Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
¸Å´ÏÄÚ¾î ½Ã½ºÅÛÀ» À§ÇÑ MPI ³ëµå ³» Åë½Å ÇÁ¸®¹ÌƼºê ¼º´É ºñ±³ |
¿µ¹®Á¦¸ñ(English Title) |
Performance Comparison of MPI Intra-node Communication Primitives for Manycore System |
ÀúÀÚ(Author) |
Á¶Áß¿¬
ÁøÇö¿í
Joong-Yeon Cho
Hyun-Wook Jin
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 25 NO. 07 PP. 0357 ~ 0362 (2019. 07) |
Çѱ۳»¿ë (Korean Abstract) |
MPI ÇÁ·Î±×·¡¹Ö ¸ðµ¨ÀÇ ³ëµå ³» Åë½ÅÀ» À§ÇØ °í¾ÈµÈ ¸Þ¸ð¸® ¸ÅÇÎ ¹æ¹ýÀº ÇÑÂ÷·ÊÀÇ µ¥ÀÌÅÍ º¹»ç¸¸À¸·Î ³ëµå ³» Åë½ÅÀ» ¼öÇàÇÒ ¼ö ÀÖ´Ù. ¸Å´ÏÄÚ¾î ½Ã½ºÅÛÀÇ ¹ßÀü¿¡ µû¶ó MPI ÇÁ·Î±×·¡¹Ö ¸ðµ¨ÀÇ ³ëµå³» Åë½Å ºñÀ²ÀÌ Áõ°¡ÇÏ°í ÀÖ¾î ¸Þ¸ð¸® ¸ÅÇÎ ¹æ¹ýÀ» À§ÇÑ Åë½Å ÇÁ¸®¹ÌƼºêµéÀÇ Æ¯Â¡À» ÀÌÇØÇÏ°í º´¸ñÁöÁ¡À» ¿ÏÈÇÑ´Ù¸é ¸Å´ÏÄÚ¾î ½Ã½ºÅÛ¿¡¼ MPI ³ëµå ³» Åë½Å ¼º´É Çâ»óÀÌ ±â´ëµÈ´Ù. º» ³í¹®¿¡¼´Â ¸Þ¸ð¸® ¸ÅÇÎ ¹æ¹ýÀ» À§ÇÑ Åë½Å ÇÁ¸®¹ÌƼºêÀÎ LiMIC, CMA, XPMEMÀ» ÀÌ¿ëÇØ Á¡´ëÁ¡ Åë½Å ¹× ºê·Îµåij½ºÆ® ÁýÇÕÅë½Å À¯ÇüÀÇ Áö¿¬½Ã°£°ú ¼ö½Å Ãø ´Ü°èº° ¿À¹öÇìµå¸¦ ºñ±³ÇÑ´Ù. ½ÇÇè °á°ú Á¡´ëÁ¡ Åë½Å À¯ÇüÀÇ °æ¿ì XPMEMÀÌ LiMIC ¹× CMA¿Í ºñ±³ÇßÀ» ¶§ ÃÖ´ë 7.4¹è ³ôÀº Áö¿¬½Ã°£À» º¸À̸ç, ºê·Îµåij½ºÆ® ÁýÇÕ Åë½ÅÀ¯ÇüÀÇ °æ¿ì XPMEMÀÌ LiMIC ¹× CMA¿Í ºñ±³ÇßÀ» ¶§ ÃÖ´ë 3.8¹è ³ôÀº Áö¿¬½Ã°£À» º¸ÀδÙ.
|
¿µ¹®³»¿ë (English Abstract) |
The memory-mapping-based mechanism designed for intra-node communication of MPI programming model can perform intra-node communication with single data copy. There has been an increase in the ratio of intra-node communication in MPI programming model due to the development of the manycore system. It is expected that improvement in MPI intra-node communication performance in the manycore system can be achieved by understanding the characteristics and mitigating overhead of communication primitives for memory-mapping-based mechanism. In this paper, we compare latency and receiver side overhead of point-to-point communication and broadcast collective communication using LiMIC, CMA, and XPMEM which communication primitive for memory-mapping-based mechanism. Based on the results it is clear that XPMEM exhibits up to 7.4 times higher latency compared to LiMIC and CMA in case of point-to-point communication and XPMEM exhibits up to 3.8 times higher latency compared to LiMIC and CMA in case of broadcast collective communication.
|
Å°¿öµå(Keyword) |
MPI
³ëµå ³» Åë½Å
¸Þ¸ð¸® ¸ÅÇÎ ±â¹Ý
Á¡´ëÁ¡ Åë½Å
ÁýÇÕ Åë½Å
MPI
intra-node communication
memory-mapping-based
point-to-point communication
collective communication
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|