Remote Procedure Call (RPC)

Readings• Tanenbaum and van Steen: 2.2 and 2.3• Coulouris: Chapter 5• A good book: “Power Programming with

RPC” by John Bloomer• Another good book: “Practical UNIX

Programming: A Guide to Concurrency, Communication and Multithreading”

• Check out the links on the 402 web page

Outline• Difficulties in Socket Programming• Remote Procedure Calls (RPC) concepts• SUN RPC• Remote Object Invocation

Difficulties in Socket Programming• Data representation• Binding

Data Representation• Complex data structures must be converted

Sender must flatten complex data structures Receiver must reconstruct them

• Sender and receiver must agree on a common format for messages.

Exampletypedef struct { char Name[MAXNAMELENGTH]; float Salary; char JobCode[IDNUMLENGTH];} Employee

You may want to send this information to the server. Usingsend(s, (void *) &e, sizeof(e), 0)

where e is of type Employee is most likely not going to work.The sender and receiver must agree on a format for the message.

Data Representation• Sender and receiver must agree on type of

message. This can can be quite difficult.• Sender must convert data to send to the agreed

upon format. This often requires a “flattening” of the data structure representing this data.

• Receiver should parse the incoming message.

Data Representation• Useful functions

sprintf() – Used to convert data items to characters. An example is the following:

sprintf(msg, “%s %f %s”,name,salary,jobcode); sscanf() – Retrieves data items from a string. An

example is the following:sscanf(msg, “%s %f %s”,name,&salary,jobcode);

Data Representation• Code segment for marshalling the Employee structure

char *msg; char name[MAXNAMELENGTH];char jobcode[MAXNAMELENGTH];float salary;int msglength;Employee e;….salary = GetSalaryFromEmployee(e);GetJobCodeFromEmployee(e,jobcode);GetNameFromEmployee(e,name);

msg = (char *) malloc(sizeof(Employee));sprintf(msg,"%s %f %s",name,salary,jobcode);…..msglength = sizeof(Employee);send(s, (void *) msg, msglength));

Data Representation• Code segment for unmarshalling the Employee data sent

char *msg; char name[MAXNAMELENGTH];char jobcode[MAXNAMELENGTH];float salary;int msglength;Employee e;…msg = (char *) malloc(sizeof(Employee));…msglength = sizeof(Employee);recv(connectfd, (char *) msg, msglength,0);sscanf(msg, “%s %f %s”, name, &salary, jobcode);…

Other Representational Issues• Usually in a client-server model, the server provides

a set of services. Each service can be invoked by a procedure. For example, in an employee management system you would have services that include:

Insert an employee record for a new employee Get the salary of an employee Etc;

• The client must identify the service that it wants.• This is part of the message.

Other Representational Issues• Thus each service must have an identifier e.g.,

in the previous code segment examples we may have something like this:sprintf(msg,“%d %s %f %s",methodidentifier, name,salary,jobcode); sscanf (msg,“%d %s %f %s",&methodidentifier, name,&salary,jobcode);

Other Representational Issues• What if we have a list of Employees that we

want to send in a message. We do not know ahead of time how many employee records will be sent in a message. There are different ways to handle this. One way

is to send a message with the service identifier and the number of employee records being sent. You then send the number of employee records.

Other Representational Issues• What was just described works fine if the client and

server machines have similar machine types.• However, it is common that there are multiple

machine types. IBM mainframes use the EBCDIC character code, but

IBM PCs use ASCII. It would be rather difficult to pass a character parameter

from an IBM PC client to an IBM mainframe server using what has just been described.

Similar problems occur with integers (1’s complement vs two’s complement).

Other Representational Issues• Need a canonical form• For the UNIX OS there is XDR(eXternal Data

Representation). For Java RMI, there is Java Remote Method Protocol (JRMP).

Binding• Binding refers to determining the location and

identity (communication identifier) of the callee In UNIX, a communication identifier is a socket

address containing host’s IP address and port number.

Binding• Strategies for binding

Static binding (which binds the host address of a server into the client program at compilation time) is undesirable. The client and server programs are compiled separately and often

at different times. Services may be moved from one host to another.

Could pass host name and port by reading a file or through the command line. You do not need to recompile You still need to know the location of the server ahead of time.

Binding • Strategies for binding (cont)

Always run the binder on a “well-known” address (i.e., fixed host and port)

The operating system supplies the current address of the binder (e.g., via environment variable in UNIX). Users need to be informed whenever the binder is relocated Client and server programs need not be recompiled

Use a broadcast message to locate the binder The binder can be easily relocated Client/Server programs need not be recompiled

Binding• Dynamic binding service is desirable. Ideally,

a binding service would have the following characteristics: Allows servers to register their exporting services Allows servers to remote services Allows clients to lookup the named service We will come back to this later when looking at

specific examples.

Difficulties in Socket Programming

• Using sockets does not conceal communication which is important to achieve access transparency (defined as hiding differences in data representation and how a resource is accessed).

• Little was done until a paper by Birell and Nelson (1984). They suggested:

Allow programs to call procedures located on other machines.

Sounds simple, but the implementation is actually quite difficult.

This approach is known as Remote Procedure Call (RPC).

Introduction to RPC• An extension of conventional procedure call

(used for transfer of control and data within a single process)

• Allows a client program to call procedures in a different address space in the same or remote machine.

Introduction to RPC• Ideal for the client-server modeled

applications Higher level of abstraction for interprocess

communication• The goal is to make distributed programming

easier. Want transparent integration with the

programming language.

Conventional Procedure Call• Associated with each activation of a procedure

is storage for the variables declared in the procedure. This storage is called the activation record.

• Compilers (e.g., C) often use a stack to organize storage for local variables.

• Procedure activations are treated as a unit for storage allocation.

Conventional Procedure Call• Consider a C like call as follows:

count = read(fd,buf,nbytes)• To make this call, C pushes the parameters

onto the stack in order, last one first, as shown in figure (b) of the next page. Before the call, the stack is as seen in figure (a).

Conventional Procedure Call

(a) The stack before the call to read(b) The stack while the called procedure is active

Conventional Procedure Call• Call-by-value parameters are treated as an

initialized local variable.• A reference parameter in C is a pointer to a

variable. In the call to read, the second parameter is a

reference parameter. buf is actually changed (buf is assumed to be a character array).

• This distinction is important in remote procedure calls.

Software Support for RPCs• The idea behind RPC is to make a remote

procedure call look as much as possible like a local one.

• The calling procedure should not be aware that the called procedure is executing on a different machine.

Client/Server Interaction• Principle of RPC between a client and server program.

RPC System Components• Stub procedures

A stub is a communications interface that implements the RPC protocol and specifies how messages are constructed and exchanged

Responsible for packing and unpacking of arguments and results; this is referred to as marshalling.

Automatically generated by “stub generators” or “protocol compilers” (more later).

RPC System Components• Client stub

Marshalling: Packs the arguments with the procedure name or identifier into a message (this is instead of activation records)

Sends the message to the server and then awaits a reply message Unpacks the results and returns them to the client

• Server Stub Receives a request message Unmarshalling: Unpacks the arguments and calls appropriate

procedure When it returns, packs the result and sends a reply message back

to client.

send

receive

RPC System Components

client stub stubIPCruntime

IPCruntime Server

Interface

invoke

return

packargs

unpackresult

receive

send

unpackargs

packresult

invokework

return

Interface

Client and Server Stubs• In a synchronous RPC the client waits until

the sender returns a reply.• In an asynchronous RPC the client does not

wait.

Passing Pointers• What about pointers?

Very difficult. Pointers are meaningful only within the address space of

the process in which it is being used. Could eliminate all together, but this is not necessary. One strategy:

Assume a parameter that is a pointer to an array of characters and that this array’s size is known.

Copy the array into the message and send to server. Server changes the data, sends it back to the client and the client copies.

Can’t deal with complex data structures e.g. a graph

RPC Interface Definition Language and Compiler

• How are stubs created?• Servers provide one or more services to client

programs• Services are encapsulated and their clients

interact with them only via interfaces• An interface definition language (IDL) is used

to define these interfaces which are also known as service interfaces.

RPC Interface Definition Language and Compiler

• Each service interface includes names of the procedures plus the types of their input and output parameters.

• The interface compiler (or stub generator) automatically generates client and server stubs.

RPC Interface Definition Language and Compiler

Sun RPC• Also called ONC (Open Network Computing) RPC• Best known and most widely available RPC• Originally designed for client-server communication

in Sun Network File System (NFS)• Provides an RPC interface language (through XDR)

and compiler called rpcgen • Can use either UDP or TCP.

Interface Definition Language• An interface contains a program number, a

version number, procedure definitions and required type definitions.

• A procedure definition specifies a procedure signature and a procedure number.

Interface Definition Language• Only a single input and output parameter is

allowed.• rpcgen compiles interface definitions into

stubs, header files and main server source code.

Example• We will illustrate SUN RPC by converting a simple local service

for generating pseudorandom numbers into a remote service.• This is based on the drand48 and srand48 functions.• Prior to invoking drand48, a program must initialize a starting

value by calling the srand48 function with a long parameter value called the seed.

• The seed determines the starting position in a predetermined sequence of pseudorandom numbers.

• After initializing the generator by invoking srand48, call drand48 to return successive values in a sequence of pseudorandom values that are uniformly distributed in the interval [0,1).

ExampleThere is one file that we will call pseudorandom.c #include "rand.h"void initialize_random(long seed){ srand48(seed);

}double get_next_random(void){ return drand48();}

ExampleA program that uses these functions is in main.c and a

segment looks like this:……. myseed = (long)atoi(argv[1]); iters = atoi(argv[2]); initialize_random(myseed); for (i = 0; i < iters; i++) printf("%d: %f\n",i, get_next_random()); exit(0);

Please note that the seed value and the number of iterations are command line arguments.

Example• Now let us see how to make initialize_random

and get_next_random remote functions.• We first provide a specification (XDR) file for

the remote service.• The XDR file has a .x extension.

Example

rand.x is the following:

program RAND_PROG { version RAND_VERS{ void INITIALIZE_RANDOM(long) = 1; void GET_NEXT_RANDOM(void) = 2; } = 1;} = 0x31111111

Program Identifiers• The Sun convention for program numbers if

the following: 0x00000000 - 0x1fffffff (Sun) 0x20000000 - 0x3fffffff (User) 0x40000000 - 0x5fffffff (transient) 0x60000000 - 0xffffffff (reserved)

Procedure Identifiers &Program Version Numbers

• Procedure identifiers usually start at 1 and are numbered sequentially

• Version numbers typically start at 1 and are numbered sequentially.

• The function names are the same as those before except that are all in uppercase.

• The functions are numbered so that the initialize_random function is service number 1 and get_next_random is service number 2 in the server.

Example• Now to use rpcgen as follows:

rpcgen –C –a rand.x

• The –C option indicates ANSI C and the –a option tells rpcgen to generate all of the supporting files.

Example• Files generated include:

makefile.rand: This file is the makefile for compiling all of the client and server code.

rand_clnt.c: This file contains the client stub, which is usually not motified.

rand_svc.c: This file contains the server stub, which is usually not modified.

rand.h: This header file contains all of the XDR types from the specification.

rand_client.c: This file contains a skeleton client main program with dummy calls to the remote service. You insert code to set up the argument values for the remote service before the dummy call.

Example• Files generated include (continued):

rand_server.c: This file contains the stubs for the remote services. Insert the code for the local version of the services into these stubs.

rand_xdr.c: If this file is generated, it contains XDR filters (routines) needed by the client and server stubs.

Example• You can now modify the rand_client.c file to

contain the client code.• You then modify the rand_server.c file to

contain the functions to be called remotely.

Example• This is the rand_client.c generated by rpcgen.#include "rand.h"#include <stdio.h>#include <stdlib.h> /* getenv, exit */

voidrand_prog_1(char *host){

CLIENT *clnt;void *result_1;long initialize_random_1_arg;double *result_2;char * get_next_random_1_arg;

Example#ifndef DEBUG

clnt = clnt_create(host, RAND_PROG, RAND_VERS, "netpath");if (clnt == (CLIENT *) NULL) {clnt_pcreateerror(host);exit(1);}

#endif /* DEBUG */

result_1 = initialize_random_1(&initialize_random_1_arg,clnt); if (result_1 == (void *) NULL) {clnt_perror(clnt, "call failed");}result_2 = get_next_random_1((void *)&get_next_random_1_arg, clnt);if (result_2 == (double *) NULL) {clnt_perror(clnt, "call failed");}

#ifndef DEBUGclnt_destroy(clnt);

#endif /* DEBUG */}

Creates a handle forthe remote service

Has version number appended to function

name

clnt pointer is deallocated

Examplemain(int argc, char *argv[]){

char *host;

if (argc < 2) {printf("usage: %s server_host\n", argv[0]);exit(1);

}host = argv[1];rand_prog_1(host);

}

Example• In the rand_prog_1, we take note of the following:

The clnt_create call generates a handle for the remote service. If it fails, a NULL pointer is returned. Returns the clnt pointer which is the handle (communication information) for the remote service.

The RAND_PROG and RAND_VERS parameters are the program and version names specified in rand.x

The “netpath” parameter indicates that the program should look for an available network transport mechanism as specified by the NETPATH environment variable.

The converted remote calls to initialize_random and get_next_random have the version number appended to the function names i.e., initialize_random is called as initialize_random_1.

In the remote calls, the parameters and return values are designated by the pointers.

The clnt pointer is deallocated by clnt_destroy.

ExampleThe following is a revised version of the main function

in rand_client.c#include "rand.h"#include <stdio.h>#include <stdlib.h> /* getenv, exit */

main(int argc, char *argv[]){ int iters, i; long myseed; CLIENT *clnt; void *result_1; double *result_2; char *arg;

Exampleif (argc != 4) { fprintf(stderr, "Usage: %s host seed iterations\n”, argv[0]); exit(1);}clnt =

clnt_create(argv[1],RAND_PROG,RAND_VERS,"netpath");

if (clnt == (CLIENT *) NULL) { clnt_pcreaterror(argv[1]); exit(1);}

Create handle;Hostname is passed as the first commandline argument

Examplemyseed = (long)atoi(argv[2]);iters = atoi(argv[3]);

result_1 == initialize_random_1(&myseed,clnt);if (result_1 == (void *) NULL) { clnt_perror(clnt, "call failed");}

for (i = 0; i < iters; i++) {

result_2 = get_next_random_1((void *)&arg, clnt); if (result_2 == (double *) NULL) { clnt_perror(clnt, "call failed"); } else

printf("5d: %f\n", i, *result_2);}

clnt_destroy(clnt);exit(0);

Convert local to remote calls

clnt pointer id deallocatec

Example• The revised version of rand_client.c is a combination of the

main program of the generated rand_client.c and the main program used in the local version.

Start with the original program for local service and insert the call to create_client near the beginning and clnt_destroy at the end.

The host name is passed as the first command line argument. The main program calls the remote functions directly, so there is no

need for rand_prog_1. The next change is to convert the calls to initialize_random and

get_next_random from local to remote calls. The remote functions pass their parameters by pointer and return a

pointer to the return value. The clnt handle is passed as an additional parameter in the class.

ExampleThe skeleton code for rand_server.c is presented here.#include "rand.h"#include <stdio.h>#include <stdlib.h> /* getenv, exit */#include <signal.h>void *initialize_random_1_svc(long *argp, struct svc_req *rqstp){ static char * result;

/* * insert server code here */

return((void *) &result);}

Exampledouble * get_next_random_1_svc(void *argp, struct svc_req *rqstp){ static double result;

/* * insert server code here */

return (&result);}

ExampleServer code that gets inserted for

initialize_random_1 is the following:

srand48(*argp);result = (void *) NULL;

Server code that gets inserted for get_next_random_1 is the following:

result = drand48();

Example• You can now create two executables rand_client and

rand_server for the client and server respectively using the following:make –f makefile.rand

• The following registers the server:rand_serverThis causes the service to be registered on the current host

and ready to receive remote requests.

Exactly with “what” does the server register with?

Dynamic Port Mapping• Servers typically do not use well known

protocol ports!• Clients know the program identifier (and host

IP address).• SUN RPC includes support for looking up the

port number of a remote program.

Port Lookup Service• A port lookup service runs on each host that

contains RPC servers.• RPC servers register themselves with this

service: "I'm program 17 and I'm looking for requests on

port 1736"

The portmapper• Each system which will support RPC servers runs a

port mapper server that provides a central registry for RPC services.

• Servers tell the portmapper what services they offer when they register.

• Basically, an RPC server registers itself with this service and pass information along the following lines:

“I am program 17, version 5 and I’m looking for requests on port 1736”

The port number is often randomly chosen.

More on the portmapper• Clients ask a remote port mapper for the port

number corresponding to a program identifier and version.

• The portmapper is available on a well-known port (111).

RPC Semantics• Major difference between an RPC and a

conventional procedure call is the number of ways the RPC may fail.

• During an RPC, problems may occur: Request message may be lost Reply message may be lost Server and/or client may crash

In the last two cases the procedure may have been called.

RPC SemanticsSome strategies for different RPC message delivery

guarantees Retry request message -- retransmit the request message

until either a reply is received or the server is assumed to have failed.

Duplicate filtering -- Filtering duplicate requests at the server when retransmissions are used.

Retransmission of replies -- Keep a history of reply messages to enable lost replies to be retransmitted without re-executing the server operations

RPC Semantics• RPC mechanisms usually include timeouts to prevent

clients waiting indefinitely for reply messages• RPC call semantics

Semantics achieved depends on how failures are dealt with “maybe” call semantics “at-least-once” call semantics “at-most-once” call semantics cannot achieve “exactly-once” call semantics

RPC Semantics• “maybe” call semantics

No retransmission of request messages Not certain whether the procedure has been

executed No fault-tolerance measures Generally not acceptable

RPC Semantics• “at-least-once” call semantics

Request messagesare repeatedly sent after timeouts until it either gets a reply message or some maximum number of retries have been made.

No duplicate request message filtering The remote procedure is called at least once if server not down The client does not know how many times the remote

procedure has been called Unless called procedure is “idempotent” (i.e. repeatable), this

could produce undesirable results (e.g., money transfer). A function such as deposit(DavesAccount,$100)is

not idempotent.

RPC Semantics• “at-most-once” call semantics

Retransmission of request messages Duplicate request message filtering If server does not crash and client receives result of call, then

the procedure has been called exactly once, otherwise an exception is reported and the procedure will have been called either once or not at all.

This works for both idempotent and non-idempotent procedures.

Complex support needed. Request identifiers ensure only retransmissions filtered and

not new sendings of the same request.

SUN RPC Semantics• The use of UDP provides “maybe” semantics.• The use of TCP provides “at most once”

semantics.

DCE RPC• Distributed Computing Environment (DCE) is

from OSF.• DCE is a middleware system that is designed

to execute as a layer of abstraction between existing (network) operating systems and distributed applications.

• Initially designed for UNIX it has now been ported to all major operating systems including VMS and Windows NT.

DCE RPC• Services provided by DCE

Distributed file service – provides a transparent way of accessing any file in the system in the same way.

Directory service – is used to keep track of the location of all resources in the system.

Security service – allows resources of all kinds to be protected.

Distributed time service – attempts to keep clocks on the different machines globally synchronized.

Goals of DCE RPC• The RPC systems makes it possible for a

client to access a remote service by simply calling a local procedure. This is similar to SUN RPC.

• Semantic options: At-most-once operation Idempotent

Binding a Client to a Server in DCE• Client-to-server binding in DCE.

2-15