Browse Prior Art Database

Method for guaranteed resources for incoming transactions from a scalable node/switch by aborting its unused speculative memory read transactions

IP.com Disclosure Number: IPCOM000010053D
Publication Date: 2002-Oct-16
Document File: 3 page(s) / 72K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a method for guaranteed resources for incoming transactions from a scalable node/switch by aborting its unused speculative memory read transactions. Benefits include improved performance.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 81% of the total text.

Method for guaranteed resources for incoming transactions from a scalable node/switch by aborting its unused speculative memory read transactions

Disclosed is a method for guaranteed resources for incoming transactions from a scalable node/switch by aborting its unused speculative memory read transactions. Benefits include improved performance.

Background

               The speculative read is a performance enhancement that can be discarded without affecting computational accuracy.

 

A confirmation or cancel may be sent later to terminate the speculative read, but the protocol does not guarantee that every speculative read is terminated over time. Unterminated speculative reads can accumulate, consuming resources. No function exists to determine at the memory controller if a particular speculative read will be confirmed or cancelled. Performance suffers if confirmed speculative reads are discarded in favor of speculative reads that will never be cancelled.

              The resources consumed by speculative reads are shared with other inbound transactions. Each resource is assigned an index.

Description

      The disclosed method is a simple mechanism that discards older unterminated speculative reads and releases their resources.

              In a distributed shared memory system (see Figure 1), multiple memory node controllers (each controlling their own memory) are connected directly (see Figure 2) or through a switch (see Figure 3). To reduce the latency of accessing remote me...