Browse Prior Art Database

Combining of Consecutive Processor Stores

IP.com Disclosure Number: IPCOM000061616D
Original Publication Date: 1986-Aug-01
Included in the Prior Art Database: 2005-Mar-09
Document File: 1 page(s) / 12K

Publishing Venue

IBM

Related People

Brandt, HR: AUTHOR [+3]

Abstract

Back-to-back store requests from a CPU which constitute the even and odd doublewords of a quadword may be serviced as a single write operation in the next lower level in the storage hierarchy. Store-through CPU designs commonly send each processor store data unit from the cache to the next level of the memory hierarchy. The bandwidth required of the memory is often a function of the anticipated store-type activity and becomes a problem when CPU cycle times improve at a faster rate than memory cycle times. It is therefore desirable to reduce the store traffic to the memory in order to conserve memory bandwidth for line fetches and channel requests. This problem can be reduced by combining the data units of two successive processor store requests to the next level in the memory hierarchy into a single store operation.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 51% of the total text.

Page 1 of 1

Combining of Consecutive Processor Stores

Back-to-back store requests from a CPU which constitute the even and odd doublewords of a quadword may be serviced as a single write operation in the next lower level in the storage hierarchy. Store-through CPU designs commonly send each processor store data unit from the cache to the next level of the memory hierarchy. The bandwidth required of the memory is often a function of the anticipated store-type activity and becomes a problem when CPU cycle times improve at a faster rate than memory cycle times. It is therefore desirable to reduce the store traffic to the memory in order to conserve memory bandwidth for line fetches and channel requests. This problem can be reduced by combining the data units of two successive processor store requests to the next level in the memory hierarchy into a single store operation. For example, two consecutive processor store requests may be for even and odd doublewords (DWs) which can be combined into a quadword that is stored in a single operation in main storage or in an L2 cache, taking no longer than a store operation for either DW store would have required alone. This function may use a special tag on each CP store request as well as a facility within the next lower level store array to accumulate two doublewords for each quadword write. This "dual store operation for the price of one" requires a structure of the next lower level store array that can store a quadword as a unit operation. The CPU execution element places storage operand results in a store buffer as part of the process of instruct...