Optimization of Message Passing Libraries – Two Examples

The computation of large problems in scientifical and industrial fields needs for efficient computer systems. A rising number parallel computers are used to deliver necessary computation resources. The physical restrictions of circuit integration limit the speed of single processor solutions. The use of consumer components is a cost–effective way to built a parallel computer. The deployment of a high–speed network enables most parallel applications to run fast. Optimized communication libraries are important to enable the use of high–speed networks with parallel applications. This field is subject to active research and development. 1 An optimized Library on Top of SISCI Parallel computations need for information exchange to compute meaningful results. The quality of the network may have a significant impact on the execution speed of some applications. The Scalable Coherent Interface (SCI) was designed to satisfy demands for high bandwidth and low latency. Applications running on clusters of workstations may benefit from SCI interconnections. As of today a single SCI link can transmit up to 320 MByte per second [15]. Also network latancy is much lower than expected from conventional network links. The latency period for transmitting a 4 byte message is considered to be about 1.4 μs. Compared with conventional ethernet communication over TCP/IP these values are realy impressive. The small delay of network transmissions is considered to be the highlight of SCI technology. Even the upcoming Infiniband technology has a latency larger than 4 μs [16]. Depending on the behaviour of the applications the use of SCI networks can lead to a significant speedup. 1.1 General Overview Many cluster computers are equipped with a SCI network. The native interface to the SCI services is known as Software Infrastructure for SCI (SISCI). This specification defines a programming interface that follows the Distributed Shared Memory (DSM) paradigm. The programmer operates on memory areas that are shared by several processes. The participating processes may run on different nodes located anywhere inside the SCI cluster. In contrast to the SISCI interface the message passing paradigm exposes inter–process communication. The data transfer has to be handled by executing dedicated send functions and receive functions. This paradigm is widely used. Many distributed applications rely on the presence of message passing libraries. To enable the use of this kind of applications with SCI the VIA2SISCI library was developed. The aim was to implement a message passing library that uses SISCI functions to provide fast data transfer. The Virtual Interface Architecture (VIA) contains the definition of a message passing interface.