The first test uses a recursive algorithm which computes the result of applying an associative operation to all elements of an array. Two examples of such operations are summation and maximization. Given that every block will be treated by its own processor and the transmission delays are zero, the algorithm will require O(logm(n)) operations, where m is the parameter of the algorithm (data transmission overhead increases with m). Testing this implementation on MVS-1000M with 100 processors exhibited a 40 times speedup on an array sized 109.

Next, an algorithm of multiple sequence alignment was implemented. It was a part of the project supported by CRDF grant No.RB01227-MO-2. The program was used to align all human-specific LTR class 5 in the EMBL data bank [1]. The test demonstrated a 2.4 times speedup on 12 processors on a Prime Power850 machine.

Third, PARUS was used to simulate a three layer perceptron with a maximum of 18,500 neurons in each layer. The maximum acceleration achieved was over 7 times.

The fourth test is a solution of the problem which is important in molecular biology. This is a problem of multiple sequence alignment. The parallel implementation was based on the MUSCLE package. The procedure of construction of an alignment based on the alignment profiles and cluster sequence tree was parallelized. The data used for testing were all human-specific long terminal repeats (LTR) class 5.

PARUS was installed and tested on some multiprocessors: MVS1000-M (cluster 768 Alpha processors), MVS-15000BM, IBM pSeries690 (SMP 16 processors Power4+ ), Sun Fujitsu PRIMEPOWER 850 Server (SMP 12 processors SPARC64-V).

[1] Alexeevski A.V., Lukina E.N., Salnikov A.N., Spirin S.A. Database of long terminal repeats in human genome: structure and synchronization with main genome archives //Proceedings of the fours international conference on bioinformatics of genome regulation and structure, Volume 1. BGRS 2004, pp 28-29 Novosibirsk (link)