Research On Performance Tuning Methodology Based On SMP Cluster

In the early of 1980s, Intel 8088 microprocessor is running at 4.77 MHz. Twenty years later, a 3.6-gigahertz Intel Pentium XERO processor powers the fastest personal computers-a speed increase of over 600 times.Today's software is more complex and is packed with more features than the twenty-year old simple text-base applications. From games and educational software to databases and operating systems, everything uses more computing resources than before and today's software is still hungry for more. Software performance can vary dramatically depending upon how the programmer chooses to solve a problem or implement a feature. Highly optimized applications can run tens of times faster than poorly written ones. A combination of using efficient algorithms and well-designed implementation implementations leads to great high performance applicationsAdopting the imaging theory of synthetic aperture telescope in 21 CMA, the voltage signal received by the aerial can be processed through FFT and then correlated. At last, complex visibility function can be obtained for the further mapping. The amount of the signals received is so great that correlator should be real-time, which can compress the great number of storage from 1. 38PB to 890GB, supposing the save time is 10 seconds' interval. And the minimal unit can be obtained at the same time. The former formally used correlator is based on hardware, with the weak points of long-last developing period, poor flexibility and high cost. This thesis mainly discusses the creative design, with correlator realized through general software on cluster, which is able to overcome the above weak points. And it can provide a good example for the realization of the imaging theory of synthetic aperture telescope.A set of high-performance cluster on the basis of one pair of networks (Giga Ethernet and Infiniband network) has been set up, and the nodal performance and corresponding management has been optimized. It is most import to make a full and pertinent developing on cluster performance, and to design the software with suitable parallel algorithm.