Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
ÀÌÁ¾ ÇÁ·Î¼¼¼ ȯ°æ¿¡¼ÀÇ º¹¼öÀÇ µö ·¯´× ¾îÇø®ÄÉÀÌ¼Ç ½ºÄÉÁÙ¸µ ±â¹ý |
¿µ¹®Á¦¸ñ(English Title) |
A Scheduling Technique for Multiple Deep Learning Applications on Heterogeneous Processors) |
ÀúÀÚ(Author) |
À̼®¿ø
¹Ú°æÈÆ
ÀÌÇü¿ø
Seokwon Lee
Kyeonghun Park
Hyungwon Lee
¿ÀÁø¿ì
Çϼøȸ
Jinwoo Oh
Soonhoi Ha
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 26 NO. 07 PP. 0303 ~ 0311 (2020. 07) |
Çѱ۳»¿ë (Korean Abstract) |
ÀÌÁ¾ ÇÁ·Î¼¼¼ ȯ°æ¿¡¼ÀÇ º¹¼öÀÇ µö ·¯´× ¾îÇø®ÄÉÀÌ¼Ç ½ºÄÉÁÙ¸µÀº ±âº»ÀûÀ¸·Î NP-³ÇØ(NP-Hard) ¹®Á¦¿¡ ¼ÓÇÏ¿© ¸Å¿ì Å« ¹®Á¦ °ø°£À» °¡Áø´Ù. ±×·¡¼ ÀϹÝÀûÀ¸·Î À¯Àü ¾Ë°í¸®Áò(GA, Genetic Algorithm)°ú °°Àº ¸ÞŸ ÈÞ¸®½ºÆ½ÀÌ Àû¿ëµÉ ¼ö ÀÖÁö¸¸ ÀÌ´Â ¼öÇà ½Ã°£ÀÌ ±æ¾î¼ ·±Å¸ÀÓ¿¡ Àû¿ëÇϱ⠾î·Æ´Ù´Â ´ÜÁ¡À» Áö´Ñ´Ù. µû¶ó¼ º» ¿¬±¸¿¡¼´Â ÀÌ·¯ÇÑ ´ÜÁ¡À» º¸¿ÏÇÏ¸é¼ ¼º´É ¶ÇÇÑ Å©°Ô ¶³¾îÁöÁö ¾Ê´Â »õ·Î¿î ±â¹ýÀÇ ½ºÄÉÁÙ¸µ ÈÞ¸®½ºÆ½À» Á¦¾ÈÇÏ¿´´Ù. Á¦¾ÈÇÏ´Â ÈÞ¸®½ºÆ½Àº º¹¼ö ÀÀ¿ë ½ºÄÉÁÙ¸µÀÇ ½ºÄÉÁÙ °¡´É¼º ¹®Á¦¸¦ °í·ÁÇÏÁö ¸ø ÇÏ´Â ±âÁ¸ÀÇ ¸®½ºÆ® ½ºÄÉÁÙ¸µ ¹æ½Ä ÈÞ¸®½ºÆ½µéÀÇ ÇѰ踦 ±Øº¹ÇÏ¿© ¡®ÇÕ¼º°ú ¹Ýº¹ °³¼±¡¯À̶ó´Â »õ·Î¿î ¹æ½ÄÀ» µµÀÔÇÏ¿´´Ù. ±×¸®ÇÏ¿© CPU, GPU, NPU·Î ±¸¼ºµÇ´Â ÀÌÁ¾ ÇÁ·Î¼¼¼ ȯ°æ¿¡¼ ¿©·¯ µö ·¯´× ³×Æ®¿öÅ©µéÀ» ´ë»óÀ¸·Î ÇÏ´Â ¼º´É ºñ±³ ½ÇÇèÀ» ÅëÇØ Á¦¾ÈÇÏ´Â ÈÞ¸®½ºÆ½ÀÌ ºü¸¥ ½Ã°£ ³»¿¡ ÁÁÀº ½ºÄÉÁÙ¸µÀ» »ý¼ºÇÔÀ» È®ÀÎÇÏ¿´´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
The scheduling of multiple deep learning applications on heterogeneous processors is basically an NP-hard problem with a very large problem space. Meta-heuristics such as GAs (Genetic Algorithms) may be applied, but these have the disadvantage of having too long an execution time to be applied at run time. Therefore, this study proposes a new scheduling heuristic, which complements this shortcoming and does not significantly degrade scheduling performance. The proposed heuristic overcomes the limitations of traditional list scheduling techniques that fail to take into account the schedulability issue in the scheduling of multiple applications and introduces a new approach called ¡®synthesis and iterative improvement¡¯. It is confirmed through experiments with different deep learning networks on heterogeneous processors (including CPUs, GPUs, and NPUs) that the proposed heuristic produces good scheduling results that are sufficiently fast to apply at run time.
|
Å°¿öµå(Keyword) |
¹Ì¼¼¸ÕÁö¾Û
´ë±âÁú
°ø°øµ¥ÀÌÅÍ
À̵¿°æ·Î
À§Ä¡Á¤º¸
¹Ì¼¼¸ÕÁö ÈíÀÔ·®
fine dust app
air quality
open data
movement path
location information
fine dust intake
ÀÓº£µðµå ½Ã½ºÅÛ
µö ·¯´×
ÀÌÁ¾ ÇÁ·Î¼¼¼
½ºÄÉÁÙ¸µ
ÈÞ¸®½ºÆ½
embedded system
deep learning
heterogeneous processors
scheduling
heuristic
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|