TY - GEN
T1 - Reviving Zombie Pages on SSDs
AU - Elyasi, Nima
AU - Sivasubramaniam, Anand
AU - Kandemir, Mahmut T.
AU - Das, Chita R.
N1 - Publisher Copyright:
© 2018 IEEE.
PY - 2018/12/11
Y1 - 2018/12/11
N2 - SSDs are being increasingly adopted in the storage hierarchy as they provide substantially lower latencies compared to HDDs. However, there is still a tremendous need for faster service to satisfy the low latency and high bandwidth demands of large-scale computing platforms. Long latency write operations and Garbage Collection (GC) costs are detrimental to the SSD performance as they negatively affect the responsiveness of the drive. Reducing write traffic to the SSD can help significantly lower the response times and minimize the GC costs. In this paper, we show that a majority of pages written to SSD turn into garbage pages. To date, these pages have been immediately (or very soon at the next GC invocations) removed from the device to clean up space for the incoming writes. However, our study in this paper suggests that delaying removal of some garbage pages can help. Specifically, leveraging value popularity to estimate their rebirth can be used to postpone collecting these pages. We design an efficient buffering mechanism to store garbage pages, thereby reducing overheads for GC, and using some of this garbage content to service subsequent writes. Our enhancements considerably reduce the number of writes and lead to lower latency, with an average improvement of 29% and 24.5%, respectively.
AB - SSDs are being increasingly adopted in the storage hierarchy as they provide substantially lower latencies compared to HDDs. However, there is still a tremendous need for faster service to satisfy the low latency and high bandwidth demands of large-scale computing platforms. Long latency write operations and Garbage Collection (GC) costs are detrimental to the SSD performance as they negatively affect the responsiveness of the drive. Reducing write traffic to the SSD can help significantly lower the response times and minimize the GC costs. In this paper, we show that a majority of pages written to SSD turn into garbage pages. To date, these pages have been immediately (or very soon at the next GC invocations) removed from the device to clean up space for the incoming writes. However, our study in this paper suggests that delaying removal of some garbage pages can help. Specifically, leveraging value popularity to estimate their rebirth can be used to postpone collecting these pages. We design an efficient buffering mechanism to store garbage pages, thereby reducing overheads for GC, and using some of this garbage content to service subsequent writes. Our enhancements considerably reduce the number of writes and lead to lower latency, with an average improvement of 29% and 24.5%, respectively.
UR - http://www.scopus.com/inward/record.url?scp=85060276185&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85060276185&partnerID=8YFLogxK
U2 - 10.1109/IISWC.2018.8573481
DO - 10.1109/IISWC.2018.8573481
M3 - Conference contribution
AN - SCOPUS:85060276185
T3 - 2018 IEEE International Symposium on Workload Characterization, IISWC 2018
SP - 156
EP - 167
BT - 2018 IEEE International Symposium on Workload Characterization, IISWC 2018
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2018 IEEE International Symposium on Workload Characterization, IISWC 2018
Y2 - 30 September 2018 through 2 October 2018
ER -