Desperately Seeking ... Optimal Multi-Tier Cache Configurations

on YouTube ↗15m

" summary

Extends PyMimircache into an n‑level I/O simulator with write‑through/write‑back support, trace upscaling and reverse miniature sampling, and stores experiment logs in MySQL and Elasticsearch to compute metrics like throughput per dollar. Runs simulations across 36 MSR traces with nine device models (three DRAM/SSD/HDD classes) and evaluates full configs of L1–L2–backend. Finds that buying more low‑end DRAM can outperform high‑end DRAM at the same cost (example: ~65 KB/s gain) and that SSD aging (2–3x latency) often negates SSD benefits.

AI-written summary · about

" chapters (13)

1. 0:00Intro
2. 0:42Motivation
3. 1:22Key Problems
4. 2:03Challenges Physical experiments are $costly and time-consuming
5. 2:55Cache Analysis Metrics
6. 3:58Multi-Tier Analysis Performance gains do not translate to $cost efficiency • Behavior in one tier cascades throughout the system • Device characteristics of each tier can widely vary
7. 5:34Multi-Tier I/O Simulator Features
8. 6:51Reverse Miniature Simulations
9. 8:34PyMimircache Extension
10. 10:08Outline
11. 11:01Evaluation 1: DRAM Performance 300
12. 12:44Evaluation 2: SSD Degradation
13. 13:55Conclusion