Alibaba Group Holding Limited
METHODS AND SYSTEMS FOR FETCHING DATA FOR AN ACCELERATOR

Abstract:

The present disclosure provides methods, systems, and non-transitory computer readable media for fetching data for an accelerator. The methods include detecting an attempt to access a first page of data that is not stored on a primary storage unit of the accelerator, wherein the first page of data corresponds to part of an array with a number of dimensions; and responsive to detecting the attempt to access the first page of data: dividing the array into sub-arrays by: dividing the array into a number of 1st level sub-arrays, and dividing a first 1st level sub-array into a number of 2nd level sub-arrays, wherein the first 1st level sub-array contains the first page of data; selecting pages for prefetching, wherein selecting pages for prefetching includes selecting all pages in a first 2nd level sub-array for prefetching if the first 2nd level sub-array meets a page access volume condition, wherein the first 2nd level sub-array contains the first page of data; and transferring the first page of data and any pages of data selected for prefetching from a memory system connected to the accelerator into the primary storage unit.