Multithreaded Programming Guide

Chapter 1 Covering Multithreading Basics

The word multithreading can be translated as multiple threads of control or multiple flows of control. While a traditional UNIX process always has contained and still does contain a single thread of control, multithreading (MT) separates a process into many execution threads, each of which runs independently.

Multithreading your code can

This chapter explains some multithreading terms, benefits, and concepts. If you are ready to start using multithreading, skip to Chapter 2, Basic Threads Programming.

Defining Multithreading Terms

Table 1-1 introduces some of the terms used in this book.

Table 1-1 Multithreading Terms




The UNIX environment (such as file descriptors, user ID, and so on) created with the fork(2) system call, which is set up to run a program.


A sequence of instructions executed within the context of a process. 

pthreads (POSIX threads) 

A POSIX 1003.1c compliant threads interface. 

Solaris threads 

A Sun MicrosystemsTM threads interface that is not POSIX compliant. A predecessor of pthreads.


Restricting access to a single thread. 


Allowing access to two or more threads. 

User- or Application-level threads

Threads managed by the threads library routines in user (as opposed to kernel) space. 

Lightweight processes

Threads in the kernel that execute kernel code and system calls (also called LWPs). 

Bound threads

Threads that are permanently bound to LWPs. 

Unbound threads

A default Solaris thread that context switches very quickly without kernel support. 

Attribute object 

Contains opaque data types and related manipulation functions used to standardize some of the configurable aspects of POSIX threads, mutual exclusion locks (mutexes), and condition variables. 

Mutual exclusion locks 

Functions that lock and unlock access to shared data. 

Condition variables 

Functions that block threads until a change of state. 

Counting semaphore

A memory-based synchronization mechanism. 


A condition that arises when at least two threads are executing simultaneously.


A condition that exists when at least two threads are making progress. A more generalized form of parallelism that can include time-slicing as a form of virtual parallelism.

Meeting Multithreading Standards

The concept of multithreaded programming goes back to at least the 1960s. Its development on UNIX systems began in the mid-1980s. While there is agreement about what multithreading is and the features necessary to support it, the interfaces used to implement multithreading have varied greatly.

For several years a group called POSIX (Portable Operating System Interface) 1003.4a has been working on standards for multithreaded programming. The standard has now been ratified. This Multithreaded Programming Guide is based on the POSIX standards: P1003.1b final draft 14 (realtime), and P1003.1c final draft 10 (multithreading).

This guide covers both POSIX threads (also called pthreads) and Solaris threads. Solaris threads were available in the Solaris 2.4 release, and are not functionally different from POSIX threads. However, because POSIX threads are more portable than Solaris threads, this guide covers multithreading from the POSIX perspective. Subjects specific to Solaris threads only are covered in the Chapter 8, Programming With Solaris Threads.

Benefiting From Multithreading

Improving Application Responsiveness

Any program in which many activities are not dependent upon each other can be redesigned so that each activity is defined as a thread. For example, the user of a multithreaded GUI does not have to wait for one activity to complete before starting another.

Using Multiprocessors Efficiently

Typically, applications that express concurrency requirements with threads need not take into account the number of available processors. The performance of the application improves transparently with additional processors.

Numerical algorithms and applications with a high degree of parallelism, such as matrix multiplications, can run much faster when implemented with threads on a multiprocessor.

Improving Program Structure

Many programs are more efficiently structured as multiple independent or semi-independent units of execution instead of as a single, monolithic thread. Multithreaded programs can be more adaptive to variations in user demands than single-threaded programs.

Using Fewer System Resources

Programs that use two or more processes that access common data through shared memory are applying more than one thread of control.

However, each process has a full address space and operating environment state. The cost of creating and maintaining this large amount of state information makes each process much more expensive than a thread in both time and space.

In addition, the inherent separation between processes can require a major effort by the programmer to communicate between the threads in different processes, or to synchronize their actions.

Combining Threads and RPC

By combining threads and a remote procedure call (RPC) package, you can exploit nonshared-memory multiprocessors (such as a collection of workstations). This combination distributes your application relatively easily and treats the collection of workstations as a multiprocessor.

For example, one thread might create child threads. Each of these children could then place a remote procedure call, invoking a procedure on another workstation. Although the original thread has merely created threads that are now running in parallel, this parallelism involves other computers.

Understanding Basic Multithreading Concepts

Concurrency and Parallelism

In a multithreaded process on a single processor, the processor can switch execution resources between threads, resulting in concurrent execution.

In the same multithreaded process in a shared-memory multiprocessor environment, each thread in the process can run on a separate processor at the same time, resulting in parallel execution.

When the process has fewer or as many threads as there are processors, the threads support system in conjunction with the operating environment ensure that each thread runs on a different processor.

For example, in a matrix multiplication that has the same number of threads and processors, each thread (and each processor) computes a row of the result.

Looking at Multithreading Structure

Traditional UNIX already supports the concept of threads--each process contains a single thread, so programming with multiple processes is programming with multiple threads. But a process is also an address space, and creating a process involves creating a new address space.

Creating a thread is much less expensive when compared to creating a new process, because the newly created thread uses the current process address space. The time it takes to switch between threads is much less than the time it takes to switch between processes, partly because switching between threads does not involve switching between address spaces.

Communicating between the threads of one process is simple because the threads share everything--address space, in particular. So, data produced by one thread is immediately available to all the other threads.

The interface to multithreading support is through a subroutine library, libpthread for POSIX threads, and libthread for Solaris threads. Multithreading provides flexibility by decoupling kernel-level and user-level resources.

User-Level Threads

Threads are the primary programming interface in multithreaded programming. User-level threads [User-level threads are named to distinguish them from kernel-level threads, which are the concern of systems programmers, only. Because this book is for application programmers, kernel-level threads are not discussed.] are handled in user space and avoid kernel context switching penalties. An application can have hundreds of threads and still not consume many kernel resources. How many kernel resources the application uses is largely determined by the application.

Threads are visible only from within the process, where they share all process resources like address space, open files, and so on. The following state is unique to each thread.

Because threads share the process instructions and most of the process data, a change in shared data by one thread can be seen by the other threads in the process. When a thread needs to interact with other threads in the same process, it can do so without involving the operating environment.

By default, threads are very lightweight. But, to get more control over a thread (for instance, to control scheduling policy more), the application can bind the thread. When an application binds threads to execution resources, the threads become kernel resources (see "System Scope (Bound Threads)" for more information).

To summarize, user-level threads are:

Lightweight Processes

The threads library uses underlying threads of control called lightweight processes that are supported by the kernel. You can think of an LWP as a virtual CPU that executes code or system calls.

You usually do not need to concern yourself with LWPs to program with threads. The information here about LWPs is provided as background, so you can understand the differences in scheduling scope, described on "Process Scope (Unbound Threads)".

Note -

The LWPs in the Solaris 2, Solaris 7, and Solaris 8 operating environments are not the same as the LWPs in the SunOSTM 4.0 LWP library, which are not supported in the Solaris 2, Solaris 7, and Solaris 8 operating environments.

Much as the stdio library routines such as fopen() and fread() use the open() and read() functions, the threads interface uses the LWP interface, and for many of the same reasons.

Lightweight processes (LWPs) bridge the user level and the kernel level. Each process contains one or more LWP, each of which runs one or more user threads. (See Figure 1-1.) The creation of a thread usually involves just the creation of some user context, but not the creation of an LWP.

Figure 1-1 User-level Threads and Lightweight Processes


Each LWP is a kernel resource in a kernel pool, and is allocated (attached) and de-allocated (detached) to a thread on a per thread basis. This happens as threads are scheduled or created and destroyed.


POSIX specifies three scheduling policies: first-in-first-out (SCHED_FIFO), round-robin (SCHED_RR), and custom (SCHED_OTHER). SCHED_FIFO is a queue-based scheduler with different queues for each priority level. SCHED_RR is like FIFO except that each thread has an execution time quota.

Both SCHED_FIFO and SCHED_RR are POSIX Realtime extensions. SCHED_OTHER is the default scheduling policy.

See "LWPs and Scheduling Classes" for information about the SCHED_OTHER policy, and about emulating some properties of the POSIX SCHED_FIFO and SCHED_RR policies.

Two scheduling scopes are available: process scope for unbound threads and system scope for bound threads. Threads with differing scope states can coexist on the same system and even in the same process. In general, the scope sets the range in which the threads scheduling policy is in effect.

Process Scope (Unbound Threads)

Unbound threads are created PTHREAD_SCOPE_PROCESS. These threads are scheduled in user space to attach and detach from available LWPs in the LWP pool. LWPs are available to threads in this process only; that is threads are scheduled on these LWPs.

In most cases, threads should be PTHREAD_SCOPE_PROCESS. This allows the threads to float among the LWPs, and this improves threads performance (and is equivalent to creating a Solaris thread in the THR_UNBOUND state). The threads library decides, with regard to other threads, which threads get serviced by the kernel.

System Scope (Bound Threads)

Bound threads are created PTHREAD_SCOPE_SYSTEM. A boundthread is permanently attached to an LWP.

Each bound thread is bound to an LWP for the lifetime of the thread. This is equivalent to creating a Solaris thread in the THR_BOUND state. You can bind a thread to give it an alternate signal stack or to use special scheduling attributes with Realtime scheduling. All scheduling is done by the operating environment.

Note -

In neither case, bound or unbound, can a thread be directly accessed by or moved to another process.


Thread cancellation allows a thread to terminate the execution of any other thread in the process. The target thread (the one being cancelled) can keep cancellation requests pending and can perform application-specific cleanup when it acts upon the cancellation notice.

The pthreads cancellation feature permits either asynchronous or deferred termination of a thread. Asynchronous cancellation can occur at any time; deferred cancellation can occur only at defined points. Deferred cancellation is the default type.


Synchronization allows you to control program flow and access to shared data for concurrently executing threads.

The four synchronization models are mutex locks, read/write locks, condition variables, and semaphores.

Using the 64-bit Architecture

For application developers, the major difference between the Solaris 64-bit and 32-bit operating environments is the C-language data type model used. The 64-bit data type uses the LP64 model where longs and pointers are 64-bits wide. All other fundamental data types remain the same as the in the 32-bit implementation. The 32-bit data type uses the ILP32 model where ints, longs, and pointers are 32-bits.

The following summary briefly describes the major features and considerations for using the 64-bit environment: