Our great sponsors
-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
Yes, I think that is my main concern in that often research papers do not disclose the weaknesses of their approaches and the opposing tradeoffs. There is no silver bullet.
The stress workload that I use is to chain corda-large [1], 5x loop [2], corda-large at a cache size of 512 entries and 6M requests. This shifts from a strongly LRU-biased pattern to an MRU one, and then back again. My solution to this was to use hill climbing by sampling the hit rate to adaptively size of the admission window (aka your FIFO) to reconfigure the cache region sizes. You already have similar code in your CACHEUS implementation which built on that idea to apply it to a multi-agent policy.
Caffeine adjusts the frequency comparison for admission slightly to allow ~1% of losing warm candidates to enter the main region. This is to protect against hash flooding attack (HashDoS) [3]. That isn't intended to improve or correct the policy's decision making so should be unrelated to your observations, but an important change for real-world usage.
I believe LIRS2 [4] adaptively sizes their LIR region, but I do not recall the details as a complex algorithm. It did very well across different workloads when I tried it out and the authors were able to make a few performance fixes based on my feedback. Unfortunately I find LIRS algorithms to be too difficult to maintain for an industry setting because while excellent, the implementation logic is not intuitive which makes it frustrating to debug.
[1] https://github.com/ben-manes/caffeine/blob/master/simulator/...