> I looked at the two examples and conclude that the pattern for running a BSP > job is creating a subclass of HamaCluster that initializes various stuff and > then starts a number of BSP threads that do the actual work. Is that about > right? It would seem to imply that a BSP job can only use one compute node, > because each process is just a Java thread. On the other hand, it hardly > makes sense to use Hadoop for just one node, so I wonder what I am missing > here.
Nope, it can be run on multi-node cluster. The BSP job handles a list of peer servers, each peer server is run in each node. and threads will be handled by peer server. In other words, a BSP job is a multi-core multi-thread program on distributed system. On Mon, Feb 8, 2010 at 9:02 PM, Konrad Hinsen <[email protected]> wrote: > On 08.02.2010, at 12:04, Edward J. Yoon wrote: > >> I would recommend you to read below test codes. >> >> svn_trunk/src/test/org/apache/hama/bsp/BSPPeerTest and SerializePrinting. > > Thanks, that's a starting point! > > I looked at the two examples and conclude that the pattern for running a BSP > job is creating a subclass of HamaCluster that initializes various stuff and > then starts a number of BSP threads that do the actual work. Is that about > right? It would seem to imply that a BSP job can only use one compute node, > because each process is just a Java thread. On the other hand, it hardly > makes sense to use Hadoop for just one node, so I wonder what I am missing > here. > > Konrad. > > -- Best Regards, Edward J. Yoon @ NHN, corp. [email protected] http://blog.udanax.org
