Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
´ÙÁßÂ÷¿ë ÀΰøÁö´É ½Ã½ºÅÛÀ» À§ÇÑ ¸Þ¸ð¸® ½ºÄÉÁÙ·¯ ÃÖÀûÈ |
¿µ¹®Á¦¸ñ(English Title) |
Optimizing the Memory Scheduler for Multi-tenant Deep Learning Accelerator |
ÀúÀÚ(Author) |
±èÅÂÇö
ÀÌÇõÀç
ÀÌÁøÈ£
Taehyun Kim
HyukJae Lee
Jinho Lee
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 45 NO. 01 PP. 2125 ~ 2127 (2022. 06) |
Çѱ۳»¿ë (Korean Abstract) |
|
¿µ¹®³»¿ë (English Abstract) |
Neural Processing Unit(NPU) is a cheap and practical device-of-choice for future deep learning platform providers. In realistic datacenter settings, many NPUs may share a common memory system while different processes are executed on one or more separate NPUs. Application running on such systems may experience unfair slowdowns due to memory-sharing and fixed memory scheduling policy, which can harm service QoS and throughput. This paper points out that the commonly used First-Ready, First-Come, First-Serve (FRFCFS) policy causes unfairness in memory services among simultaneously-executed processes. Motivated byour findings, we propose an improved policy that can mitigate this problem. |
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|