A Better Crystal Ball

Pages: 1 2 3 4 5 6 7 8

Testing the Model

To test the model, I used the 26 individual benchmark programs that comprise the SPECbase2k scores for the Pentium 4 (P4) processor. The P4 scores were chosen because there are results available for four different clock frequencies submitted over a relatively short time span with no ostensible changes to the system and software configurations [2]. This kind of ‘pure’ frequency scaling data is very useful to test my model, but unfortunately occurs only rarely, even for PC oriented processors that experience frequent clock rate boosts.

The input data points to my performance scaling model are the benchmark scores at 1.4 and 1.5 GHz. I used the model to predict P4 performance at lower (1.3 GHz) and higher (1.7 GHz) clock frequencies and compared the predictions to the actual SPEC2k scores that Intel submitted for those clock frequencies. But first let’s examine the hardware and software configurations:

<b>Table 2 Hardware and Software Configuration for Test Data Measurements</b>
&nbsp;

P4 / 1.3 GHz Submission

P4 / 1.4 GHz Submission

P4 / 1.5 GHz Submission

P4 / 1.7 GHz Submission

Motherboard

D850GB

D850GB

D850GB

D850GB

Chipset

i850

i850

i850

i850

Bus Frequency

400 MHz

400 MHz

400 MHz

400 MHz

Memory

PC800, non-ECC

PC800, non-ECC

PC800, non-ECC

PC800, non-ECC

Operating

System

Win 2000

(build 2195)

Win 2000

(build 2195)

Win 2000

(build 2195)

Wind 2000

(build 2195)

Toolset

IRC 5.0

VC++ 6.0 (lib)

SmartHeap 5.0

IRC 5.0

VC++ 6.0 (lib)

SmartHeap 5.0

IRC 5.0

VC++ 6.0 (lib)

SmartHeap 5.0

IRC 5.0

VC++ 6.0 (lib)

SmartHeap 5.0

Base C options

-QxW -Qipo shlW32M.lib +FDO

-QxW -Qipo shlW32M.lib +FDO

-QxW -Qipo shlW32M.lib +FDO

-QxW -Qipo shlW32M.lib +FDO

Base C++ options

-QxW – Qipo -GX -GR

-QxW – Qipo -GX -GR

-QxW – Qipo -GX -GR

-QxW – Qipo -GX -GR

Base FORTRAN

options

-Qipo -QxW -O3 +FDO

-Qipo -QxW -O3 +FDO

-Qipo -QxW -O3 +FDO

-Qipo -QxW -O3 +FDO

It appears that the hardware and software factors for the four SPECbase2k submissions are identical, except for processor clock frequency. The application of the 1.4 GHz and 1.5 GHz performance data points for my performance scaling model is shown below in Table 3.

<b>Table 3 Fitting the Model to the 1.4 GHz and 1.5 GHz Data Points</b>

Program

P1

P2

m

Program

P1

P2

m

164.gzip

515

553

0.000*

168.wupwise

719

759

0.209

175.vpr

290

298

0.597

171.swim

1243

1244

0.988

176.gcc

573

588

0.617

172.mgrid

530

558

0.247

181.mcf

473

473

1.000

173.applu

630

641

0.743

186.crafty

465

497

0.034

177.mesa

520

553

0.105

197.parser

451

472

0.333

178.galgel

522

537

0.581

252.eon

608

650

0.031

179.art

513

514

0.971

253.perlbmk

660

703

0.083

183.equake

720

739

0.614

254.gap

673

708

0.258

187.facerec

427

451

0.202

255.vortex

699

735

0.265

188.ammp

356

366

0.590

256.bzip2

403

420

0.393

189.lucas

740

764

0.529

300.twolf

394

403

0.665

191.fma3d

406

427

0.262

SPECint_base2k

502

524

0.351

200.sixtrack

240

257

0.008

F1 =

F2 =

301.apsi

411

427

0.438

* see note in text

1.4 GHz

1.5 GHz

SPECfp_base2k

529

549

0.460

What is surprising is how close to the minimum and maximum values of 0 and 1 the m parameter was observed. In fact beyond – the value of m for 164.gzip was actually calculated to -0.03, an apparent effect of individual SPEC2k scores being rounded to integral amounts. A negative amount is impossible in practice as it implies a supra-linear performance scaling with respect to processor clock frequency so m was set to exactly zero for 164.gzip.


Pages: « Prev   1 2 3 4 5 6 7 8   Next »

Be the first to discuss this article!