Culler d patterson ucb s ziavras how summarize suite

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: han Bandwidth • • • Secure transfers require involvement of the O/S in the form of software interrupts/exceptions/traps overhead per transfer Long messages amortize overhead; overhead bigger part of short messages DMA-like transfers: same overhead shared by many individual data items Transfers may be delayed until enough data has been accumulated for a bulk transfer Some material is adapted from D. Culler & D. Patterson (UCB) S. Ziavras How Summarize Suite Performance (1/4) • Arithmetic average of execution time of all pgms1 – But they vary by 4X in speed, so some would be more important than others in arithmetic average • Could add a weight per program, but how pick weight? 2 – Different companies want different weights for their products • Using the arithmetic mean of performances: WRONG • SPECRatio: Normalize execution times to reference computer, yielding a ratio proportional to performance: time on reference computer SPECRatio = time on computer being rated Some material is adapted from D. Culler & D. Patterson (UCB) S. Ziavras Mean Computer Performance • Similar problems with n computers and 1 program, or 1 computer and n exclusive execution modes (such as vector, sequential, parallel), or 1 computer and a suite of n programs • Assume n programs & 1 computer programs computer • 1 Arithmetic mean time of a given computer for a given set of n programs (ti for program i) T= 1/n Σi=1n ti • 2 Weighted arithm. mean time (fi: weight of program i) T*= Σi=1n fi*ti Some material is adapted from D. Culler & D. Patterson (UCB) S. Ziavras Arithmetic Mean Performance Arithmetic Mean Performance • Ri: execution rate for program i, for 1 ≤ i ≤ n Ri ~ 1/Ti = a R 1 n i • If π ={f i, for i=1,2,...,n} are the program weights for i=1 are the program weights – Weighted arithmetic mean execution rate n R a = ∑ (f i • R i ) * W WRONG N R n ∑ i= i =1 It’s proportional to the sum of the inverses of execution times times 1/R*...
View Full Document

{[ snackBarMessage ]}

Ask a homework question - tutors are online