Usage of cache admission algorithms with plugin cache?

[AlexSutila]

Hello,

I am aware from the original LibCacheSim repo that it implements various cache admission algorithms. I was wondering, what is the feasability of simulating workloads using PluginCache which incoorporate these admission algorithms?

If it is doable, what is the best way to approach this? If any code changes are required, I would also be willing to assist in any way I can. Thanks in advance!

[1a1a11a]

We do not have support for admission in the Python binding at the moment. If you are interested, you can propose the change and we can review it @haochengxia .

[haochengxia]

Thanks for your interest. We’re open to adding support for admission algorithms. The simplest approach, in my view, is to provide an additional optional hook for an admissioner, allowing us to directly follow the logic of cache_get_base.

[AlexSutila]

Thanks for your interest. We’re open to adding support for admission algorithms. The simplest approach, in my view, is to provide an additional optional hook for an admissioner, allowing us to directly follow the logic of cache_get_base.

This makes sense and is pretty in line with what I'm thinking at the moment, though there are a few more subtle details I wanna get approval of before I take the time to actually go through with implementation.

My current plan is to expose the admissioner struct via pybind and in a very similar fashion to what you guys already do in libcachesim/cache.py and existing caching algorithms, have an abstract base class which provides an interface for existing admission algorithms within libcachesim.

My intention is also to include something along the lines of a CustomAdmissioner (naming is subject to change of course) allowing for custom admission algorithms to be developed conveniently from within python. Does this sound reasonable?

[1a1a11a]

This sounds good to me! A few comments

• The admissioner module in libCacheSim was not carefully designed. If you see a need to change, e.g., not exposing enough information to support different algorithms, please create a separate issue and PR to update it.
• One thing we need to be careful about is overhead, because admission is not used in a few common use cases. We want to limit the overhead, so keep this in mind and share some performance numbers after you finish.

[AlexSutila]

Quick update. I have a an initial draft ready for a PR which includes support for all existing admission algorithms and also the ability to create custom admission policies through hooks like I described before. I was able to do all of this without needing any modifications to libCacheSim itself. I also ended up using the existing script to run some benchmarks.

I did two runs on a metaCDN trace the latest commit of my fork and on the latest commit before I made any changes. I've included both resulting screenshots below.

1. First image is from the latest commit on my fork
2. Second image is from the older commit before any changes I introduced

• 
• 

If you feel it is necessary, I can also get numbers from some more traces and/or also look into modifying the benchmarking script to test the performance while utilizing an admission policy before submitting a PR.

[1a1a11a]

The performance looks reasonable to me.
I will leave it to @haochengxia to review and decide.

[1a1a11a]

BTW, I like the figures you created, maybe we can have a PR to add them to readme.

[1a1a11a]

What does miss ratio consistency mean?

[haochengxia]

Added by #54

[haochengxia]

What does miss ratio consistency mean?

We use a box plot to illustrate whether a method produces the same miss ratio across multiple runs. While results differ between the CNATIVE and Python plugins due to their settings, each method yields a deterministic miss ratio. We plan to refine the benchmark scripts further.