THE COMPUTER MODEL OF CLUSTER OF PARALLEL CALCULATION.
Zhoranova N. Zh., Tulegenova S.N.
The increase of
compound calculation in using imitative models in various system leads to
considerably was tings of tame at making investigation. To shorten the tame at
making investigation could be, if it is used parallel calculation on
specialized cluster for this purpose.
In whole case,
calculation cluster-it is the collection of computers calculation bundles which
have own operative memory and work under the defection of operative system.
Most popular is the using of similar clusters, where all the bundles are
absolutely the same in its architecture and productivity.
There is the
dedicated computer for each cluster-the leading machine frontend. The
installation of programmer guarantee in this machine directs the starting of programmer on cluster. In fact, calculation process of users set on
calculation bundles, therefore they are divided so that each processor has no
more than one calculation process. It isn’t allowed to start calculation process on the leading machine
(frontend).
The users have
terminal access to the leading machine cluster, and there is no need to enter
the bundles of cluster. The starting of programmer on cluster realized in
«packet regime»- that is the user has no immediately internationally
cooperation with programmer ,the programmer. Car not wait for the leading in
facts from cavitations and sending out immediately on screen. The fore the programmer
of users can work when the user does not connect the cluster.
Calculation cluster, as a rule, works
under the direction of one of various OC Unix – mule using mute purposing net
of operative system or one of version set OC Windows. Particularly, moss of
clusters work under the direction of OC-Linux- freely widening variant Unix. Unix, particularly, has number of
differences from Windows, which usually
works on personal computers, where there differences concern mainly interface
of course the work of processor and file system.
All bundles of
cluster have access to common file system, which is fileserver. That is, the
file might be created, for example, on the leading computer or on one of the
bundles and then it can be read under the same name on the other bundle. The
note or one file on the same tame from different bundles is impossible. Besides
common file system, there may be local disks on bundles of cluster. They might
be used with programmers for keeping temporary files. After finishing (exactly,
immediately before concluding) the work these files must be send away.
No special
parallel completers for clusters are existed. Common optimistically completers
are used with the language Cu and Fortran-GHU, Intel or others, can create
fulfilling programs OC Linux. As a
rule, for the compilation of parallel MPI program special scripts are used
(mpicc, mpit 77, mpit 0 and others )
which turn out to be superstructures over comparators and allow to
include necessary libraries.
There existed
several ways of acting calculation power of cluster.
1. To start great
number of single processed tasks. In may reasonable variant, if it needs to set
great deal of independent calculating
experiments with different entry facts, at the same tame the date of leading of
each separate calculation doesn’t
match and all the facts might be placed in amount of memory, available for one processor.
2. To start
prepared parallel programmer. As a rule it is enough in order the programmer
will be available in outcome texts, realized with using interface MPI on
language C/C++ or Fortran. The signs of
freely spreading parallel programmer, realized with the help of MPI: Games-US
(quantum chemistry) POV Ray-MPI (reading rays ).
3. Calling in own
program parallel libraries. As well as for some field, such as line algebra,
available libraries, which allow to
solve wide sphere of standards super tasks using possibilities of parallel
process. If addressing to such super tasks forms major part calculative
operation program, then using such parallel library make it possible to get
parallel program practically without writing own parallel code.
4. Sitting up own
parallel programmer. This is the most difficult, but the most universal way.
There exist two
main variants:
1.Putting parallel
constructions in existing parallel programs.
2. Settling up
parallel programs fro “zero” (0).
Parallel
programs on calculating cluster work in
the model of message passing. That means the program consist of great number of
process, each of them works on its own
processors and has its own address space. The direct access to the memory of other processor is impossible, but
exchange of facts between processors takes place with the help of operation
receptions and sending messages. The process which must receive facts, call the operation “Receive
(to get message) and points exactly from what process if must receive facts to
other call operation “send” (send a message) and points exactly which processor
must be send these facts. This model is realized with the help of standard
inter face MPI, and particularly the packer of programmers. MPICH2-quickly
acting and widely porting realization of standard MPI (realized both standards
MPI-1 and MPI-2). There exiled different realization of charge and commercial,
transferred, and oriented on the concrete communicative net.
As a rule MPI-
programmers are based on the model SPMD (one programmer – many facts )
that is for all kinds of
processes only one code programmer, and
different processes keep different facts and fulfill their action depending on
ordinal number of process.
With the problem
of construction of imitative model on
the base of parallel calculation, in any case, everyone may come against, who
sets and, tries to solve calculative lacks on cluster system.
It s necessary to
take account, that the organization of cluster offers sufficiently complicated
tasks and not all the universities have them. That is why in all cases
mathematic-calculator in preparing programmer limits with empirical valuation
fulfilled according to personal experience and intuition, seeing that the given
aspect of staging task don’t play an important role and may be passed by on
missed at all. In solving tests or not large tasks in size, one can act by
guesswork, but in planning enormous
calculative experiment, the time of
count is calculated daily and weekly, the choice of the method of dividend, the
size of super tasks or the way of organization actually the calculation may
vitally affect the time of count, and of course the result of experiment at
all.
That is why on
working out the complicated program for the construction of imitative model on
cluster, it is necessary to carry out beforehand control of working ability of
program, imitating cluster on usual computer with the help of complete set of
virtual machine.
·
construction of such model parallel calculation giver
the opportunity to provide fugitive and quantities analyses of process of
calculation;
·
to fulfill apron and a posterior valuation of
effeteness of parallel calculation.
·
to make a
valid choice algorithm of decision, ways and parameters of dividing of this
algorithm for reaching optimal meaning of speeding or effect.
The
worked out in such way, parallel program may be realized on the real cluster,
with the help of passed by access.
In
real case it may be separated by two points of view for model: these point are
“from instrument” and “from task”.
The method “from instrument” includes the analysis of characteristic of
cluster as the program – apparatus system carrying out some complete set of
operation.
The method “from task” consists of examining calculative process from
the point of view of its main function solving calculative tasks, in known
degree abstractedly from program –
apparatus guarantee.
At the same time, adequate constructive model of parallel calculates ion
are the actual and difficult solving problem yes, in all, demanding vital
spending for the organization of cluster. That is why, offering method of
creating on the special computer the complete set of virtual machine and net
system, imitating the work of cluster with the following use of passed by
access to real cluster, allow to shorten the time of preparing the program of
parallel calculation and to raise the quality of realizing models.
The leading empirical way of planning nowadays and staging of parallel
calculative experiments must by and most likely will by force out by using
theoretically well-founded and experimentally tests models, especially no more
principle hindrance for the construction of such model are not found out yes.
The main parameters used in the method “from instrument “ in most cases are
well come hear to analytically cover. When it isn’t so, one may try to
construct the special king or to use the probability methods of valuation.
Considering “from task” may lean upon well – known methods following
calculative mathematic, completed or overworked according to specifically
parallel calculation.
The list of literature:
1.Parallel.Ru Information -analytical materials on
parallel calculations (Electronic resurce)
access:www.parallel.ru/2001-2013.
2. Malyshev L.V Some aspects of modeling of parallel
and cluster calculations/ Sibirial shool- seminar on parallel and cluster cflculation/. Ediled by professon A.V.Starchenko. Tomsk, publishing nouse of.Tomsk University
2002-129