RAIDR: Retention-Aware Intelligent DRAM Refresh 🔗
This work proposed Retention-Aware Intelligent DRAM Refresh (RAIDR), a low-cost mechanism that can identify and skip unnecessary refreshes using knowledge of cell retention times. Our key idea is to group DRAM rows into retention time bins and apply a different refresh rate to each bin.
Strengths
- The author grouped the previous mechanisms over smart retention time into two types: hardware-only and hardware-software.
- RAIDR stores retention time bins with Bloom filters which allows low storage overhead and ensures that bins never overflow. This function is independent to specific DRAM chips.
- RAIDR is flexible, so a system designer can stirke a balance between implementation overhead and refresh reduction.
Weaknesses
- Flexibility always leads to troubles, especially how to adjust the parameters to balance the tradeoff between overhead and refresh reduction, and eventually lead to better memory system performance.
- The DRAM organization figures are not as cute as the ones in other DRAM papers such as Ambit.
- RAIDR only cares about energy and performance, while security may also be important, especially the RowHammer problem, which is ignored.
Can you do better?
To better visualize, I would draw some timeline figures to present how RAIDR handles the various retention time between different DRAM cells.
I would consider RowHammer in this work, as it is a really significant problem of DRAM related to retention. It would be better if there is some discussion over how it defense RowHammer.
Takeways
- RAIDR’s main benefits might be the robust to variation in different DRAM systems environments, and increase as memory volume arises.
Other Comments
- The utilization of bloom filter is brilliant and I think this is the essence of this work. I might consider embed this design into other works.