Demystifying Cache Policies for Photo Stores at Scale: A Tencent Case Study

Abstract

  • a study on internet-scale photo caching algorithms in the case of QQPhoto from Tencent Inc.: even advanced cache algorithms can only perform at a similar level as simple baseline algorithms
  • with appropriate prefetching, improve the cache hit ratio and reduce the average access latency at a marginal cost of 4.14% backend network traffic

Introduction

  • It is particularly challenging to design effective cache algorithms for photo caching workloads, as the access patterns are rather complicated and extremely difficult to predict due to the nature of multi-tenancy and extensive-sharing in the cloud web environment.
  • We find that the major factor limiting hit ratio improvement is the compulsory misses or cold misses. Moreover, we have observed that a majority of photos are requested in a limited period of the time following the photo uploading, which we call the immediacy feature. Leveraging this feature, we propose to augment a prefetcher to the photo cache to reduce compulsory misses.

Background and Motivation

QQPhoto Preliminaries

In QQPhoto, photo writes and reads are routed through separate upload and download channels and the QQPhoto cache stack specifically refers to the download cache path.

Photo Traces and Sampling Method

reservoir sampling 选 1/100 的 key，之后只观察这些 key 的 reference（跟 SHARDS 有点像）

Advanced Algorithms

Motivation

因为高级算法没用，不如另辟蹊径。

Photo Cache Prefetching

不同 resolution 可能会影响命中，所以只需要 prefetch 最热的那个 resolution。因为 prefetch 一些照片到 cache，也涉及到在新的情况下如何做 eviction。

Evaluation

暂时没看

评论

不明白 resolution 是什么。文章看得我晕晕的。感觉太针对这个 traces 了，不够 general and open-minded。但有一个神奇的好处是，photo caching trace 可以比较方便地改成 object storage traces。