situations.
Yes it's expected because your infrastructure use a ConcurrentHashMap for async invocations as a Work queue. > The question is is it too intrusive and do you have a feeling for where we > get the best payback in reducing the overhead. You mention caching the URL. Yes. As you can see the grain is middle/coarse and if you dont' cache the URL, the situation get worse. I cache the url and I cache the callableReference, which i resolve for calling back. CallableReference --> proxy costs around 500ms too much for me. > In other discussions over the last few weeks [1] [2] we are looking at a > slightly simplified approach to deploying the nodes where the topology is > calculated ahead of time and hence endpoint information can be provided with > deployed composites. Rather than the nodes having to make calls across the > information to find this information. This is initially less dynamic than > the situation we have now but you could argue that it is more predictable. > Do you see problems for you application in taking this slightly more static > approach? No idea. I'll start to review all next week. > I still have the previous code you attached to the open JIRA [3] sitting on > my machine. My intention was to check it in as a demonstration but haven't > done so yet. Should I wait until you have ported to 1.1? Yes. Wait, because i'm going to refactor inorder to have something more loosely decoupled. I saw the latest callback issue, so maybe i'll able to use infrastructure callback better. Cheers, Giorgio. --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
