EFFECTIVE AND SCALISABLE BUILDING AND SPECIFICATIONS OF HASH TABLES USING MORE GPUs

摘要

The approaches described allow the effective and scalable use of several GPUs to build and probe hash tables and to materialize the results of surveys. Direct memory access through the GPUs for building and/or probing a hash table can be distributed over the GPUs and used(i) the development of global storage location identifiers. A global location identifier can be calculated from the data of an entry and identify a global location for an insertion and/or exploration using the entry. The global location identifier can be used by a GPUto determine whether an insertion or exploration is to be carried out using an entry and/or where the insertion or exploration is to be carried out. To coordinate the GPUs when materializing the results of a hash table, a global offset to the global output buffer can be held in a memory to which each of the GPUs can access, or the GPUs can calculate the global offsets using an exclusive sum of the local output buffer sizes.