4 ms·
The LIRS papers do a good job of taking different workload patterns, such as loop and zig-zag and mixing multiple patterns together, to validate that their algo
by NovaX 3y ago
The LIRS papers do a good job of taking different workload patterns, such as loop and zig-zag and mixing multiple patterns together, to validate that their algorithm is robustly good as a general purpose cache. This seems to take traces with similar workload patterns that are distinguished by their excessive size to assert the algorithm's quality. When papers do this it looks good from an academic sense, but it is frustrating from an industry perspective because the workload may not be known upfront or the one must cherry-pick an algorithm for a bespone, consistent pattern. I find it more interesting when algorithm designers try to stress their work and discuss how/why it fails or succeeds. That gives more confidence to those of us who want to implement and use the work in a practical setting.
- 1a1a11a 3y agoThe large datasets from 2007 to 2020 cover a variety of workload patterns. If you have a workload that you would like to try out, our code and data are all open-sourced. Also feel free to drop us an email :)
- 1a1a11a 3y agoBTW, the 1% queue in LIRS makes it similar to W-TinyLFU (but less robust than W-TinyLFU)
- NovaX 3y agoYou have a large corpus of similar workload patterns. Simply using a 750GB trace or many samples from a cluster of machines doing the same work does not make it a meaningful variety. It would be reasonable to have also tested with the traces that your comparison algorithms used, which shows your implementations are valid and how they differ on those authors terms. Skipping that may come across as cherry-picking because I find that QDLP under performs by -20% in zigzag, -15% in DS1, -33% in blockchain mining. Your algorithm is good in specific cases, like many others are, but there are many real workloads where it is not usable.