KSC 2020
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
ÄÁÅ×ÀÌ³Ê ±â¹ÝÀÇ È®À强 ÀÖ´Â Fabric-Attached Memory °ü¸® Ç÷§Æû |
¿µ¹®Á¦¸ñ(English Title) |
Scalable Container-based Software Platform for Fabric-Attached Memory Pool |
ÀúÀÚ(Author) |
ÀÌÈñ¶ô
ÀÌâ±Ô
°Çö±¸
¸¶Áø¼®
¼®¼º¿ì
¿À¸íÈÆ
±è¿µÀç
Heerock Lee
Chang-Gyu Lee
Awais Khan
Hyeongu Kang
Jinseok Ma
Song-Woo Suk
Myeonghoon Oh
Youngjae Kim
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 47 NO. 02 PP. 0997 ~ 0999 (2020. 12) |
Çѱ۳»¿ë (Korean Abstract) |
|
¿µ¹®³»¿ë (English Abstract) |
Recent progress in Non-Volatile Memory (NVM) and Gen-Z Interconnect introduce a new type of memory pool called Fabric-Attached Memory(FAM). Until today, the massive memory pools are built via shared memory but often with severe overhead. Unlike existing solutions, FAM provides scalable memory pool with a near-local access latency. However, such hardware changes entail huge changes in software platform. To this end, we propose a scalable software platform based on the container which is widely deployed in many scalable systems. First, our platform provides extensible control plane to containers by employing broker container that accepts OpenFAM APIs. Second, we designed the data plane using bind mount to keep FAM¡¯s direct access with memory semantic. Our control plane showed only average 7.62% drops in IOPS and 8.13% latency increase compared to without container. Also there is only negligible overhead in our data plane.
|
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|