Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
En prestandajämförelse av minnesbaserade schemaläggningsmetoder
University West, Department of Engineering Science, Division of Computer and Electrical Engineering.
2014 (Swedish)Independent thesis Basic level (professional degree), 5 credits / 7,5 HE creditsStudent thesisAlternative title
A performance comparison of memory aware scheduling methods (English)
Abstract [sv]

När antalet kärnor per processor ökar ställer det högre krav på minnesbussbandbredden. Minnesbussen ansvarar för den trafik som skickas mellan processorn och minnesmoduler-na. För att undvika att program påverkas av den begränsade minnesbussbandbredden, används schemaläggning. En schemaläggare består av två delar, en algoritm och en karaktä-riseringsmetod.

I detta projekt testas följande fyra karaktäriseringsmetoder, Stack distance profile, Memory bandwidth usage, Miss rate och Slowdown based classification och dessa schemaläggs med tio olika program från NASA advanced super computing division [6]. Målet är att se hur bra eller dåligt karaktäriseringsmetoderna presterar och om prestandan skiljer sig åt på de olika processorerna.

Alla program körs först utan schemaläggning, enskilt samt parvis på processorerna. Efter körningarna skapas en baslinje som innehåller alla programkombinationer samt den pre-standaförsämring som programmen erhållit vid körningen. Denna baslinje kan sedan användas för att jämföra karaktäriseringsmetoderna med varandra genom att se vilken pre-standaförsämring som karaktäriseringsmetoderna fick i förhållande till baslinjens körningar. Hp3, T7500 och Kraken är namnen på de tre maskiner som innehåller processorerna och på dessa schemalades alla tio program med hjälp av de fyra karaktäriseringsmetoderna.

Slowdown based classification var den karaktäriseringsmetod som överlag presterade bäst och hade ett medelvärde som hamnade bland 30% av de bästa körningarna. Miss rates me-delvärde hamnade bland 60% av de bästa körningarna och Memory bandwidth usage bland 30% av de bästa körningarna. Stack distance profile presterade sämst och hamnade bland 70% av de bästa körningarna

Abstract [en]

When the number of cores per processor increases the demand on the memory buss bandwidth increase as well. The memory buss contains the traffic that flows between the processor and the memory modules. In an attempt to avoid that programs become affect-ed by the limited memory buss bandwidth, co-scheduling is used. A co-scheduler consists of two parts, an algorithm and a characterization method.

Four characterization methods are tested in this project, Stack distance profile, Memory bandwidth usage, Miss rate and Slowdown based classification and they are co-scheduled with ten programs from the NASA advanced super computing division [6]. The goal is to see how good or bad the characterization methods perform and if the performance differ on the different processors.

All programs are first executed without co-scheduling, solo and pairwise on the processors. After the executions are finished a baseline is created which contains all the combinations of programs and the slowdown value they received. The baseline can then be used to com-pare the characterization methods with each other, by seeing what slowdown the character-ization methods got relative to the baseline executions. The names of the three machines that contains the different processors are Hp3, T7500 and Kraken and on these, ten pro-grams are co-scheduled with four different characterization methods.

Slowdown based classification was the characterization method who performed best out of the four characterization methods. It got an average value among 30% of the best execu-tions. Miss rates got an average value among 60% of the best executions and Memory bandwidth usages average value was among 30% of the best executions. Stack distance profile performed the worst and got an average value among 70% of the best executions

Place, publisher, year, edition, pages
2014. , 17 p.
Keyword [sv]
Processorer, program, schemaläggning, minnesbussbandbredd
National Category
Electrical Engineering, Electronic Engineering, Information Engineering
Identifiers
URN: urn:nbn:se:hv:diva-6794Local ID: EXM301OAI: oai:DiVA.org:hv-6794DiVA: diva2:754066
Subject / course
Computer enigeering
Educational program
Nätverksteknik med IT-säkerhet
Supervisors
Examiners
Available from: 2014-10-10 Created: 2014-10-09 Last updated: 2014-10-10Bibliographically approved

Open Access in DiVA

No full text

By organisation
Division of Computer and Electrical Engineering
Electrical Engineering, Electronic Engineering, Information Engineering

Search outside of DiVA

GoogleGoogle Scholar

Total: 74 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf