Parallel Systems

EPPA for pvm: User Documentation

Code Instrumentation

The files & their purpose

On our LINUX system, the files are under the PARDIR directory (set to /home/parallel/software).
Or download the files from the pvm example with EPPA.
The files:

the executable EPPA
the header files: EPPAInstrumentation.h, standardlib.h
the library: libpar.a (contains the necessary object files)

The compiler needs:
    pvm, mysql : pvm and database
    parlib.a : library with the class measure
    EPPAInstrumentation.h & standardlib.h : headers for compiling

Include them in the projectfile as follow:

INCLUDEPATH = /usr/include/g++ $(PVM_ROOT)/include $(MYSQLDIR)/include $(PARDIR)/include
LIBS = -L$(PVM_ROOT)/lib/LINUX -L$(MYSQLDIR)/lib -L$(PARDIR)/lib-lpar -lpvm3 -lgpvm3 -lmysqlclient

If you put the EPPA files in a different directory as your code, add this path to the INCLUDEPATH and change -L$(PARDIR)/libinto -L plus the path

Class EPPAProbe (see EPPAInstrumentation.h)

Start:

EPPAProbe(const string& alg_name, int machines, DO_WITH_SIMILAR_EXP doWithSimilarExp=REPLACE_IF_FASTER, int W=1, const string& comment_str = " ")

constructor for the master
alg_name: specify a name of the parallel algorithm under study
machines = # processors. If your master doesn't do anything useful, it must NOT be included, then it will not be used in the performance calculation. To write the sequential experiment to the database, set machines to 1.
doWithSimilarExp: experiments of the same user, with the same name, #processors and problem size W are considered to be similar, so there are no reasons to keep multiple measurements in the database. Specify here what should be done with similar experiments

possible values: enum DO_WITH_SIMILAR_EXP {REPLACE_IF_FASTER, REPLACE_IF_SUCCESFULL, ASK_USER, DO_NO_REPLACE_EXP};
REPLACE_IF_FASTER (default): experiment replaces similar experiments, if the runtime is less.
REPLACE_IF_SUCCESFULL: the new experiment will always replace similar ones (if the master ends without crashing).
ASK_USER: the user will be asked what should happen
DO_NO_REPLACE_EXP: do not replace anything.

(optional) W = the problem size (e.g. # cities for TSP)
(optional) comment text
each experiment gets a unique number: the experimentID

EPPAProbe();

this will not activate the measuring. Use this for the master.


EPPAProbe(bool activateMeasurements);

constructor for the slaves. Specify whether to activate the measuring, be aware that measuring should be on or off for all processes.

void StartMeasuring();

(optional) This restarts the time measurement, which starts automatically after the constructor. Don't call it when there are already phases instrumented.

void SlaveIsSpawned(int slaveTid);

Mandatory! Call this function whenever a process is spawned, pass its pvm task id (tid). A phase of type PH_SPAWN is automatically added to the measurements.

Destructor:

~EPPAProbe();

Ends the parallel measurement, then calculates and writes the totals to the database.

void Stop();

(optional) Use if you want to stop the measurents before the destructor is called.

Measure:

TIME EndOfPhase(int phaseType, int nbrQuantums=0, int nbrOperations=0, string comment="");

Indicate the end of a phase of your parallel program.

Specify the type of the phase :

 enum BasicPhaseType {PH_IDLE=2, PH_WORK=3, ..., PH_PART=8, PH_SYNC=9};

  PH_IDLE: whenever the process was idle (at blocking receive points)
  PH_WORK: when the process has performed useful work
  PH_PART: the partitioning of the work
  PH_SYNC: for extra synchronisation overhead, specific for your parallel algorithm

(optional) nbrQuantums: value for the size of work, the number of basic 'quantums' processed. For the moment, this is just for information, it is shown on the execution profile of EPPA..
(optional) nbrOperations: the number of operations performed, if different than nbrQuantums (like the number of sort operations). NOT USED FOR THE MOMENT
(optional) comment: for informative reasons only. Can be changed with EPPA.
Returns the duration of the elapsed phase in us. TIME is defined in standardlib.

TIME EndOfCommPhase(int phaseType, int commPartnerTid=-1, int nbrQuantums=0, string comment="");

Indicate the end of a communication phase of your parallel program.

Specify if sending or receiving:

 enum BasicPhaseType {..., PH_COMM_SEND=5, PH_COMM_RECV=6, ...};

(optional) add with comm_partner_tid the task_id of the communication partner, so that the communication can be shown with EPPA!
(optional) nbrQuantums: value for the size of communication, like the number of bytes or packets you send. For the moment, this is just for information, it is shown on the execution profile of EPPA..
(optional) comment: for informative reasons only. Can be changed with EPPA..
Returns the duration of the elapsed phase in us. TIME is defined in standardlib.h

Information:

int GetExperimentID(); // database experiment ID
TIME ViewExpChrono(); // time since start of experiment (us)
TIME ViewPhaseChrono(); // time of current phase (us)

<>See also pvm example program with EPPA

<>

- Back to the top -