NXS Message-Passing Library

The NXS library is a small subset of the NX library provided by Intel Corporation for their parallel computers. The NXS library was developed for use with the UNH C* system. However, it can also be used for "directly" programming the Galaxy cluster.

The NXS library provides the following primitives:

• long numnodes(void): returns the number of processors participating in the parallel computation.

• long mynode(void): returns the logical number (0 ... numnodes()-1) of the processor executing the call.

• void csend(long type, char *buf, long len, long node, long pid): sends a message to another processor. This is a non-blocking send: the message will be copied and queued for transmission; return of this function does not imply the message has been received at the destination processor.

  A description of the parameters follows:

  • type - user specified message type that can be used by the receiver to selectively receive messages.
  • buf - source of message, which must be a string of contiguous bytes.
  • len - length of the message, which must not exceed CS__MESSAGE_LIMIT.
  • node - logical number of the destination processor.
  • pid - should always be set to zero.

  Note: It is considered an error for a node to send a message to itself.

• void crecv(long typesel, char *buf, long len): receive a message. This is a blocking send: the function will not return until a message with the matching type is received and copied to the given buffer.

  A description of the parameters follows:

  • typesel - message type selector.
  • buf - buffer in which to place the message.
  • len - length of the buffer.

  Note: it is an error to receive a message that is longer than the length of the buffer.

• long iprobe(long typesel): check for the availability of a message matching the given type selector. Returns 1 is a matching message is available; returns 0 otherwise.

• long infocount(void): returns the length of the last message received or the length of the last message found by iprobe.

• void gsync(void): the function synchronizes all processors participating in the parallel computation. All processors must call this function together, otherwise the computation will block forever.

You should include nxs.h in any source file using the NXS library.

Programs using NXS should be compiled using the UNH C* driver. The UNH C* driver will locate the NXS header file, will call the C compiler, and will link with the NXS library. See the Galaxy cluster homepage for details on accessing the UNH C* driver.

CAUTION: The UNH C* driver, by default, provides C* versions of the ANSI stdio routines. To use the conventional stdio routines with NXS programs, define the symbol CS__NO_STDIO before including stdio.h.

The C* versions of the stdio routines must be executed by all processors. The processors are synchronized for each operation so that only processor 0 actually does the I/O. The side effects of the I/O operation are then broadcast to the other processors. If you use the C* stdio library with a NXS program, you must call numnodes() or mynode() prior to invoking any I/O operation.

Comments and questions should be directed to pjh@cs.unh.edu