Programming Model

The programming model is defined as a set of library calls which can be called from C and C++. We choose these languages for the following reasons:

• Only languages which have some kind of a `pointer' make it possible to implement the routines, which access the shared-memory, as library calls. Otherwise costly extensions to the languages would have been needed.
• The C compiler is available on every UNIX system. As it is also used to develop the operating system itself, more effort is taken by the manufacturer to make this compiler bug-free, stable and have it generate optimized code.
• Compared to programs using index references, programs using pointer references can lead to more efficient machine code.

Mechanisms

Shared-Memory

The use of the shared-memory is based on the concept of `segments', very much like the original shared-memory mechanism provided by UNIX System V. A process which wants to share data with another process (possibly on another node) first has to create a shared-memory segment of the needed size. To have the operating system select the correct location for this memory segment, the process has to specify with which neighbouring nodes this segment needs to be shared.

After the segment has been created, other processes may map the same segment into their address space by the means of an `attach' operation. The addresses in these address spaces are totally unrelated, so pointers may not be passed between different processes. The segments may also be unmapped and destroyed dynamically.

There is one disadvantage to the shared-memory implementation on the MEMSY system. To ensure a consistent view over all nodes the caches of the processors must be disabled for accesses to the shared-memory. But the application programmer may enable the caches for a single segment if he is sure that inconsistencies between the caches on different nodes are not possible for a certain time period. The inconsistencies are not possible if only one node is using this segment or if this segment is only being read.

Messages

The message mechanism allows the programmer to send short (2 word) messages to another processor. The messages are buffered at the receiving side and can be received either blocking or non-blocking. They are mainly used for coordination. They are not especially optimized for high-volume data transfer.

Semaphores

To provide a simple method for global coordination, semaphores have been added to the programming model. They reside on the node on which they have been created, but can be accessed uniformly throughout the whole system.

Spinlocks

Spinlocks are coordination variables which reside in shared-memory segments. They can be used to guard short critical sections. In contrast to the other mechanisms this is implemented totally in user-context using the special machine instruction `XMEM'. The main disadvantage of the spinlocks is the `busy-wait' performed by the processor. This occurs if the process fails to obtain the lock and must wait for the lock to become free. To minimize the effects of programming errors on other applications, a time-out must be specified, after which the application is terminated (there is a system-imposed maximum for this time-out).

I/O

Traditional UNIX-I/O is supported. Each processing element has a local data storage area. There is one global data storage area which is common to all processing nodes.

Parallelism

To express parallelism the programmer has to create multiple processes on each processing element by a special variant of the system call `fork'. Parallelism between nodes is handled by the configuration to be defined: One initial process is started by the application environment on each node that the application should run on. In the current implementation these processes are identical on all nodes.

Information

The processes can obtain various information from the system regarding their positions in the whole system and the state of their own or other nodes.