THREAD CACHE ALLOCATION

摘要

Systems and techniques are described for thread cache allocation. A described technique includes monitoring input and output accesses for a plurality of threads executing on a computing device that includes a cache comprising a quantity of memory blocks, determining a respective reuse intensity for each of the threads, determining a respective read ratio for each of the threads, determining a respective quantity of memory blocks for each of the partitions by optimizing a combination of cache utilities, each cache utility being based on the respective reuse intensity, the respective read ratio, and a respective hit ratio for a particular partition, and resizing one or more of the partitions to be equal to the respective quantity of the memory blocks for the partition.

主权项

1. A computer-implemented method comprising:
monitoring input and output accesses for a plurality of threads executing on a computing device that includes a cache comprising a quantity of memory blocks, each of the threads being associated with a respective partition including a first respective quantity of the memory blocks, each memory block being included in only one of the respective partitions; determining a respective reuse intensity for each of the threads, the reuse intensity being based on, at least, a quantity of unique memory blocks in the respective partition of the thread that have been read during a first period of time; determining a respective read ratio for each of the threads, the read ratio being based on the input and output accesses for the thread; determining a second respective quantity of memory blocks for each of the partitions by optimizing a combination of cache utilities, each cache utility being based on the respective reuse intensity, the respective read ratio, and a respective hit ratio for a particular partition wherein the respective hit ratio is a ratio of a total number of cache hits given a variable quantity of memory blocks and a total number of input and output accesses during a second period of time for the thread associated with the particular partition; and resizing one or more of the partitions to be equal to the second respective quantity of the memory blocks for the partition.