DefaultSorter is the same sorter implementation used in MapReduce world and is single threaded. PipelinedSorter on the other hand works based on divide/conquer approach and works on multiple sort-spans which can be sorted by different threads. More details can be found in http://people.apache.org/~gopalv/PipelinedSorter.pdf.
It is not possible to increase sort.mb to greater than 2 GB with defaultsorter implementation. With pipelinedsorter, it is possible to allocate more than 2 GB as sort buffer. This could be useful in scenarios where you have large containers and can allocate more than 2 GB for sort buffer to avoid potential disk spills. It is possible to control the number of threads allocated for sorting in PipelinedSorter using "tez.runtime.pipelined.sorter.sort.threads" (defaults to 2). Setting this to lot higher value might not be useful as it depends on the number of processors available in the system and the number of containers running on the system. Depending on workloads, 2-4 could be a sweetspot. Starting Tez 0.7, PipelinedSorter has been made the defacto-sorter, though users can switch back to DefaultSorter (mapreduce world implementation) by setting "tez.runtime.sorter.class=LEGACY" ~Rajesh.B On Wed, Jun 3, 2015 at 7:18 AM, [email protected] <[email protected]> wrote: > In OrderedPartitionedKVOutput ,I see > if (this.conf.getInt(TezRuntimeConfiguration.TEZ_RUNTIME_SORT_THREADS, > TezRuntimeConfiguration.TEZ_RUNTIME_SORT_THREADS_DEFAULT) > 1) { > sorter = new PipelinedSorter(getContext(), conf, > getNumPhysicalOutputs(), > memoryUpdateCallbackHandler.getMemoryAssigned()); > } else { > sorter = new DefaultSorter(getContext(), conf, > getNumPhysicalOutputs(), > memoryUpdateCallbackHandler.getMemoryAssigned()); > } > > When set tez.runtime.sort.threads >1 will choose PipelinedSorter . > ------------------------------ > [email protected] > -- ~Rajesh.B
