21.01.2013 Views

Lecture Notes in Computer Science 4917

Lecture Notes in Computer Science 4917

Lecture Notes in Computer Science 4917

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Average leakage-power sav<strong>in</strong>g (%)<br />

Variation-Aware Software Techniques for Cache Leakage Reduction 235<br />

60<br />

50<br />

40<br />

30<br />

20<br />

10<br />

Sav<strong>in</strong>g by cache-<strong>in</strong>itializ<strong>in</strong>g<br />

Sav<strong>in</strong>g by reschedul<strong>in</strong>g & register-renam<strong>in</strong>g<br />

0<br />

16KB 8KB 4KB<br />

16KB 8KB 4KB 16KB 8KB 4KB 16KB 8KB 4KB 16KB 8KB 4KB 16KB 8KB 4KB<br />

mpeg2 fft jpeg compress fir dct<br />

Fig. 8. Effect of cache-size on average leakage-sav<strong>in</strong>g results<br />

Set-associative caches take better advantage of the available cache-l<strong>in</strong>es and reduce<br />

the number of unused ones. Fig. 9 shows the leakage sav<strong>in</strong>gs <strong>in</strong> an 8KB cache when<br />

the number of ways changes from 1 (direct-map) to 8. The leakage-sav<strong>in</strong>g by cache<strong>in</strong>itialization<br />

reduces <strong>in</strong> caches with higher associativity, and f<strong>in</strong>ally total sav<strong>in</strong>g<br />

reduces to that obta<strong>in</strong>ed by the reschedul<strong>in</strong>g and register-renam<strong>in</strong>g technique as is<br />

aga<strong>in</strong> the case for MPEG2 and JPEG <strong>in</strong> Fig. 9.<br />

Furthermore, <strong>in</strong> set-associative caches, the location of each <strong>in</strong>struction <strong>in</strong> the cache<br />

cannot be precisely determ<strong>in</strong>ed s<strong>in</strong>ce there are multiple cache-l<strong>in</strong>es <strong>in</strong> the cache-set that<br />

corresponds to the address of the <strong>in</strong>struction. This uncerta<strong>in</strong>ty is expected to decrease<br />

the sav<strong>in</strong>g results of the reschedul<strong>in</strong>g algorithm, however, our cache simulator gives<br />

separate per-way residence-times for each <strong>in</strong>struction so as to direct the match<strong>in</strong>g<br />

process toward the cache-ways with higher probability of host<strong>in</strong>g the <strong>in</strong>struction.<br />

Average leakage-power sav<strong>in</strong>g (%)<br />

60<br />

50<br />

40<br />

30<br />

20<br />

10<br />

0<br />

Sav<strong>in</strong>g by cache-<strong>in</strong>itialization<br />

Sav<strong>in</strong>g by reschedul<strong>in</strong>g & register-renam<strong>in</strong>g<br />

1 2 4 8 1 2 4 8 1 2 4 8 1 2 4 8 1 2 4 8 1 2 4 8<br />

#ways #ways #ways #ways #ways #ways<br />

mpeg fft jpeg compres fir dct<br />

Fig. 9. Effect of set-associative caches on total leakage sav<strong>in</strong>g

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!