Inter-host load splitting divides the computation load across hosts. You
can use methods similar to intra-processing element (PE) splitting with the
Split operator except that there is no need for threaded ports on the
downstream operators. You can alternatively use the parallel annotation and user-defined
parallelism.
When operators in a parallel region are fused only with operators within that parallel region,
Teracloud®
Streams replicates the PEs and achieves inter-host load splitting. For more information, see User-defined
parallelism.
This following graphic shows inter-host load splitting across PEs by using the
Split operator:
For small tuple sizes, network processing on the output ports can become a
bottleneck.
You can ensure that PEs are placed on separate hosts by specifying exlocation
constraints. For example,
