TY - GEN
T1 - ABACa: Access Based Allocation on Set Wise Multi-Retention in STT-RAM Last Level Cache
AU - Agarwal, Sukarn
AU - Chakraborty, Shounak
PY - 2021/8/23
Y1 - 2021/8/23
N2 - Exhibition of potential advantages of high density, non-volatility, and low static power consumption makes STTRAM a credible successor to SRAM in caches. However, higher write energy and latency of the STT-RAM limit its potential towards commercial usage. Relaxation of STT-RAM’s retention time can be a viable solution to alleviate these obstacles by reducing both write time and energy. However, significant reduction in retention time might lead to premature expiry of the blocks requiring frequent refreshes or write-backs, which can incorporate unnecessary stalls along with the increased miss-rate.This paper proposes ABACa, an approach that logically bifurcates a cache set-wise for two different retention times where cache blocks are segregated upon their arrival and placed in the corresponding set, accordingly. In particular, if a block’s arrival is triggered by a read miss, the block is placed into a set with a higher retention time, called as read-set. On the other hand, the block is placed into a write-set having a lower retention time, if the block’s arrival is caused by a write miss. Our empirical analysis shows that, ABACa achieves a significant improvement of 40.75% in miss-rate and 61.35% EDP (Energy Delay Product) gain compared to baseline multi-retention STT-RAM-based and SRAM-based last level caches, respectively.
AB - Exhibition of potential advantages of high density, non-volatility, and low static power consumption makes STTRAM a credible successor to SRAM in caches. However, higher write energy and latency of the STT-RAM limit its potential towards commercial usage. Relaxation of STT-RAM’s retention time can be a viable solution to alleviate these obstacles by reducing both write time and energy. However, significant reduction in retention time might lead to premature expiry of the blocks requiring frequent refreshes or write-backs, which can incorporate unnecessary stalls along with the increased miss-rate.This paper proposes ABACa, an approach that logically bifurcates a cache set-wise for two different retention times where cache blocks are segregated upon their arrival and placed in the corresponding set, accordingly. In particular, if a block’s arrival is triggered by a read miss, the block is placed into a set with a higher retention time, called as read-set. On the other hand, the block is placed into a write-set having a lower retention time, if the block’s arrival is caused by a write miss. Our empirical analysis shows that, ABACa achieves a significant improvement of 40.75% in miss-rate and 61.35% EDP (Energy Delay Product) gain compared to baseline multi-retention STT-RAM-based and SRAM-based last level caches, respectively.
KW - multi-retention time
KW - STT-RAM
KW - cache memory
KW - block placement
U2 - 10.1109/ASAP52443.2021.00032
DO - 10.1109/ASAP52443.2021.00032
M3 - Conference contribution
SN - 978-1-6654-2702-9
SP - 171
EP - 174
BT - 2021 IEEE 32nd International Conference on Application-specific Systems, Architectures and Processors (ASAP)
PB - Institute of Electrical and Electronics Engineers
ER -