Á¤º¸°úÇÐȸ ³í¹®Áö A : ½Ã½ºÅÛ ¹× ÀÌ·Ð
ÇѱÛÁ¦¸ñ(Korean Title) |
ºÐ»ê °øÀ¯ ¸Þ¸ð¸® ½Ã½ºÅÛ¿¡¼ °ÅÁþ °øÀ¯ Á¦°Å ¹× Åë½Å·® °¨¼Ò ±â¹ý |
¿µ¹®Á¦¸ñ(English Title) |
Techniques for Eliminating False Sharing and Reducing Communication Traffic in Distributed Shared Memory Systems |
ÀúÀÚ(Author) |
ÇѺÎÇü
Á¶¼ºÁ¦
Á¶À¯±Ù
Boohyung Han
Seongje Cho
Yookun Cho
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 25 NO. 10 PP. 1100 ~ 1108 (1998. 10) |
Çѱ۳»¿ë (Korean Abstract) |
°ÅÁþ °øÀ¯(False sharing)´Â ÇϳªÀÇ ÀÏ°ü¼º À¯Áö ´ÜÀ§ ³»¿¡ Á¸ÀçÇÏ´Â ¿©·¯ °³ÀÇ ´Ù¸¥ ¿öµå¸¦ ¼·Î ´Ù¸¥ 󸮱âµéÀÌ ÂüÁ¶ÇÒ ¶§ ¹ß»ýµÇ¸ç, ½Ã½ºÅÛ ¼º´ÉÀ» ÀúÇϽÃÅ°´Â ÁÖ¿ä ¿øÀÎÀÌ´Ù. º» ³í¹®¿¡¼´Â ÆäÀÌÁö ±â¹Ý ºÐ»ê °øÀ¯ ¸Þ¸ð¸® ½Ã½ºÅÛ »ó¿¡¼ º´·Ä ÇÁ·Î±×·¥µéÀÇ °øÀ¯ ¸Þ¸ð¸® ÂüÁ¶ ÆÐÅÏÀ» ºÐ¼®ÇÏ°í, °ÅÁþ °øÀ¯¿Í Åë½Å·®À» °¨¼Ò½ÃÄÑ ÁÖ´Â È¿À²ÀûÀÎ µ¿Àû °øÀ¯ ¸Þ¸ð¸® ÇÒ´ç ±â¹ýÀ» Á¦½ÃÇÑ´Ù. Á¦½ÃÇÑ ÇÒ´ç ±â¹ý¿¡¼´Â ÂüÁ¶ ÆÐÅÏÀÌ ¼·Î ´Ù¸¥ µ¥ÀÌŸ °´Ã¼°¡ µ¿ÀÏ ÆäÀÌÁö¿¡ °øÁ¸ÇÏ´Â °ÍÀ» ¹æÁöÇϱâ À§ÇØ, µ¥ÀÌŸ °´Ã¼ÀÇ Å©±â°¡ °¡»ó ÆäÀÌÁö Å©±âº¸´Ù ÀÛÀº °æ¿ì¿Í Å« °æ¿ì·Î ³ª´©¾î °¢°¢ ´Ù¸¥ Á¤Ã¥À» »ç¿ëÇÑ´Ù. °´Ã¼°¡ ÆäÀÌÁö Å©±âº¸´Ù ÀÛÀº °æ¿ì¿¡´Â µ¿ÀÏ ÆäÀÌÁö¿¡ ¼·Î ´Ù¸¥ ³ëµå°¡ ¿äûÇÑ °´Ã¼µéÀÌ ¼¯ÀÌÁö ¾Êµµ·Ï ÇÏ°í, µ¥ÀÌŸ °´Ã¼°¡ ÆäÀÌÁö Å©±âº¸´Ù Å« °æ¿ì¿¡´Â °´Ã¼¸¦ ¿©·¯ ÆäÀÌÁö °æ°è¿¡ °ÉÄ¡Áö ¾Êµµ·Ï Á¤·ÄÇÔÀ¸·Î½á ÇϳªÀÇ °´Ã¼°¡ ¿©·¯ ÆäÀÌÁö¿¡ °ÉÃļ ÇÒ´çµÇ´Â °æ¿ì¸¦ ÃÖ¼ÒÈ ÇÏ¿© °ÅÁþ °øÀ¯ÀÇ °¡´É¼ºÀ» ÁÙÀδÙ. º» ³í¹®¿¡¼´Â ½ÇÇèÀ» ÅëÇØ »õ ÇÒ´ç ±â¹ýÀÇ °ÅÁþ °øÀ¯ ÆúÆ® ¼ö¿Í Åë½Å ¿À¹öÇìµå¸¦ ÃøÁ¤ÇÏ¿´´Ù. Á¦½ÃÇÑ ÇÒ´ç ±â¹ýÀ» ½ÇÇèÇÑ °á°ú ±âÁ¸ÀÇ °øÀ¯ ¸Þ¸ð¸® ÇÒ´ç ±â¹ýµéº¸´Ù °ÅÁþ °øÀ¯¿Í Åë½Å·®ÀÌ Å©°Ô °¨¼ÒµÇ¾ú´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
False sharing is a problem which occurs when different processors reference disjoint sets of words within a coherence block, and it degrades system performance. We analyze shared memory reference patterns in parallel programs, and then propose an efficient dynamic shared memory allocator for reducing false sharing in page-based distributed shared momory systems. To prevent the data objects that have different reference patterns from being placed in the same page, our allocator adopts one of two policies according to the size of an object. If the size of an object is smaller than the page size, our allocator allocates each object requested by different processors to different pages. Otherwise, the allocator allocates memory space by aligning the object not to span the page boundaries. We measure the number of false sharing faults and communication traffic overhead by experiments.
Experimental results show that our scheme reduces false sharing and communication traffic greatly commpared to the existing shared memory allocation schemes. |
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|