CSE480 Lecture Notes
February 1, 1996

Remote Procedure Call

picture

Goal: make RPC transparent (i.e., behave just like proc call)

Potential Problems
- address space
- parameters
- failtures

Basic RTC
- normal procedure call
- count = read(fd,buf,nbytes)   Fig 2-17

Parameters
- call-by-value
- call-by-reference
- call-by-copy/restore  (e.g., Ada's in/out)


RPC implementation uses client stubs in the place
acutal procedure code (Fig 2-18)

10 steps at client

Parameter Passing

- parameter marschalling -- packing paramenters into a msg
- n = sum(4,7)   Fig 2-19

- how to handle different data represenations
  EBCDIC, ASCII, Universial ASCII
  floating point numbers
  little endian (bytes from right to left, 486)
  big endian (bytes from left to right, SPARC)
  e.g., Fig 2-20    5  "JILL"
  two approaches
  i) direct conversion (n * (n-1) translators)
 ii) cannonical represenation; e.g., XDR

- how to handle pointers
  address at 1000
  copy and use internal offset
  tree structure
  recursive structure

Dynamic binding

- how to locate a server
- how to register a server
- (name, version, handle, unique-id)

RPC Semantics in the Presence of Failures

- goal: RPC behave like local procedure call

1) cannot locate the server

- server down
- version mismatch
- raise an exception

2) request msg is lost

- use timer and resend request msg

3) reply msg is lost

- use timer and resent another request
- server need to tell whether the request is duplicate
  unless the request is idempotent
  i) make all request idempotent
     e.g., read (fd, buf, n)
           read (fd, buf, pos, n)
           deposit (money) -- not possible to make it idempotent
 ii) assign request nos and keep track

4) server crashes after receiving a request

- unique no not enough since cannot know whether or not RPC done
- Fig 2-24, (b) and (c) cannot be distinguished
  raise an except if RCP not complete
  retransmit the reply/request

- Possible Semantics
  0) do nothing
  i) at least once
 ii) at most once
iii) exactly once (not possible to implement)

5) client crashes after sending a request

- make request and then crash
- orphan -- computation active not no parent
- potential problems -- waste cycles, etc.
- how to deal with orphans
  i) extermination
 ii) reincarnation
iii) expiration

- killing orphans may not be desirable (e.g., hold locks)