一、Big Storage

• 为big storage system设计一个接口
• how to design innards of the storage system

1. WHY HARD

• performance -> sharding(分片)
• faults -> tolerance
• tolerance -> replication
• replication -> inconsistency
• consistency -> low performance

2. STRONG CONSISTENCY

image.png

单个服务器的容错能力很差

3. BAD REPL DESIGN

image.png

4.1 GFS

• Big, Fast
• Global
• sharding
• Automatic recovery

4.2 Single data center

• Intern use
• Big sequential access
• single master

4.3 master data细节讲解

image.png

• filename -> array of chunk handles(NV非易失存储)
• handle ->
  1.list of chunk server(V)
  2.version number for each chunk(NV)
  3.primary(V)
  4.lease expiration(V)
• LOG, CHECKPOINT -> DISK

4.4 master opration细节讲解

• READ
  1.name,offset -> master
  2.M chunk handle, serverlist
  cached
  3.c(line) -> cs(chunk server)
• WRITES

1. NO PRIMATY? on master
   find up to date repl
   pick p , s
   increment v#
   tells p , s v#
   M writers v# k, dish