EAGER: Elastic Multi-layer Memcached Tiers

Facebook和YouTube拥有包含许多对象的海量数据库。对象(如图片、视频)从这些数据库中读取并呈现给用户。这些数据库使用速度较慢且成本较低的I/O设备。重复的数据库查询速度也很慢，因此公司会缓存流行的内存对象(RAM)。由于RAM的速度大约是I/O的1000倍，因此对流行对象的查询得到服务的速度更快。单台计算机不能有效地使用太多的RAM，因此使用了一种名为“Memcached”的分布式内存缓存系统。Memcach创建一个节点集群，在更快的内存中形成一个键值数据库。Memcach的简单架构导致了它的流行，但也是它的致命弱点。当负载增加时，需要更多的缓存节点，但很难添加新的节点并将缓存的数据重新分发到节点。同样，当负载减少时，关闭一些节点并将其缓存的数据迁移到更少的节点变得更加困难；缩减集群规模有助于降低它们消耗的大量能源成本。Memcated节点数量的任何更改都会导致丢弃大多数缓存数据，导致分布式缓存必须从后端数据库缓慢重新预热的较长时间。这个早期概念探索性研究(AGER)项目研究了平稳扩展Memcach集群的技术，而不会造成任何短暂的性能降级。该项目还在内存节点和后端数据库之间引入了一个基于闪存的中间存储层，以帮助减少后端数据库的高延迟，并帮助平稳扩展Memcached集群。

Facebook and YouTube have massive databases containing many objects.Objects (e.g., pictures, videos) are read from these databases and presentedto users. These databases use slower and cheaper I/O devices. Repeateddatabase queries are also slow, so companies cache popular objects involatile memory (RAM). As RAM is about 1000x faster than I/O, queries topopular objects are serviced faster. A single computer cannot efficientlyuse too much RAM, so a distributed memory caching system called "Memcached"is utilized. Memcached creates a cluster of nodes, forming a key-valuedatabase in faster memory.Memcached's simple architecture led to its popularity, but is also itsAchilles' heel. When loads increase, more caching nodes are needed, but itis difficult to add new nodes and redistribute the cached data to morenodes. Similarly, when loads reduce, it is harder to shut down some nodesand migrate their cached data to fewer nodes; scaling down the cluster helpsreduce the large energy costs they consume. Any changes to the number ofMemcached nodes result in throwing out most cached data, leading to lengthyperiods where the distributed caches have to be slowly re-warmed up from thebackend database.This EArly-concept Grants for Exploratory Research (EAGER) project investigates techniques to scale Memcached clusterssmoothly, without any transient performance degradations. The project alsointroduces an intermediate Flash-based storage tier between the memory nodesand the backend database, to help reduce high latencies to the backenddatabase and help the smooth scaling of the Memcached cluster.