Remote Procedure Call(RPC)

Hiding the Network

Client-Server Model

• server exports an interface, client invokes it

• server acts like an abstract data type

• 服务端提供接口，client 端负责调用

• 服务端视为一种实现抽象数据类型的实体。

Interfaces

• wide range of possible interfaces 多种可能的接口

• any function prototype is potentially a candidate

Historical Roots of RPC

• potential parallelism in a distributed system 分布式系统潜在的并行性

• client and server on different machines

• no reason for client to block

• A sends invocation message to B

• A does other work

• A polls for B’s reply or is interrupted

• A receives B’s reply

• often little for A to do while awaiting B’s Reply 在等待B的答复时通常几乎没有什么可做的

• explicit matching of replies complicates code 回复的显式匹配使代码复杂化

• parallelism better exploited via threads 通过线程更好地利用并行性

• message passing offers greater generality 消息传递提供了更大的通用性

• but that generality not useful in common cases 但这种通用性在一般情况下没有用

• well suited for the client server model 非常适合client server模型

• sacrifices some generality 牺牲一些通用性

• can be made very efficient 可以变得非常高效

• simple to use

• emulates familiar procedure call paradigm 模拟熟悉的过程调用范例

Characteristics of RPC

Two aspects

• control flow 控制流

• invocation syntax 调用语法

 

 

 

 

 

 

 

 

Control flow

• caller makes request and blocks 调用者发起请求并阻塞

• callee services request and replies 被调用方请求和回复

• caller resumers 调用者回复

There is thus synchronous transfer of controller

在进行远程调用中即调用者的状态在调用期间（服务器执行时）被冻结，然后控制权返回，之后本地状态可能会再次发生变化。这是理想状况，实际上，要保证本地状态不发生变化是非常非常困难的，而分布式系统中最棘手的错误，往往就发生在这一假设被违背的时候。在你不知情的情况下，一部分本地状态悄然发生改变，而你对此毫无察觉。因此，当回复返回时，实际情况已与预期不符合。

迅速和零延迟是两个不同的概念。本地环境中延迟虽非零，但其值已及其接近零。

远程调用需要时间，如果调用一次消耗 60ms 的时间，在这 60毫秒内，其他进程或线程都可能会改变本地状态。因此在分布式系统中，要考虑延迟或传输时间。

Invocation syntax

• Local syntax : y = foo( x1, x2, ……)

• Remote syntax : y = foo(cid, x1, x2,……)

• cid is called connection handle

调用语法通常由编程语言决定过程调用的形式。

 

 

Limitations of RPC

Client and Server do not share address space

• pointers don’t mean the same thing at the two ends

• many implications
can’t share global data
can’t use call-by-reference
only call-by-value-result
large data structures can be expensive
procedure parameters don’t work easily
up-level addressing doesn’t work easily

Delayed binding in RPC

• conceptually similar to dynamic linking 概念上类似于动态链接

• hence runtime linking errors possible

• multiple instances of callee
many servers may export same of callee
client may wish to talk to more than one server 客户端可能希望与多个服务器通信
no counterpart in local procedure call 本地过程调用中没有对应项

Failure independence of clients and Services

• in local case ,client and server live or die together

• in remote case ,client see new failure modes
• network failure 网络故障
• server machine crash （服务 机器崩溃）
• server process crash （服务 进程崩溃）

• nontransitive communication
• A can talk to B , and B to C, but not c to A

• failure handling code has to be more thorough(and more complex)