I think you may be right. The library author said he made many changes because implementing the eviction policy following the paper's design was pretty awful.

I think you may be right. The library author said he made many changes because implementing the eviction policy following the paper's design was pretty awful. > First, I tried to implement cache on hash table and lock free queues, as the authors wrote. I was annoyed, s3-fifo was indeed many times faster than lru cache implemented on hash table with global mutex, but still lost to ristretto and theine, which use bp-wrapper techniques. The profiler also showed that it was the eviction policy that was spending the bulk of the time. And I was only able to beat ristretto and theine after rewriting the implementation using bp-wrapper. > In summary, I completely disagree about the scalability advantage of S3-FIFO over the other policies > Though to be honest, what was causing the hit ratio to go down in DS1 I still don't understand https://github.com/Yiling-J/theine-go/issues/29#issuecomment-1841356725 https://github.com/Yiling-J/theine-go/issues/29#issuecomment...