XIV（4）--Flash caching

1，Flash Cache简介

Flash cache 自己是Facebook的一个开源项目，（准确的说是一个Linux的模块），能够动态加载。Flashcache经过在文件系统（VFS）和设备驱动之间新增了一次缓存层，来实现对热门的缓存。Flashcache是另外一种缓存，通常用SSD做为介质的缓存（通常的缓存用的是内存），经过将传统硬盘上的热门数据缓存到SSD上，而后利用SSD优秀的读性能，来加速系统。这个方法较以内存缓存，没有内存快，可是空间能够比内存大不少。以下图：

  如今不少硬件厂商也会在存储设备中增长这个功能来，例如IBM XIV的Flash cache和 EMC的VFCache

2， XIV Flash Cache

从11.1.0开始，XIV Gen3开始支持可选的Flash caching feature，极大的提升小数据块，随机读的性能，适用于Data patterns一直改变的环境。

XIV flash caching is implemented as an extension of the primary cache layer.   每一个模块400GB Cache，那一个系统就是6TB(400GB*15)，从11.4开始，每一个Module能支持800GB。Flash cache 只用于读操做，当再也不须要Cache中的数据时，直接drop掉就好了。

XIV flash caching overview

  Flash caching算法是嵌入在XIV Firmware中的，能自动适应相应的IO类型，对用户透明，不须要管理员手动的作performance turning。    

XIV系统中有2种类型的Cache：main 和 extended  
--The main cache handles host write I/Os and then destages them directly to the disk drive. //Main cache主要处理主机写IO而后直接Flush到Disk    
--The extended cache handles the caching of random read miss operations less than 64 KB. //Extended cache主要处理小于64KB的随机读操做    

Sequential read prefetches (larger than 64 KB) are handled in main dynamic random access memory (DRAM) cache. //大于64KB的顺序读操做是由main cache处理的

3，Flash Cache learning

A flash cache map is built as read misses occur in the DRAM cache. The process, known as flash cache learning

关于XIV中Main cache和Extended Cache（Flash Cache）是如何扮演各自角色的，请看下图：

 

XIV flash cache learning

 

The cache node immediately checks the extended cache for the requested I/O. If the requested I/O exists in the extended flash cache, it is served to the host through the main cache. The I/O operation is now complete and is recorded as a flash cache read hit.

If the operation results in a true read miss (not in the DRAM cache and not in extended flash cache), the request is forwarded in an unmodified state to the disk drive (SAS layer). The I/O is retrieved from the disk drive and served to the host through the main cache. From a host perspective, the I/O operation is now complete and is recorded as a read miss. The related pages are copied into reserved buffers in the main cache.

 

Important: Any read larger than 64 KB bypasses the extended flash cache.

当Buffer中的data达到512KB时，会顺序地写到Flash cache中。这种方式延长了Flash cache中的寿命。

Note: XIV在系统重启和Firmware升级中能保存Flash Cache中的数据    

XIV Storage System software Version 11.2 introduced improved flash caching algorithms,providing a performance boost of up to 4.5 times over systems without flash cache for random database-type workloads. This boost is accomplished by storing and computing all flash cache-related data integrity checking tasks in DRAM rather than on the flash cache.

4，Approaches for Using SSDs in a Storage System

Approach of the tier. With policies that moved the data … or As a caching layer. This is the approach of XIV.

5，SSD failure

• No redistribution in case of failure/phase-out

• SSD Failure:

  -Reinitialize the metadata of the SSD (so it is not used)
  -The degraded module continue to server reads from its DRAM cache and large sequential reads from its disks
  -Small read misses are redirected to the secondary

• SSD Phase-out and not failed

- Its data is invalidated (on writes)
=> if phased-in not all data is lost

During a rebuild, following a SAS disk failure in a module, the data on SSD in that module is

not invalidated. Rather, it is gradually updated to contain the new data blocks (the same way

the DRAM does).

------------------------------------------------------------------------------

后记：关于EMC VFCache和IBM Flash Cache，感兴趣的能够看看比特网的下图对比。

参考文章：http://storage.chinabyte.com/223/12261223.shtml