Thread-safety for SGI STL

SGI STL provides what we believe to be the most useful form of thread-safety. This explains some of the design decisions made in the SGI STL implementation.

Client must lock shared mutable containers

The SGI implementation of STL is thread-safe only in the sense that simultaneous accesses to distinct containers are safe, and simultaneous read accesses to to shared containers are safe. If multiple threads access a single container, and at least one thread may potentially write, then the user is responsible for ensuring mutual exclusion between the threads during the container accesses.

This is the only way to ensure full performance for containers that do not need concurrent access. Locking or other forms of synchronization are typically expensive and should be avoided when not necessary.

It is easy for the client or another library to provide the necessary locking by wrapping the underlying container operations with a lock acquisition and release. For example, it would be possible to provide a locked_queue container adapter that provided a container with atomic queue operations.

For most clients, it would be insufficient to simply make container operations atomic; larger grain atomic actions are needed. If a user's code needs to increment the third element in a vector of counters, it would be insuffcient to guarantee that fetching the third element and storing the third element is atomic; it is also necessary to guarantee that no other updates occur in the middle. Thus it would be useless for vector operations to acquire the lock; the user code must provide for locking in any case.

This decision is different from that made by the Java designers. There are two reasons for that. First, for security reasons Java must guarantee that even in the presence of unprotected concurrent accesses to a container, the integrity of the virtual machine cannot be violated. Such safety constraints were clearly not a driving force behind either C++ or STL. Secondly, performance was a more important design goal for STL then it was for the Java standard library.

On the other hand, this notion of thread-safety is stronger than that provided by reference-counted string implementations that try to follow the CD2 version of the draft standard. Such implementations require locking between multiple readers of a shared string.

Lock implementation

The SGI STL implementation removes all nonconstant static data from container implementations. The only potentially shared static data resides in the allocator implementations. To this end, the code to implement per-class node allocation in HP STL was transformed into inlined code for per-size node allocation in the SGI STL allocators. Currently the only explicit locking is performed inside allocators.

Many other container implementations should also benefit from this design. It will usually be possible to implement thread-safe containers in portable code that does not depend on any particular thread package or locking primitives.

Alloc.h uses three different locking primitives depending on the environment. In addition, it can be forced to perform no locking by defining _NOTHREADS. The three styles of locking are:

• Pthread mutexes. These are used if _PTHREADS is defined by the user. This may be done on SGI machines, but is not recommended in performance critical code with the currently (March 1997) released versions of the SGI Pthreads libraries.
• Win32 critical sections. These are used by default for win32 compilations with compiler options that request multi-threaded code.
• An SGI specific spin-lock implementation that is usable with both pthread and sproc threads. This could serve as a prototype implementation for other platforms. This is the default on SGI/MIPS platforms.

It would be preferable if we could always use the OS-supplied locking primitives. Unfortunately, these often do not perform well, for very short critical sections such as those used by the allocator.

Allocation intensive applications using Pthreads to obtain concurrency on multiprocessors should consider using pthread_alloc from pthread_alloc.h. It imposes the restriction that memory deallocated by a thread can only be reallocated by that thread. However, it often obtains significant performance advantages as a result.