Today’s paper is “FIFO can be Better than LRU: the Power of Lazy Promotion and Quick Demotion” authored by Juncheng Yang and co. With FIFO, the oldest object in the cache is the first one to be evicted; with LRU, it’s the least recently used object The advantages of FIFO over LRU were already teased out in their previous paper, what we’re getting here is a much more extensive evaluation plus two new-ish techniques.
When analyzing cache efficiency, the two key metrics usually considered are miss ratio and throughput (number of requests served per a given period). With the latter metric, FIFO has the following advantages over LRU [1]:
- Less computation particularly on cache hits
- Fewer metadata to keep track of and update per object cached
- More scalable - less likely to be the bottleneck as number of threads increases
However, it’s often assumed that plain FIFO has worse performance (higher miss ratio) compared to LRU. The question then is how can this performance gap be addressed. The authors propose two techniques: lazy promotion and quick demotion.
To understand these 2 techniques, let’s consider a generic cache. As users, we can insert objects into the cache and remove them either explicitly or implicitly via TTLs. Regardless of the underlying algorithm, all caches maintain some internal notion or measurement for which objects are more valuable and which ones are less valuable and can be evicted if need be. Whenever the cache increases its valuation of some object, we say that the object has been promoted; when the value is decreased, we say it’s been demoted.
Let’s consider textbook LRU. Whenever there’s a cache hit, the object accessed is moved to the head of the list. As such, LRU carries out promotion explicitly (moving the object to the head) - demotion is implicit, passive and kind of drawn-out (all other objects are shifted backwards).
The authors argue that we should gun for the opposite: lazy promotion and quick demotion.
Lazy Promotion #
Let’s start with Lazy Promotion. With plain FIFO, regardless of how popular an object is, as newer objects get inserted, the popular object gets closer and closer to eviction. We don’t want to carry out some form of promotion with every cache hit since it means we’re doing more work per request hence we end up handling fewer requests per second. Instead, we keep around some tiny metadata, even a bit or two-bits that indicate the object is popular. Then when it’s about to get evicted, we instead re-insert it (i.e. promote it) and reset the popularity metrics. In fact, if you took LRU, applied lazy promotion and instead of a queue, you’ve got a circular buffer and a ‘hand’, you’d end up with Clock.
Quick Demotion #
Quick Demotion entails removing low value objects as soon as possible rather than keeping them around in the cache. The authors make the following observation: “Because demotion happens passively in most eviction algorithms, an object typically traverses through the cache before being evicted. Such traversal gives each object a good chance to prove its value to be kept in the cache. However, cache workloads often follow Zipf popularity distribution with most objects being unpopular … We believe the opportunity cost of new objects demonstrating their values is often too high: the object being evicted at the tail of the queue may be more valuable than the objects recently inserted” [1]. I’d rephrase it as follows: newer objects are more likely to be of low value compared to older objects hence caching algorithms should first consider newer ones for eviction. Also, the authors point out that QD delivers the most gain when the cache size is large, objects are short-lived and the underlying cache algorithm is not quite efficient.
Implementing QD #
The diagram above provides an overview of how QD can be implemented. A couple of points worth adding: the main cache uses 90% of the space and the probationary queue uses 10%. Also, the ghost queue stores as many entries as the main cache. On configuring the size of the ghost queue, you’d have to assume all objects have the same size; veering off from this assumption seems to be left as an exercise for the dear reader :)