Browse Prior Art Database

A Method and System for Computing a Single Thread Performance in a Simultaneous Multithreading Environment

IP.com Disclosure Number: IPCOM000237737D
Publication Date: 2014-Jul-08
Document File: 2 page(s) / 32K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a system and method for computing a single thread performance in a simultaneous multithreading environment.

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

Page 01 of 2

A Method and System for Computing a Single Thread Performance in a Simultaneous Multithreading Environment

Disclosed is a method and system for computing a single thread performance in a simultaneous multithreading environment.

The method and system estimates performance of each independent thread in two environments. In one environment, the thread is executed by itself whereas in another environment the thread is executed by a Simultaneous Multi-Threading (SMT). In order to estimate the performance, the system measures behavior of the thread within a pipeline and estimates an amount of variance in behavior of the thread in the two environments. Here, the amount of variance is estimated based on current behavior of the thread executed in the SMT environment and by itself.

Further, the behavior of thread executed by itself on a core is analyzed based on time consumed by the thread. Subsequently, features of the thread are mapped to a model to estimate the performance of the thread executed by itself. Here, the thread is mapped to the model by analyzing a plurality of specific operations in the pipeline. Further, the thread is mapped to the model to determine linear equations for representing the performance of the thread executed by itself.

A model of linear equation is determined by using a formula

Pred_Cy = c1*Atr1 + … + cn*Atrn.

The formula simplifies the process of computing reasonable number of terms and is amenable to implement Single Instruction Multiple Data (SIMD). However, for an optimized reduction of overall total error and maximum error for any workload, a single linear formula is determined in a plurality of estimations. Here, the single linear formula represents a subtraction of the number of SMT from the number of SMT cycles. Thus, an average error and the maximum error are determin...