Re: W-TinyLfu for cache eviction

On 03.12.2015 10:27, Vladimir Sitnikov wrote:
> I've recently noticed W-TinyLfu cache admission policy (see [1]) being
> used for caffeine "high performance caching library for Java 8".
> It demonstrates high cache hit ratios (see [2]) and enables to build
> high-throughput caches (see caffeine in [3])
> Authors explicitly allow implementations of the algorithm (see [4]).
>
> Does it make sense to evaluate the algorithm for buffer replacement?

I expect synchronization issues with implementation of this algorithm.
It seems to be hard to avoid some global critical section which can
cause significant performance degradation at MPP systems (see topic
"Move PinBuffer and UnpinBuffer to atomics").

> [1]: http://arxiv.org/pdf/1512.00727v1.pdf
> [2]: https://github.com/ben-manes/caffeine/wiki/Efficiency
> [3]: https://github.com/ben-manes/caffeine/wiki/Benchmarks
> [4]: https://github.com/ben-manes/caffeine/issues/23#issuecomment-161536706
>
> Vladimir Sitnikov
>
>