Andrew Tanenbaum - Distributed operating systems

Normally, a pointer is followed (dereferenced) by putting it in a register and indirecting through the register. When this special technique is used, a pointer is dereferenced by sending a message back to the client stub asking it to fetch and send the item being pointed to (reads) or store a value at the address pointed to (writes). While this method works, it is often highly inefficient. Imagine having the file server store the bytes in the buffer by sending back each one in a separate message. Still, it is better than nothing, and some systems use it.

An issue that we have glossed over so far is how the client locates the server. One method is just to hardwire the network address of the server into the client. The trouble with this approach is that it is extremely inflexible. If the server moves or if the server is replicated or if the interface changes, numerous programs will have to be found and recompiled. To avoid all these problems, some distributed systems use what is called dynamic bindingto match up clients and servers. in this section we will describe the ideas behind dynamic binding.

The starting point for dynamic binding is the server's formal specification. As an example, consider the server of Fig. 2-9(a), specified in Fig. 2-22. The specification tells the name of the server (file_server), the version number (3.1), and a list of procedures provided by the server (read, write, create, and delete).

#include

specification of file_server, version 3.1:

long read(in char name[MAX_PATH], out char buf[BUF_SIZE], in long bytes, in long position);

long write(in char name[MAX_PATH], in char buf[BUF_SIZE], in long bytes, in long position);

int create(in char[MAX_PATH], in int mode);

int delete(in char[MAX_PATH]);

end;

Fig. 2-22.A specification of the stateless server of Fig. 2-9.

For each procedure, the types of the parameters are given. Each parameter is specified as being an in parameter, an out parameter, or an in out parameter. The direction is relative to the server. An in parameter, such as the file name, name, is sent from the client to the server. This one is used to tell the server which file to read from, write to, create, or delete. Similarly, bytes tells the server how many bytes to transfer and position tells where in the file to begin reading or writing. An out parameter such as buf in read, is sent from the server to the client. Buf is the place where the file server puts the data that the client has requested. An in out parameter, of which there are none in this example, would be sent from the client to the server, modified there, and then sent back to the client (copy/restore). Copy/restore is typically used for pointer parameters in cases where the server both reads and modifies the data structure being pointed to. The directions are crucial, so the client stub knows which parameters to send to the server, and the server stub knows which ones to send back.

As we pointed out earlier, this particular example is a stateless server. For a UNIX-like server, one would have additional procedures open and close, and different parameters for read and write. The concept of RPC itself is neutral, permitting the system designers to build any kind of servers they desire.

The primary use of the formal specification of Fig. 2-22 is as input to the stub generator, which produces both the client stub and the server stub. Both are then put into the appropriate libraries. When a user (client) program calls any of the procedures defined by this specification, the corresponding client stub procedure is linked into its binary. Similarly, when the server is compiled, the server stubs are linked with it too.

When the server begins executing, the call to initialize outside the main loop [see Fig. 2-9(a)] exportsthe server interface. What this means is that the server sends a message to a program called a binder, to make its existence known. This process is referred to as registeringthe server. To register, the server gives the binder its name, its version number, a unique identifier, typically 32 bits long, and a handleused to locate it. The handle is system dependent, and might be an Ethernet address, an IP address, an X.500 address, a sparse process identifier, or something else. In addition, other information, for example, concerning authentication, might also be supplied. A server can also deregister with the binder when it is no longer prepared to offer service. The binder interface is shown in Fig. 2-23.

Fig. 2-23.The binder interface.

Given this background, now consider how the client locates the server. When the client calls one of the remote procedures for the first time, say, read, the client stub sees that it is not yet bound to a server, so it sends a message to the binder asking to importversion 3.1 of of the file_server interface. The binder checks to see if one or more servers have already exported an interface with this name and version number. If no currently running server is willing to support this interface, the read call fails. By including the version number in the matching process, the binder can ensure that clients using obsolete interfaces will fail to locate a server rather than locate one and get unpredictable results due to incorrect parameters.

On the other hand, if a suitable server exists, the binder gives its handle and unique identifier to the client stub. The client stub uses the handle as the address to send the request message to. The message contains the parameters and the unique identifier, which the server's kernel uses to direct the incoming message to the correct server in the event that several servers are running on that machine.

This method of exporting and importing interfaces is highly flexible. For example, it can handle multiple servers that support the same interface. The binder can spread the clients randomly over the servers to even the load if it wants to. It can also poll the servers periodically, automatically deregistering any server that fails to respond, to achieve a degree of fault tolerance. Furthermore, it can also assist in authentication. A server could specify, for example, that it only wished to be used by a specific list of users, in which case the binder would refuse to tell users not on the list about it. The binder can also verify that both client and server are using the same version of the interface.

However, this form of dynamic binding also has its disadvantages. The extra overhead of exporting and importing interfaces costs time. Since many client processes are short lived and each process has to start all over again, the effect may be significant. Also, in a large distributed system, the binder may become a bottleneck, so multiple binders are needed. Consequently, whenever an interface is registered or deregistered, a substantial number of messages will be needed to keep all the binders synchronized and up to date, creating even more overhead.

The goal of RPC is to hide communication by making remote procedure calls look just like local ones. With a few exceptions, such as the inability to handle global variables and the subtle differences introduced by using copy/restore for pointer parameters instead of call-by-reference, so far we have come fairly close. Indeed, as long as both client and server are functioning perfectly, RPC does its job remarkably well. The problem comes in when errors occur. It is then that the differences between local and remote calls are not always easy to mask. In this section we will examine some of the possible errors and what can be done about them.