pingora/tinyufo

TinyUFO

TinyUFO is a fast and efficient in-memory cache. It adopts the state-of-the-art S3-FIFO as well as TinyLFU algorithms to achieve high throughput and high hit ratio as the same time.

Usage

See docs

Performance Comparison

We compare TinyUFO with lru, the most commonly used cache algorithm and moka, another great cache library that implements TinyLFU.

Hit Ratio

The table below show the cache hit ratio of the compared algorithm under different size of cache, zipf=1.

cache size / total assets	TinyUFO	TinyUFO - LRU	TinyUFO - moka (TinyLFU)
0.5%	45.26%	+14.21pp	-0.33pp
1%	52.35%	+13.19pp	+1.69pp
5%	68.89%	+10.14pp	+1.91pp
10%	75.98%	+8.39pp	+1.59pp
25%	85.34%	+5.39pp	+0.95pp

Both TinyUFO and moka greatly improves hit ratio from lru. TinyUFO is the one better in this workload. This paper contains more thorough cache performance evaluations S3-FIFO, which TinyUFO varies from, against many caching algorithms under a variety of workloads.

Speed

The table below shows the number of operations performed per second for each cache library. The tests are performed using 8 threads on a x64 Linux desktop.

Setup	TinyUFO	LRU	moka
Pure read	148.7 million ops	7.0 million ops	14.1 million ops
Mixed read/write	80.9 million ops	6.8 million ops	16.6 million ops

Because of TinyUFO's lock-free design, it greatly outperforms the others.

Memory overhead

The table below show the memory allocation (in bytes) of the compared cache library under certain workloads to store zero-sized assets.

cache size	TinyUFO	LRU	moka
100	39,409	9,408	354,376
1000	236,053	128,512	535,888
10000	2,290,635	1,075,648	2,489,088

Whether these overheads matter depends on the actual sizes and volume of the assets. The more advanced algorithms are likely to be less memory efficient than the simple LRU.