Class OverlapAccumulatorKernel

Inheritance Relationships

Base Type

Class Documentation

class ral::batch::OverlapAccumulatorKernel : public ral::cache::distributing_kernel

The OverlapAccumulatorKernel assumes three input caches:

  • ”batches”

  • ”preceding_overlaps”

  • ”following_overlaps” It assumes the previous kernel will fill “batches” N cacheData that is sorted and the batches are in order It assumes that preceding_overlaps and following_overlaps will contain N-1 cacheData that corresponds to the preceding and following overlaps copied from the batches The idea is that part of batches[x] will be copied to fill preceding_overlaps[x+1] and part will be copied to fill following_overlaps[x-1] The preceding_overlaps and following_overlaps will contain metadata to indicate if the overlaps are complete or incomplete (DONE_OVERLAP_STATUS or INCOMPLETE_OVERLAP_STATUS) For example, preceding_overlaps[x+1] would be set to INCOMPLETE_OVERLAP_STATUS if batches[x] was not big enough to fulfill the required preceding_value which is defined by the window frame (i.e. ROWS BETWEEN X PRECEDING AND Y FOLLOWING)

The purpose of OverlapAccumulatorKernel is to ensure that all INCOMPLETE_OVERLAP_STATUS overlaps coming from the previous kernel are COMPLETED by copying from other batches. The other purpose is to fill the overlaps of preceding_overlaps[0] and following_overlaps[N] with data that has to come from the neighboring nodes (or make a blank overlap if there is no neighbor) The OverlapAccumulatorKernel will comunicate with other nodes by sending overlap_requests (PRECEDING_REQUEST, FOLLOWING_REQUEST) which when received and responded to as PRECEDING_RESPONSE and FOLLOWING_RESPONSE

Right before outputting, OverlapAccumulatorKernel will combine preceding_overlaps[x], batches[x] and following_overlaps[x] together to make one batch pushed to the output. The following kernel, will then have in one batch with number of rows (preceding_overlaps[x]->num_rows() + batches[x]->num_rows() + following_overlaps[x]->num_rows()), which is the data necessary to procude a batches[x]->num_rows() worth out final output rows.

This kernel uses ConcatenatingCacheDatas a lot to try to reduce and postpone the materialization of data.

Public Functions

OverlapAccumulatorKernel(std::size_t kernel_id, const std::string &queryString, std::shared_ptr<Context> context, std::shared_ptr<ral::cache::graph> query_graph)
inline virtual std::string kernel_name()
ral::execution::task_result do_process(std::vector<std::unique_ptr<ral::frame::BlazingTable>> inputs, std::shared_ptr<ral::cache::CacheMachine> output, cudaStream_t stream, const std::map<std::string, std::string> &args) override
virtual kstatus run() override

Executes the batch processing. Loads the data from their input port, and after processing it, the results are stored in their output port.


kstatus ‘stop’ to halt processing, or ‘proceed’ to continue processing.