InfiniCache: Exploiting Ephemeral Serverless Functions to Build a Cost-Effective Memory Cache

非常有意思（虽然不一定实用），用 lambda function 作 object caching，在 light load 下省钱效果佳。

Abstract

  • exploit serverless functions’ memory resources to enable elastic pay-per-use caching
    • erasure coding
    • intelligent billed duration control
    • data backup mechanism

关注 large-object-only production workload

Introduction

In-Memory Object Caches (IMOCs) are largely used only as a small cache for buffering small-sized objects.

Large objects (MBs-GBs) consume significant memory capacity and network bandwidth + cause cache churn with evictions of many small objects.

IBM Docker registry traces: large objects also have good locality.

Providers only charge tenants when a function is invoked. 这个 cache 如果被 access 的频率不高，那会有可能比按整个时间段存的 cache 更划算。

Background and Motivation

large object caching 的研究比较少（相比 Redis / Memcached 这些）。

Large Object Caching

  • Extreme Variability in Object Size
  • Tension between Small and Large Objects
    • large objects 加起来传输占比超过 95%，但 IOPS < 3500 GETs/hour
  • Caching Large Object Matters
    • 37%-46% large objects are reused within 1 hour since last time they were accessed

Building a Memory Cache on Cloud Functions: Opportunities and Challenges

opportunities:

  • pay-per-use pricing
  • short-term caching

challenges of lambda function:

  • a limited CPU and memory capacity
  • run at most 900sec
  • only allow outbound TCP network connections (not stateful)
  • lack of quality-of-service (QoS) control

InfiniCache Design

略，参考以下即可：