Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
ÇѱÛÁ¦¸ñ(Korean Title) |
°¡»óÈ È¯°æ¿¡¼ÀÇ NUMA ±â¹Ý ¶ô ¼º´É ºÐ¼® |
¿µ¹®Á¦¸ñ(English Title) |
Performance Analysis of NUMA-aware Locks in Virtualized Systems |
ÀúÀÚ(Author) |
±è½ÂÀ±
ÀÌÅÂÇü
¾ö¿µÀÍ
Seungyun Kim
Taehyung Lee
Young Ik Eom
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 27 NO. 02 PP. 0104 ~ 0109 (2021. 02) |
Çѱ۳»¿ë (Korean Abstract) |
ÃÖ±Ù CPU ¹× ¸Þ¸ð¸® Ãø¸éÀÇ È®À强À» À§ÇØ µ¥ÀÌÅÍ ¼¾ÅÍ ¹× Ŭ¶ó¿ìµå Ç÷§Æû¿¡¼´Â NUMA ¼¹ö¸¦ ¸¹ÀÌ È°¿ëÇÏ°í ÀÖ´Ù. ÀÌ¿¡ µû¶ó ½Ã½ºÅÛÀÇ ¼º´É ÃÖÀûȸ¦ À§ÇØ ºñ°¡»óÈ ½Ã½ºÅÛÀ» ±â¹ÝÀ¸·Î NUMA ±¸Á¶ÀÇ Æ¯Â¡À» È°¿ëÇÑ ¿¬±¸°¡ ¸¹ÀÌ ¼öÇàµÇ¾ú´Ù. °¡»óÈ ½Ã½ºÅÛ ¿ª½Ã NUMA ½Ã½ºÅÛ¿¡¼ÀÇ ¼º´É ÃÖÀûȸ¦ À§ÇØ vNUMA ±â´ÉÀ» °¢ °¡»ó¸Ó½Å¿¡ Á¦°øÇÏ°í ÀÖÀ¸¸ç, À̸¦ ÅëÇØ °Ô½ºÆ® ¿î¿µÃ¼Á¦°¡ NUMA ÃÖÀûÈ ±â¹ýÀ» È°¿ëÇÒ ¼ö ÀÖµµ·Ï ÇÏ°í ÀÖ´Ù. ÇÏÁö¸¸, ´ëºÎºÐÀÇ NUMA ÃÖÀûÈ ±â¹ýµéÀº ³×ÀÌƼºê ȯ°æ¿¡¼ ¼öÇàµÇ¾ú±â ¶§¹®¿¡ ¿©·¯ °¡»ó¸Ó½ÅÀÌ ÄÄÇ»Æà ÀÚ¿øÀ» °øÀ¯Çϴ ȯ°æ¿¡¼´Â ÀÌ·± ±â¹ýµéÀÇ ¼º´É ÀÌÁ¡ÀÌ Å©°Ô ÁÙ¾îµé ¼ö ÀÖ´Ù. º» ³í¹®¿¡¼´Â vNUMA ±â¹Ý °¡»ó¸Ó½Å ȯ°æ¿¡¼ NUMA ¶ôÀÇ ¼º´ÉÀ» Ä¿³Î ¿µ¿ª ¹× À¯Àú ¿µ¿ª¿¡¼ ºÐ¼®ÇÏ¿´´Ù. ½ÇÇè °á°ú, ´ÜÀÏ °¡»ó¸Ó½Å ȯ°æ¿¡¼´Â NUMA ¶ôÀÇ ¼º´ÉÀÌ È£½ºÆ®¿Í À¯»çÇÏ°Ô ³ªÅ¸³ªÁö¸¸, ´ÙÁß °¡»ó¸Ó½Å ȯ°æ¿¡¼´Â LWP ¹®Á¦·Î ÀÎÇØ ¼º´É ÀÌÁ¡ÀÌ Å©°Ô °¨¼ÒÇÏ´Â °ÍÀ» È®ÀÎÇÏ¿´´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
Recently, NUMA-based servers have been used in data centers and cloud platforms for scalability of the CPU and memory resources. Accordingly, many studies have been conducted in native systems in attempts to optimize the system performance by exploiting the characteristics of the NUMA architecture. The virtualized system also provides virtual NUMA(vNUMA) to each virtual machine to allow guest operating systems to utilize the NUMA-related optimization technique. Unfortunately, since most NUMA-related optimization studies have been conducted in native systems, the performance benefits of these techniques are likely to be significantly reduced in environments where multiple virtual machines(VMs) share computing resources. In this paper, we analyze the performance of the NUMA-aware lock, which is applied to the kernel-space lock and the user-space lock, on vNUMA-based VM environments. The experimental results show that, in a single VM environment, the performance of the NUMA-aware lock is similar to that of the host, whereas in multiple virtual machine environments, the performance benefits of the NUMA-aware lock are significantly reduced due to LWP problems.
|
Å°¿öµå(Keyword) |
NUMA ±¸Á¶
¶ô
°¡»ó¸Ó½Å
vCPU ¼±Á¡
NUMA architecture
lock
virtual machine
vCPU preemption
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|