Papers I read in 2021 Q4

The Five-minute Rule Thirty Years Later and its Impact on the Storage Hierarchy

http://www.renata.borovica-gajic.com/data/adms2017_5minuterule.pdf

Appuswamy, Raja, et al. “The five minute rule thirty years later and its impact on the storage hierarchy.” Proceedings of the 7th International Workshop on Accelerating Analytics and Data Management Systems Using Modern Processor and Storage Architectures. No. CONF. 2017.

In 1987, five minute rule, which means data accessed for every 5 or less minutes are better to be in DRAM than HDD, has found. As the storage devices emerge, the rule has revisited for every decade. This paper revisits it again, in 2017.

This paper classifies storage devices in four tiers: performance, capacity, archival, and backup.

Data for latency-sensitive interactive analytics should be in the performance tier, while data for latency-insensitive batch analytics should be in the storage tier.

For state-of-the-art PCIe SSDs, the bread-even interval predicted by the five-minute rule is less than a minute.

The break-even interval between the devices are as below:

• DRAM-HDD: 5m in 1987, 5m in 1997, 1.5h in 2007, and 4h in 2017
• DRAM-SSD: 15m in 2007, 7m for read / 24m for write in 2017 (a sentence in the paper says this is only 40 seconds in 2017, confusing… Maybe for 3D XPoint like things only)
• SSD-HDD: 2.26h in 2007, 1d in 2017

Exploring the Design Space of Page Management for Multi-Tiered Memory Systems

https://www.usenix.org/conference/atc21/presentation/kim-jonghyeon

Kim, Jonghyeon, Wonkyo Choe, and Jeongseob Ahn. “Exploring the Design Space of Page Management for Multi-Tiered Memory Systems.” 2021 {USENIX} Annual Technical Conference ({USENIX}{ATC} 21). USENIX} Association, 2021.

This paper explores a better way for tiered memory management. It considers access tier of each page in addition to the access locality (whether the page is placed on the local memory or remote memory). That is, Linux uses local PMEM instead of remote DRAM, when local DRAM is full. This paper says it should use remote DRAM instead in the case, as remote DRAM is faster than local PMEM. Based on the finding, the authors proposes their new page placement scheme, which uses AutoNUMA facility to check access frequency and do demotion/promotion proactively. The authors says it achieves 3.5x speedup compared to Intel’s recent approach, which Huang Ying is doing.

For the evaluation, the authors has used graph500, SpecACCEL, GraphMat, and Liblinear.