Книга: Distributed operating systems

6.4.8. Synchronization

6.4.8. Synchronization

In a DSM system, as in a multiprocessor, processes often need to synchronize their actions. A common example is mutual exclusion, in which only one process at a time may execute a certain part of the code. In a multiprocessor, the TEST-AND-SET-LOCK (TSL) instruction is often used to implement mutual exclusion. In normal use, a variable is set to 0 when no process is in the critical section and to 1 when one process is. The TSL instruction reads out the variable and sets it to 1 in a single, atomic operation. If the value read is 1, the process just keeps repeating the TSL instruction until the process in the critical region has exited and set the variable to 0.

In a DSM system, this code is still correct, but is a potential performance disaster. If one process, A, is inside the critical region and another process, B, (on a different machine) wants to enter it, B will sit in a tight loop testing the variable, waiting for it to go to zero. The page containing the variable will remain on ZTs machine. When A exits the critical region and tries to write 0 to the variable, it will get a page fault and pull in the page containing the variable. Immediately thereafter, B will also get a page fault, pulling the page back. This performance is acceptable.

The problem occurs when several other processes are also trying to enter the critical region. Remember that the TSL instruction modifies memory (by writing a 1 to the synchronization variable) every time it is executed. Thus every time one process executes a TSL instruction, it must fetch the entire page containing the synchronization variable from whoever has it. With multiple processes each issuing a TSL instruction every few hundred nanoseconds, the network traffic can become intolerable.

For this reason, an additional mechanism is often needed for synchronization. One possibility is a synchronization manager (or managers) that accept messages asking to enter and leave critical regions, lock and unlock variables, and so on, sending back replies when the work is done. When a region cannot be entered or a variable cannot be locked, no reply is sent back immediately, causing the sender to block. When the region becomes available or the variable can be locked, a message is sent back. In this way, synchronization can be done with a minimum of network traffic, but at the expense of centralizing control per lock.

Оглавление книги