This is both the reference manual and the User's Guide for the p4 parallel programming system. It contains definitions of all functions for both C and Fortran, examples, a brief tutorial, and discussions of related systems.
and
P4 is a library of macros and subroutines developed at Argonne National Laboratory for programming a variety of parallel machines. A paper describing its functions and use is [(ref butler-lusk:p4-paper)]. Its predecessor was the m4-based ``Argonne macros'' system described in the Holt, Rinehart, and Winston book Portable Programs for Parallel Processors, by Lusk, Overbeek, et al., from which p4 takes its name[(ref lusk-overbeek:p4-book)]. The current p4 system maintains the same basic computational models described there (monitors for the shared-memory model, message-passing for the distributed-memory model, and support for combining the two models) while significantly increasing ease and flexibility of use. See section Getting Started for a simple example.
P4 is intended to be portable, simple to install and use, and efficient. It can be used to program networks of workstations, advanced parallel supercomputers like the Intel Touchstone Delta and the Alliant Campus HiPPI-based system, and single shared-memory multiprocessors. It has currently been installed on the following list of machines: Sequent Symmetry (Dynix and PTX), Convex, Encore Multimax, Alliant FX/8, FX/800, and FX/2800, Cray X/MP, Sun (SunOS and Solaris), NeXT, DEC, Silicon Graphics, HP, and IBM RS6000 workstations, Stardent Titan, BBN GP-1000 and TC-2000, Kendall Square, nCube, Intel IPSC/860, Intel Touchstone Delta, Intel Paragon, Alliant Campus, Thinking Machines' CM-5, and the IBM SP-1 (TCP/Ethernet, TCP/switch, EUI, and EUI-H). It is not difficult to port to new systems. Although p4 tries to be completely portable, there are a small number of specific exceptions (See section Machine-Specific Notes) that may need to be taken into account on a given machine.
You can obtain the complete distribution of p4 by anonymous ftp from info.mcs.anl.gov in the directory pub/p4. See the README file there for recent news on what is available. Take the file p4-1.4.tar.Z. The distribution contains all source code, installation instructions, this reference manual, and a collection of examples in both C and Fortran. Alog is included in the distribution with p4. The file upshot.tar.Z contains display facilities that can be used with p4 and other systems.
To ask questions about p4, report bugs, contribute examples, etc., you can send mail to p4@mcs.anl.gov.
The current release is version 1.4. You can check which version of the source code you have by looking at the file lib/p4_patchlevel.h in the distribution. You can check which version of the object code you have linked to by running any p4 program with the command-line option -p4version (See section Command-Line Arguments).
Salient features of p4 include:
The p4 source code distribution contains the following files and subdirectories:
In this section we describe how to install the p4 library, either for your own personal use or for the use of everyone at your site. In the first case you do not need any super-user privileges. In the second case, you may or may not, depending on how things are configured at your site. We also describe how to install and run the examples that come with p4, the online help system (this manual as a World Wide Web document) and how to build a working directory for your own programs yet share the installed copy of p4 with other users.
To build p4, position yourself in the top-level p4 directory (Here we refer to
this directory as p4, but you may have it as p4-1.4 or something
similar) and type:
make all P4ARCH=<machine>where <machine> is one of the machine names listed in p4/util/machines, currently:
make all P4ARCH=SYMMETRYThe all is optional, for example
make P4ARCH=SYMMETRYThis will create a machine-dependent Makefile in each subdirectory, make the p4 library, and compile and link a subset of the examples.
To add a new machine type, or to change the characteristic parameters associated with an existing one, you can edit the file p4/util/defs.all.
To save disk space, various intermediate object files can be removed with
make cleanThe system can be restored to its original, machine-independent state with
make realcleanNote that this removes the machine-dependent Makefiles in each directory, so the operation is not idempotent.
It is also possible to install (or clean) only some of the directories:
make all P4ARCH=SUN DIRS=messages make clean DIRS='monitors messages'To install only the Makefiles in all subdirectories, use:
make makefiles P4ARCH=<machine>To install the necessary library and include files in a directory everything that is needed to compile and link p4 programs, do:
make install INSTALLDIR=<dir>This will create a p4 directory in <dir>, build a minimal set of directories, copy the relevant .a and .h files into it, and test the installation by mking a small set of examples.
See section Getting Started for instructions on how to run some example programs after you have installed p4.
The directory p4/doc contains this manual as well as files that require installation. This manual was prepared with the latexinfo package from GNU emacs. The files in p4/doc are:
A good way to see how various p4 functions are used is to look at the example programs included in the distribution. The p4/monitors directory contains shared-memory examples written in C that use monitors, including one instrumented with ALOG. The p4/messages subdirectory contains message-passing examples written in C. The programs in p4/messages_f are Fortran message-passing examples, and the p4/contrib and p4/contrib_f directories contain a number of miscellaneous examples contributed by users. In each directory there is a README that describes the individual examples.
The easiest way to get started with p4 is to play with some of the sample programs provided with the system.
We will begin with a message-passing example in the sub-directory named p4/messages. The code for the program is in the files sr_test.c and sr_user.h.
As the name implies, this program is an example of p4's send/receive functionality. Briefly, it is a simple program that runs a master process and some slave processes. The master and the set of slaves form a ring of processes in which the master reads a message from stdin and sends a copy of the message to the first slave, which passes it on; the last slave passes the message back to the master. If the master receives an undamaged copy of the message, it assumes that all went well, and reads another message. Note that the ring of processes is a logical structure in which each process assumes that its predecessor in the ring is the process with the next lower id, and its successor is the process with the next higher id. The master has id 0 (zero) and has the process with the largest id as its predecessor.
The first executable p4 statement in a program should be:
p4_initenv(&argc,argv);This initializes the p4 system and allows p4 to extract any command line arguments passed to it, e.g. debugging parameters.
Similarly, the last executable p4 statement in a program should be:
p4_wait_for_end();This waits for termination of p4 processes and performs some cleanup operations.
The procedure p4_get_my_id returns the unique integer id assigned to the calling process by p4.
The statement:
p4_create_procgroup();reads a procgroup file that the user builds and creates the set of slaves described in that file. Obviously this statement must be executed before any slaves can be assumed to exist. This procedure is the method you must use to create processes that do message-passing.
The procedure p4_clock returns an integer that represents wall-clock time in milliseconds. It is typically used to retrieve the time before and after some work, the difference representing the time to do that work. Note that there is also a p4_ustimer that is useful on those machines that support a microsecond timer.
The procedures p4_send and p4_sendr are two of several p4 procedures that are available for sending messages to other processes. They take as arguments the message type, the id of the "to" process, the address of the message, and the message length.
The procedure p4_recv receives a message from another process and sets the values of all four parameters. P4_recv will automatically retrieve a buffer in which to place a received message, thus p4_msg_free may be called to free that buffer when it is no longer needed.
The procedure p4_num_total_slaves is one of several procedures that the user can invoke to determine information about the current execution.
To run this program, you need to create a procgroup file that describes
where all slave processes are to be executed
(See section Specifying Processes in the Procgroup File). We will assume that
you have an example procgroup file (named sr_test.pg) in the
p4/messages directory, and can run sr_test by merely typing:
sr_testIf the procgroup file is elsewhere, then you must type:
sr_test -pg pathname_of_procgroup_fileAnother example that is made by default is the program systest. It tests a number of the message-passing features of p4.
The procgroup file is the only portion of the interface that is very likely to change through multiple versions of p4. As new architectures are supported, it is hoped that we can merely alter the procgroup file format to reflect any new features. (Of course new procedure calls may also be required, but existing procedure calls will remain unchanged when possible). See See section Running p4 on Specific Machines for a discussion of machine dependencies in starting p4 programs.
The current format of a procgroup file is as follows:
local_machine n [full_path_name] [loginname] remote_machine n full_path_name [loginname] . . .
In some situations, the program is started via some special command executed from the host machine. In such cases, the procgroup file name can be specified to the special command line along with the program name (see for example the runcube and rundelta shell scripts in the p4/messages subdirectory). In those cases where no special command is required, no special handling is required for the procgroup filename.
The first line of a procgroup file may be ``local n'' where n is the number of slave processes that share memory with the master. The full path name on the ``local'' line is ignored on machines other than cube and mesh machines, and the IBM SP-1. The word ``local'' may be replaced by an alias for the local machine if needed, to specify an alternative transport layer. The subsequent lines contain either three or four fields:
# start one slave on each of sun2 and sun3
local 0
sun2 1 /home/mylogin/p4pgms/sr_test
sun3 1 /home/mylogin/p4pgms/sr_test
Lines beginning with # are comments.
It is also possible to have different executables on different machines. This is required, of course, when the machines don't share files or are of different architectures. An example of such a procgroup file would be:
local 0
sun2 1 /home/user/p4pgms/sun/prog1
sun3 1 /home/user/p4pgms/sun/prog2
rs6000 1 /home/user/p4pgms/rs6000/prog1
On a shared memory machine such as a KSR, in which you want all the processes
to communicate through shared memory using monitors, the procgroup file can be
as simple as:
local 50
On the CM-5, your procgroup file would look like:
local 32 /home/joe/p4progs/cm5/multiply
Next, let's assume that you have a Sequent Symmetry (named symm) and an
Encore Multimax (named mmax). We will also assume that you are working
on symm, and plan to run the master there. If you would like to run
two processes on symm (in addition to the master) and two on
mmax, then you might code a procgroup file that looks like:
local 2
mmax 2 /mmaxfs/mylogin/p4pgms/sr_test
P4 also permits you to treat the symmetry as a remote machine even when
you are running the master there. Thus, you might code a procgroup file
as follows:
local 2
symm 2 /symmfs/mylogin/p4pgms/sr_test
mmax 2 /mmaxfs/mylogin/p4pgms/sr_test
In this example, there are seven processes running. Five of the
processes are on symm, including the master. Two of the processes on
symm are in the master's procgroup and two are running in a separate
procgroup as if they were on a separate machine. Of course, the last
two are running on mmax.
Finally, suppose that you have a fiber-channel network that parallels your Ethernet, connecting the same machines, and that connections fro running TCP/IP over the fiber-channel network are obtained by connecting to sun1-fc, sun2-fc, etc. Then even if sun1 is the local machine that you are logged into, you will want your procgroup file to look like:
sun1-fc 0
sun2-fc 1 /home/user/p4pgms/sun/prog1
sun3-fc 1 /home/user/p4pgms/sun/prog2
Some notes about the contents of the procgroup file should be made at
this point. First, the value of n on the local line can be zero,
i.e. the master may have no local slaves. Second, the local machine
may be treated as if it is a remote machine by merely entering it in
some line as a remote machine. Third, a single machine may be treated
as multiple remote machines by having the same remote machine name
entered on multiple lines in the procgroup file. Fourth, if a single
machine is listed multiple times, those processes specified on each
line form a single cluster (share memory). Fifth, the cluster size
specified for a uniprocessor should be 1, because all slaves in a
cluster are assumed to run in parallel and to share memory.
We refer to the original (master) process as the ``big master''. The first process created in each cluster is the ``remote master'' or the ``cluster master'' for that cluster. All p4-managed processes (see the procedure p4_create_procgroup) have unique integer id's beginning with 0. The processes within a cluster are numbered consecutively.
The real fun associated with any computing environment arrives when you actually type in a program and run it yourself. We will assume that you have successfully installed p4 on your own system and are ready to write a small program, compile it, and run it.
We will start with a tiny program in which the worker processes do no work, and then expand its capabilities. Edit a file called p4simple.c and type:
This is one of the simplest p4 programs that you can write. Let's examine it. The #include "p4.h" statement must appear in all programs that use any p4 features. The procedure p4_initenv must be invoked before any other p4 procedures, and p4_wait_for_end must be invoked after all p4 processing is completed. The p4_get_my_id returns a unique integer id for each process, beginning with 0. The procedure p4_create_procgroup is responsible for creating all processes other than 0. It has no effect if called by any other processes than process 0. The way in which p4_create_procgroup determines how many other processes there should be, and where they should run, will be discussed shortly.#include "p4.h"
main(argc,argv) int argc; char **argv; { p4_initenv(&argc,argv); p4_create_procgroup(); worker(); p4_wait_for_end(); }
worker() { printf("Hello from %d\n",p4_get_my_id()); }
All processes that this program executes invoke the worker procedure, including process 0. Thus, in this program, the master process acts just like all other processes once it gets the environment established.
To understand how things get started, let's consider two separate situations. In the first situation, all processes are running on a single machine. Then, when process 0 starts, it executes the p4_create_procgroup procedure to start all other slaves. The other slaves are started on the same machine by means of a UNIX fork.
In the second situation, there may be slaves running both on the same machine as process 0, and slaves running on other machines as well. In this situation, the first slave running on a remote machine will need to execute the main procedure. It will discover that it is not process 0. However, as part of initialization, process 0 will direct it to fork any additional slaves required on the same machine.
In some ways, the above example can be used as a prototype for all p4 programs, just by varying the content of the worker routine.
Here is a Fortrran version of the program we just discussed.
program p4simple
include 'p4f.h'
call p4init()
call p4crpg()
call fworker()
call p4cleanup()
stop
end
subroutine fworker()
include 'p4f.h'
integer*4 procid
procid = p4myid()
print *,'Hello from ',procid
end
Now, let's make the worker process a little bit more interesting. Let's assume that we have nprocs slaves with ids 0, 1, 2, ... nprocs -1. And, we want to write a program in which every process sends a single message to every other slave, and then receives a message from every other slave. We might alter the code for the worker procedure to be the following:
worker()
{
char *incoming, *msg = "hello";
int myid, size, nprocs, from, i, type;
myid = p4_get_my_id();
nprocs = p4_num_total_ids();
for (i=0; i < nprocs; i++)
{
if (i != myid)
p4_send(100, i, msg, strlen(msg)+1);
}
for (i=0; i < nprocs - 1; i++)
{
type = -1;
from = -1;
incoming = NULL;
p4_recv(&type,&from,&incoming,&size);
printf("%d received msg=:%s: from %d",myid,incoming,from);
p4_msg_free(incoming);
}
}
This program demonstrates several features of p4's support for
message-passing. Before we get into the specifics however, let's
examine the overall logic of the program. Each process determines its
own id and the total number of processes executing in this run
(including process 0). Then, in the first for-loop, each process sends
a single message to each of the other processes. Finally, in the
second for-loop, each process receives a message from each of the other
processes.
The p4_send call requires 4 arguments:
Finally, note that p4_msg_free frees the message buffer obtained by p4_recv. The procedure p4_msg_free should be called only after the contents of the message are no longer needed. P4_msg_free should be used to free these buffers because, although a user only sees the data portion of a message, p4 internally represents a message as a structured data item.
To compile and link this program for execution, you need to create a makefile. We will assume that you have installed p4 in /usr/local/p4 and that you have typed the program above into a file name p4simple.c in the directory /home/mylogin/p4pgms.
To build your makefile, copy the file
/usr/local/p4/messages/makefile.proto
into your working directory. This is a prototype makefile that
contains machine-independent information, and which p4 can use to build a
machine-specific makefile for your program. This prototype makefile contains
information about several sample programs that demonstrate
message-passing in p4. If you edit this file, you will see information
for making a program named sr_test. Do a global change of
sr_test to p4simple. You should also change the value of
P4_HOME_DIR. It should contain the full
pathname of the p4 system, e.g. /usr/local/p4. Now change
directories to /usr/local/p4 and type:
make makefiles P4ARCH=<machine_type> DIRS=/home/mylogin/p4pgmswhere <machine_type> is the machine type that you specified when you installed p4 on your machine. Now, you should be able to change back to your directory and see a file named Makefile there. You should then be able to type:
make p4simple
There is one last piece missing before you can execute your program. Recall
that p4_create_procgroup needs to know how many processes to start and
where to start them; it reads a file (called a procgroup file) to gather
this information. P4 always assumes that you have a master process, and that
you describe the slave processes (process groups) in the procgroup file. You
can name a procgroup file any name you choose, but <progname>.pg is the
default name. For example, in this case your procgroup file should be named
p4simple.pg. The information contained in procgroup files can get
fairly involved, but if you have a computer that supports shared memory among
processes, then you can code a very simple example at first.
Let us suppose first that you want to run your program on a network of workstations. Then your procgroup should look something like:
local O
some.network.machine 1 /home/me/p4progs/p4simple
This file indicates that you wish to run only the master on the local machine
(the one you are logged into when you execute the program) and one slave on
the machine some.network.machine.
Now, all you have to do to run your program is type:
p4simple
You should see a line printed each time a process receives a message
from another process (on some machines, there may be a restriction that
only one process can do I/O, however such restrictions are not
common). Experiment by changing the number of slaves indicated in the
procgroup file.
You may notice that even a small p4 program becomes large when linked with the p4 library. You might consider using strip to reduce the size or removing -g from the CFLAGS in the makefile.
The command-line arguments to a p4 program are all optional.
-p4help get this information, then exit -p4pg <file> set procgroup file -p4dbg <level> set debug level -p4rdbg <level> set remote debug level -p4gm <size> set global memory size -p4dmn <domain> provide local domain name -p4out <file> set output file for master -p4rout <file> set output file prefix for remote masters -p4ssport <port#> set private port number for secure server -p4norem don't start remote processes -p4log enable internal p4 logging by alog -p4version print current p4 version numberIn version 1.4, these flag names are valid without their p4 prefix, for backward compatibility.
If one specifies -p4norem on the command line, p4 will not actually start the processes. The master process prints a message suggesting how the user can do it. The point of this option is to enable the user to start the remote processes under his favorite debugger, for instance. The option only makes sense when processes are being started remotely, such as on a workstation network.
In the following sections, we provide details for each p4 function in the library. The procedures are gathered into the following groups:
Most p4 functions return -1 if an error occurs. Some, however, call the function p4_error when severe errors occur. This function prints a message and then attempts to terminate all of the user's processes See section Functions for Debugging p4 Programs.
In some situations a p4 procedure will give an error message and then exit. This is typically done as a result of a failed system call and handled by calling the p4 procedure named p4_error that examines the return values from socket procedures, etc. Most of the time however, the procedures simply return a value. Some of the procedures return no value and thus are declared to return VOID. Some of the procedures return either a pointer to a character string or NULL; NULL indicates an error. The remaining procedures return an integer value; (-1) indicates an error.
In this section we describe the p4 functions needed for basic creation and termination of processes.
int p4_initenv(argc,argv) int *argc; char **argv;should be called by your program before an attempt is made to use any p4 procedures or data areas. We suggest making it the first executable statement in your program. p4_initenv parses the command line arguments and extracts the ones intended for p4 ignoring all others (see the discussion of command line arguments). Note that you pass the address of argc to p4_initenv so that it can actually remove its own arguments before your program looks at them.
int p4_create(fxn) int (*fxn)();There are two procedures that you can use to create processes in p4, p4_create_procgroup and p4_create. Processes created via p4_create are said to be ``user-managed'' whereas those created by p4_create_procgroup are ``p4-managed''. The p4-managed processes are automatically assigned unique id's (beginning with 0 for the big master), they have message queues allocated for them so that they can do message-passing, and they are able to run either on a shared-memory multiprocessor with the creating process or they can run on a separate machine. Processes created via p4_create do not have any of these advantages. They must develop their own id's, they cannot do message-passing, and they can only run on a shared-memory multiprocessor with the creating process. The only disadvantage of p4_create_procgroup is that you must build a procgroup file describing the set of required slave processes before the master program begins execution. This eliminates the possibility of determining late in the execution exactly how many processes you want to use to solve a problem. Generally, this is not a problem, especially since we can combine p4_create_procgroup and p4_create in the following way: You can use p4_create_procgroup to develop a network of processes that talk to each other via messages. Each of those processes can create further processes to help it out as necessary. The original set of processes communicate with their local slaves through shared data areas and with each other via message-passing.int p4_create_procgroup()
p4_create receives one argument that is a pointer to a function. It creates a single new process that executes the indicated function. The new process may share data areas (in shared memory) with the parent process. However, the new process is not managed by the p4 system in the sense that it is not assigned an id, it cannot pass messages, etc. The only p4 procedure that deals with user-managed slaves is p4_create. No other procedures are even aware of their existence.
p4_create_procgroup reads your procgroup file to determine the number of slave processes to create and where they are to be placed. It looks first for the file specified on the command line following the -p4pg argument if there is one. If there is no such argument, it looks for a file with the same name as the executable (master) file, with the .pg suffix. If it does not find one, it looks for a file named procgroup. It builds a procgroup table that describes all created processes and gives a copy of the table to each process. The processes then use the table to discover how to communicate with each other (processes in a cluster can send messages directly through shared memory or some other vendor-specific mechanism), others communicate via sockets). An alternative method is to build the table in memory yourself and use p4_startup.
The effect of p4_create_procgroup can be obtained in another way if a system would prefer to use its own way of specifying the locations of processes. A user may allocate the procgroup data structure and then fill it in ``by hand'' rather than by reading a file in p4 procgroup format. The following procedures support this method of starting processes.
struct p4_procgroup *p4_alloc_procgroup()allocates a procgroup data structure of the form described in p4.h. The formats of individual entries (p4_procgroup_entry) are given there as well.
int p4_startup(pg) struct p4_procgroup *pg;starts processes as specified by an an already-created procgroup data structure allocated by p4_alloc_procgroup and filled in by the user using the structures p4_procgroup_entry and p4_procgroup.
VOID p4_wait_for_end()is the p4 termination/cleanup procedure that you should invoke at the end of every execution of a program that uses p4. For the master process, it does some termination processing and then waits for slave processes to end. It should be called by all processes.
int p4_get_my_id()returns an integer value representing the id of the process assigned by the p4 system. If the process is not a p4-managed process, the value (-1) is returned.
int p4_num_total_ids()returns an integer value indicating the total number of ids started by p4 in all clusters, including the big master and all remote masters.
int p4_num_total_slaves()returns an integer value indicating the total number of processes started by p4 in all clusters, including all remote masters but not the big master.
The p4 system supports the cluster model of parallel computation, in which subsets of processes share memory with one another, with the clusters communicating via messages. A procgroup file for a program written for the cluster model might look like this:
local 4
alliant1.abc.edu 5 /home/me/myprog
alliant2.abc.edu 5 /home/me/myprog
encore.somewhere.edu 5 /usrs/me/myprog
This would specify a total of 20 processes, 5 (including the master) running
on the local machine (here assumed to be capable of supporting five processes
that share memory) together with 5 slaves each on three other shared-memory
machines. One process out of each set of remote slaves will be the ``remote
master'' for that cluster..
VOID p4_get_cluster_ids(start,end) int *start; int *end;receives pointers to two integers. It places the p4-assigned id's of the first and last id's within the current cluster into the two arguments (including the remote master).
int p4_get_my_cluster_id()returns a unique id (relative to 0) within a cluster of p4-managed processes. Thus, a cluster master will always have a cluster id of 0. It is not clear that a separate cluster id is really useful, but the functionality is provided just in case.
BOOL p4_am_i_cluster_master()returns a BOOL value indicating whether the invoking process is the ``cluster master'' process within its cluster.
int p4_num_cluster_ids()returns an integer value indicating the number of ids in the current cluster as started by p4_create_procgroup.
VOID p4_cluster_shmem_sync(cluster_shmem) VOID **cluster_shmem;This routine is used to synchronize the processes in a cluster before they begin to use shared memory.
VOID p4_get_cluster_masters(numids, ids) int *numids, ids[];This procedure fills in the values of numids and ids. It obtains the p4-id's of all ``cluster masters'' for the program, placing them in the ids array and placing the number of id's in numids.
P4 supports a set of send/receive procedures. These procedures are ``generic'' in the sense that they do not know whether a message must travel across a network or through shared memory, or via some other mechanism. They depend on a lower-level set of procedures that handle local or network (remote) communications. By default, the messages are assumed to be typed. If the user wishes to use untyped messages, he can hide the typing by coding some very simple C macros that always use a single message type.
p4_send(type,to,msg,len) p4_sendr(type,to,msg,len) p4_sendx(type,to,msg,len,datatype) p4_sendrx(type,to,msg,len,datatype) p4_sendb(type,to,msg,len) p4_sendbr(type,to,msg,len) p4_sendbx(type,to,msg,len,datatype) p4_sendbrx(type,to,msg,len,datatype)int type, to, len, datatype; char *msg;
Each of these procedures sends a message. The type argument is an integer value chosen by the user to represent a message type. The to argument is an integer value that specifies the p4-id of the process that should receive the message. The len argument contains the length in bytes of the message to be passed. Note that some of the procedures have a ``b'' in their name, e.g. p4_sendb. These procedures assume that the msg is in a buffer that the user obtained earlier via a p4_msg_alloc; otherwise, the buffer is assumed to be in the user's local space, and may cause the message to be copied internally. The procedures with an ``r'' in the name do not return until an acknowledgement is received from the to process (the ``r'' stands for rendezvous). Those procedures with an ``x'' in the name take an extra argument (datatype) that specifies the type of data in the message; these procedures will use that information to call XDR for data conversion if the message is being passed to a machine of a different architecture, i.e. where the internal representation may be different. The valid values for the datatype parameter are P4INT, P4DBL, P4FLT, P4LNG, and P4NOX. The last of these means ``no translation''.
BOOL p4_messages_available(req_type,req_from) int *req_type,*req_from;returns a BOOL value indicating whether the process has any messages available or not. The parameters req_type and req_from are both pointers to integers; they are used as both input and output arguments. On input, req_type has a value that indicates the type of message that the user wishes to check for availability (-1 indicates any type). The variable req_from is used similarly to indicate who a message is desired from.
int p4_recv(req_type,req_from,msg,len_rcvd) int *req_type,*req_from,*len_rcvd; char **msg;takes four arguments. The msg argument is a pointer to a pointer to a char. If this value is NULL, then p4 will allocate the buffer for the message according to its length. That is, one need not know ahead of time the length of a message being received. If this value is not NULL, then it points to a p4 message buffer that the user has obtained via p4_msg_alloc. The len_rcvd argument is a pointer to an integer that is assigned the length of the received message. Req_type and req_from are both pointers to integers; they are used as both input and arguments. On input, req_type has a value that indicates the type of message that the user wishes to receive (-1 indicates any type). It will block until a message of that type is available. Req_from is used similarly to indicate who a message is desired from. One important note about this procedure is that it obtains the area in which to place a message, and the user must explicitly free that area when finished with it (see p4_msg_free). There is an option available with p4_recv in which the user can provide his own buffer rather than having p4 allocate it. To do this, the user points msg to a buffer that he must obtain via a call to p4_msg_alloc (see below). No p4_msg_free should be performed if the same buffer is going to be re-used multiple times.
char *p4_msg_alloc(len) int len;
VOID p4_msg_free(m) char *m;
obtain and free a buffer area that can be used to receive a message. This procedure should be used for this task because a message has hidden information which the user is unaware of and therefore should not use malloc to obtain the area.
P4 supports a number of operations for dealing with all processes at once.
p4_broadcast(type, data, data_len) int type; char *data; int data_len;
p4_broadcastx(type, data, data_len, data_type) int type; char *data; int data_len, data_type;provide the ability to broadcast messages like p4_send and p4_sendx. These are semantically equivalent to a loop which uses p4_send or p4_sendx to individually send a message to each other process (the sender is not included.) Messages sent by one of these broadcasts are received by normal p4_recv's. The implementation of p4_broadcast is more efficient than such a loop, since it uses a ``broadcast tree''. One situation to look out for is a normal p4_broadcast followed by a p4_send. It is possible for the first message to arrive at its destination after the second one. The order of messages in this situation can be enforced with the use of the type argument.
p4_global_op(type,x,nelem,size,op,data_type) int type; char *x; int size, nelem; int (*op)(); int data_type;where op is one of:
p4_int_absmax_op() p4_int_absmin_op() p4_int_max_op() p4_int_min_op() p4_int_mult_op() p4_int_sum_op() p4_dbl_absmax_op() p4_dbl_absmin_op() p4_dbl_max_op() p4_dbl_min_op() p4_dbl_mult_op() p4_dbl_sum_op() p4_flt_absmax_op() p4_flt_absmin_op() p4_flt_max_op() p4_flt_min_op() p4_flt_mult_op() p4_flt_sum_op()and data_type is one of P4INT, P4LNG, P4FLT, or P4DBL.
This collection of routines provide the ability to do a variety of global operations. See the example program p4/messages/systest.c. They apply the commutative and associative operation op globally to x on an element-by-element basis and broadcast the result to all nodes. That is, each process ends up with
for (i=0; i<n; i++)
x[i] = x[node 0][i] op x[node 1][i] op x[node 2][i] op ...
op should be of the form
VOID op(char *x, char *y, int nelem)
{
data_type *a = (data_type *) x;
data_type *b = (data_type *) y;
while (nelem--)
*a++ operation= *b++;
}
where data_type and operation are chosen appropriately.
The order in which nodes apply the operation is undefined (hence op must be commutative and associative). The communication may be internally sub-blocked so the function op should not be hardwired to specific vector lengths.
This is still a relatively primitive version, which gathers the necessary data up a balanced binary tree and then uses p4_broadcast to send the results back. The type argument specifies the message type to be used in the communication associated with this global operation.
Strictly speaking, the size parameter, which is size in bytes of one element, is unnecessary. It is retained for backward compatibility.
VOID p4_global_barrier(type) int type;This procedure takes one argument which is the message type to be used for internal message-passing. It causes the invoking process to hang until all processes specified in the procgroup file have invoked the procedure.
Here is a simple example of a shared-memory program using monitors. In this program, each process retrieves values from a shared loop index. A monitor is used to ensure that all values are retrieved exactly once.
#include "p4.h"struct globmem { p4_getsub_monitor_t getsub; } *glob;
main(argc,argv) int argc; char **argv; { p4_initenv(&argc,argv);
glob = (struct globmem *) p4_shmalloc(sizeof(struct globmem)); p4_getsub_init(&(glob->getsub));
p4_create_procgroup(); worker(); p4_wait_for_end(); }
worker() { int i, nprocs;
nprocs = p4_num_total_ids(); i = 0; while (i >= 0) { p4_getsub(&(glob->getsub),&i,10,nprocs); p4_dprintf("I got %d\n",i); } }
The following functions are just basic memory management routines.
char *p4_malloc(n) int n;typically acts like the standard malloc, but may be rewritten for user systems that require different operation.
VOID p4_free(p) char *p;typically acts like the standard free, but may be rewritten for user systems that require different operation.
char *p4_shmalloc(n) int n;acts like the standard malloc except will obtain shared memory on machines that support sharing memory among processes. Compare with p4_malloc.
VOID p4_shfree(p) char *p;frees memory obtained with p4_shmalloc. Compare with p4_free.
The abstraction provided by p4 for managing data in shared memory is monitors. Good places to learn about the monitor concept in general are [(ref pbh:architecture)] and [(ref hoare:monitors)]. The specific approach taken by p4 is described in [(ref lusk-overbeek:p4-book)]. P4 provides several useful monitors (p4_barrier_t, p4_getsub_monitor_t, p4_askfor_monitor_t) as well as a general monitor type to help the user in constructing his own monitors (p4_monitor_t).
The following functions can be used to construct monitors. A monitor so constructed has the type p4_monitor_t.
int p4_moninit(m,i) p4_monitor_t *m; int i;initializes the monitor pointed to by m and gives it i queues for processes to wait on while they are blocked (see p4_mdelay). One queue is sufficient for most purposes. The queues are numbered beginning with 0.
VOID p4_menter(m) p4_monitor_t *m;enter the monitor pointed to by m. By the definition of a monitor, access is restricted to a single process in the monitor at a time (if everybody plays by the rules).
VOID p4_mexit(m) p4_monitor_t *m;exits the monitor pointed to by m. You are of course assumed to have previously entered that monitor.
VOID p4_mcontinue(m,i) p4_monitor_t *m; int i;checks to see if there are any processes blocked on the i-th queue of the monitor m and causes one of them to be released for entry to the monitor if so. If there are no such processes, the invoking process simply exits. Note that a process could have been blocked previously by invoking the procedure p4_mdelay. The queues are numbered beginning with 0.
VOID p4_mdelay(m,i) p4_monitor_t *m; int i;permits a process to delay itself on the i-th queue of monitor m if the process wishes to release the monitor, but wants to be waked up by another process later (via the procedure p4_mcontinue). The queues are numbered beginning with 0.
In this section we describe some of the specific monitors that are built into the p4 library. Each of them has its own pre-defined type, which can be used to allocate storage for them, which should be in shared memory. See the p4/monitors directory for examples. A lock is itself a monitor, with no extra delay queues.
VOID p4_lock_init(l) p4_lock_t *l;initializes the lock l. Must be used prior to any attempts to lock or unlock l.
VOID p4_lock(l) p4_lock_t *l;blocks if the lock l is already locked, otherwise locks l and proceeds.
VOID p4_unlock(l) p4_lock_t *l;unlocks the lock l.
VOID p4_getsub(gs,s,max,nprocs) p4_getsub_monitor_t *gs; int *s,max,nprocs;is a procedure used to obtain the next value of a shared counter (subscript). It takes as its first argument, a pointer to a getsub monitor that protects the shared counter. It assigns the current value of the counter to the integer that s points to, and then increments the counter by 1. p4_getsub_init initially sets the counter to 0. When the counter passes the value max, all nprocs processes are returned the value (-1) once, then the counter is reset to 0 for further use.
VOID p4_getsubs(gs,s,max,nprocs,stride) p4_getsub_monitor_t *gs; int *s,max,nprocs,stride;is like p4_getsub except that the counter is increased on each call by stride instead of 1.
int p4_getsub_init(gs) p4_getsub_monitor_t *gs;initializes the getsub monitor pointed to by gs; this initialization includes assigning a value of 0 to the counter that the monitor protects.
The standard barrier synchronization pattern is expressed as a monitor. There can be multiple barrier monitors, and one can wait for only some of the processes at the barrier if this is desired.
VOID p4_barrier(b,nprocs) p4_barrier_monitor_t *b; int nprocs;causes the executing process to hang until nprocs processes execute a barrier instruction with a pointer to the same barrier monitor b as an argument.
int p4_barrier_init(b) p4_barrier_monitor_t *b;initializes the barrier monitor b; this procedure should be invoked before you attempt to use the monitor in any operations.
Finally, the askfor monitor functions like a general dispatcher of work.
int p4_askfor(af,nprocs,getprob_fxn,problem,reset_fxn) p4_askfor_monitor_t *af; int nprocs; int (*getprob_fxn)(); VOID *problem; int (*reset_fxn)();requests a new ``problem'' to work on from the problem pool. The arguments are (1) a pointer to the askfor monitor that protects the problem pool, (2) the number of processes that call this procedure (with af) looking for work, (3) a pointer to the user-written procedure that obtains a problem from the pool, (4) a pointer that is filled in with the address of a user-defined representation of a problem to be solved, and (5) a pointer to a user-written procedure to reset when all problems in the pool are solved, in case the same monitor is re-used for another set of problems later. p4_askfor returns an integer indicating whether a problem was successfully obtained or not:
-1 : program is terminating (some process called p4_progend) 0 : a problem was obtained and ``problem'' points to it 1 : problem solved by exhaustion, i.e. no more problems to get n > 1 : a process found a solution and called p4_probend with code nFor a detailed discussion of the ``askfor'' monitor, see [(ref lusk-overbeek:p4-book)].
int p4_update(af,putprob_fxn,problem) p4_askfor_monitor_t *af; int (*putprob_fxn)(); VOID *problem;updates the problem pool being managed by the askfor monitor. The arguments are (1) a pointer to the askfor monitor that protects the problem pool, (2) a pointer to the user-written procedure that puts problems into the pool, and (3) a pointer to a user-defined representation of a problem to be put in the pool. Putprob_fxn should return 1 if it did indeed put a new problem into the pool, so that any delayed processes should wake up and re-examine the pool (this logic is handled by the p4_askfor) and 0 if upon entering the monitor and examining its potential problem together with the data there it decided not to add a new problem to the pool. It can be assumed that the ``putprob'' logic (defined by putprob_fxn) is executed inside the monitor.
int p4_askfor_init(af) p4_askfor_monitor_t *af;initializes the askfor monitor af; this procedure should be invoked before you attempt to use the monitor in any operations.
VOID p4_probend(af,code) p4_askfor_monitor_t *af; int code;allows the user process to mark a problem as solved early when several processes are coordinating their activities via an askfor monitor. The code is an integer value that will be returned to all processes when they ``askfor'' a new sub-problem to work on.
VOID p4_progend(af) p4_askfor_monitor_t *af;allows a process to cause a return code of (-1) to be returned to all processes using an askfor monitor. This would typically be called by a master process to indicate that no more problems are to be solved and that all slave processes should terminate.
A small number of simple functions are available for accessing various clocks and timers.
int p4_clock()returns a value in milliseconds. This is a wall-clock value, usually obtained from the system via gettimeofday. Also see p4_ustimer below.
p4_usc_time_t p4_ustimer()returns a wall-clock time value in microseconds. The precision of this number depends on the timer installed on the individual machine. In some cases the resolution may be no greater than that of p4_clock(). For arithmetic and printing purposes, the type p4_usc_time_t is an unsigned long integer.
p4_usc_time_t p4_usrollover()returns the timer value at which a microsecond timer ``rolls over''. Since p4_usc_time_t is a long integer's worth of microseconds, it is likely that the timer will roll over (become zero) during even medium-length runs (about 72 minutes on most machines).
P4 has a set of routines to aid in producing a printed trace of events, both user-defined and pre-defined in the p4 system.
VOID p4_dprintf(fmt, va_alist) char *fmt; va_dclacts just like the standard printf except that the print line is preceded by a value that identifies the process. This value is typically the string pn_u where n represents the p4-assigned id and u represents the unix-id of the process on its host. However, there are other forms of this value. For example, the big master is represented as bm_u. Also, if a process prints before it has a p4-assigned id, then its value will be something like bm_slave_n_u or rm_slave_n_u. Typically, it is not possible for a user program to print anything before being assigned an id by p4, but the p4 system itself may use this procedure to print messages from a particular process if it encounters problems getting the process initialized.
VOID p4_dprintfl(level, fmt, va_alist) int level; char *fmt; va_dclis like p4_dprintf except that the first argument is an integer indicating the debugging level that must be in effect before this message will print. A level of 0 will cause the message to always print. If you run a program with the debug level set to 5 (via command-line arguments), then all dprintfl's with level less than or equal to that debug level will print. See section Command-Line Arguments for how to set the debug level at run time.
The debug level can be examined and changed by the user during execution:
int p4_get_dbg_level()returns the current debug level for this process and its cluster.
VOID p4_set_dbg_level(level) int level;sets the current debug level for this process and its cluster. P4 itself is liberally instrumented with p4_dprintfl's of level 10 and above, leaving levels 0-9 for the user. The greater the debug level of the built-in messages, the greater understanding of p4 needed by the user to make sense of them. However, levels as high as 30 may well be useful to the user trying to debug a p4 program. Roughly speaking, the following debug levels produce messages about the indicated events.
level 10: created process
sent message
received message
level 20: creating process
sending message
receiving message
process starting
process exiting
level 30: waiting for ack
sending ack
sent ack
received ack
queueing message for later receipt
queued message for later receipt
level 40: memory management
buffer management
level 50: reading procgroup
other initialization message exchange
level 60: send-receive details, especially machine-specific traces
level 70: listener interactions:
creating listener
created listener
messages from inside listener
level 80: detailed data structures after initialization
level 90: detailed tracing of flow thru procedures
For optimum performance, the test of the debug level required by these
messages can be removed at compile time by not commenting out the
#define P4_DPRINTFL line in the OPTIONS
file (See section Introduction).
The following function is provided to deal with abnormal termination. It can be called by any process.
VOID p4_error(string, value) char *string; int value;prints string as an error message and then forcefully terminates all co-operating processes and cleans up all shared resources.
VOID p4_soft_errors(onoff) int onoff;enables/disables soft errors, returning the previous setting. The default is ``disabled'', which means that certain p4 functions will call p4_error instead of returning -1.
p4_error gets control on certain kinds of interrupts. It is automatically called for SIGSEGV, SIGBUS, and SIGFPE interrupts, to catch user programming errors and clean up, after which it returns interrupt handling to default mode and returns, so that the user may obtain a dump. It also handles SIGINT interrupts, in which case it cleans up and exits. Finally, it may be called directly by the user, in which case it cleans up (other p4 processes and IPC's) and exits.
Although p4_error is supposed to get rid of all running p4 processes, it can happen that an error is bad enough that p4 processes are left running. A primitive aid in finding and killing these processes is the shell script kj, which takes a string as an argument and then kills processes containing that string as part of their program names. Currently it only kills processes on the machine where it is run, but it can be run via rsh on remote machines. There are other useful scripts (e.g. killipc and killp4) in the p4/bin directory to do such things as clean up SYSV IPC items that may be left when a program abnormally terminates. P4 will generally cleanup these items if the abnormal termination is a type that p4 traps, otherwise the user must do the cleanup. This is an unfortunate side-effect of the way that SYSV handles things, it really should be the OS's function to take care of this.
On many machines it is possible to attach a debugger like dbx to a running process. This is one way to find out where a hanging process is stuck.
In this section are found functions that do seem to fit neatly into any of the other sections.
char *p4_version()returns a string containing the version number of p4 being run.
VOID p4_print_avail_buffs()P4 maintains an array of buffer lists of various sizes, so that it can very rapidly allocate and deallocate buffers. You can see the contents of the buffer pools at any time by calling this procedure.
VOID p4_set_avail_buff(bufidx,size) int bufidx; int size;This procedure is used to set the size of buffers in p4's buffer pools. The parameter bufidx specifies a particular buffer list, and should be a number from 0 to 7. The size parameter specifies that buffers up to that size will be managed by p4 in a particular list. It is important to maintain the buffer sizes in increasing order. The default list of buffer sizes is 64, 256, 1024, 4096, 16384, 65536, 262144, 1048576. This causes wasted space if you send only one large message, causing the allocation of a large buffer which is not reused. Savings in space can be achieved by adjusting these numbers to correspond with the message sizes of your application. If no large messages are sent at all, however, no space is wasted since the large buffers will never be allocated. If you send a message larger that the largest size in this array, p4 will allocate the buffer, and then free it back to the system as soon as it can.
In this section we describe the p4 Fortran library. All Fortran programs must include the file p4f.h from the directory lib_f. The Fortran calls to p4 procedures are analogous to their C counterparts, but have Fortran-like names. You might find the documentation for the corresponding C routine, in one of the sections above, helpful.
p4init()should be called by your program before an attempt is made to use any p4 procedures or data areas. We suggest making it the first executable statement in your program.
p4crpg()This routine should be called by the master process (the one started directly by you) to read the procgroup file and start the processes specified there. It can be called by other process, but has no effect in that case.
integer p4myid()returns an integer value representing the id of the process assigned by the p4 system.
p4cleanup()should be called by the master process to wait for the termination of the processes created by p4crpg.
p4send(type,dest,msg,len,rc) integer type, dest, len, rc real msg
p4sendx(type,dest,msg,len,data_type,rc) integer type, dest, len, data_type, rc real msg
p4sendr(type,dest,msg,len,rc) integer type, dest, len, rc real msg
p4sendrx(type,dest,msg,len,data_type,rc) integer type, dest, len, data_type, rc real msgEach of these procedures sends a message. The type argument is an integer value chosen by the user to represent a message type. The dest argument is an integer value that specifies the p4-id of the process that should receive the message. The len argument contains the length in bytes of the message to be sent. The procedures with an ``r'' in the name do not return until an acknowledgement is received from the to process (the ``r'' stands for rendezvous). Those procedures with an ``x'' in the name take an extra argument (datatype) that specifies the type of data in the message; these procedures will use that information to call XDR for data conversion if the message is being passed to a machine of a different architecture, i.e. where the internal representation may be different. p4 maintains an internal table of which pairs of machine types require conversion, so it only does the conversion when it is necessary. The valid values for the data_type parameter are P4INT, P4DBL, P4FLT, P4LNG, and P4NOX. The last of these means ``no translation''.
p4recv(type,from,buf,buflen,msglen,rc) integer type, from, buflen, msglen, rc real bufThe buf parameter is the buffer into which the message is to be received. It can be of any Fortran type. The buflen parameter specifies its length, so that p4 can check for overruns. The number of bytes actually received is given by msglen. The type and from parameters specify the message type and the source of the message. If either of these is set to -1, then screening is not applied, and the parameter is set to indicate the type and/or source of the message actually received. rc is the return code from the call.
p4probe(type,from,rc)sets rc to 1 or 0 depending on whether the process has any messages available or not. The parameters type and from are used as both input and arguments. On input, type has a value that indicates the type of message that the user wishes to check for availability (-1 indicates any type). The variable from is used similarly to indicate who a message is desired from.
p4brdcst(type,data,len,rc) integer type, len, rc real data
p4brdcstx(type,data,len,data_type,rc) integer type, len, data_type, rc real dataprovide the ability to broadcast messages like p4send and p4sendx. These are semantically equivalent to a loop which uses p4send or p4sendx to individually send a message to each other process (the sender is not included.) Messages sent by one of these broadcasts are received by normal p4recv's. The implementation of p4brdcst is more efficient than such a loop, since it uses a ``broadcast tree''.
integer p4ntotids()returns an integer value indicating the total number of processes started by the p4 master process and all remote processes. It includes the master process itself.
integer p4nslaves()returns an integer value indicating the total number of processes started by p4, not including the original master process.
integer p4nclids()returns an integer value indicating the number of ids in the current cluster as started by p4crpg.
integer p4myclid()returns a unique id (relative to 0) within a cluster of p4-managed processes. Thus, a cluster master will always have a cluster id of 0.
p4globarr(type) integer typetakes one argument which is the message type to be used for internal message-passing. It causes the invoking process to wait until all processes specified in the procgroup file have invoked the procedure.
p4getclmasts(numids,ids) integer numids, ids(*)This procedure fills in the values of numids and ids. It obtains the p4-ids of all ``cluster masters'' for the program, placing them in the ids array and placing the number of ids in numids.
p4getclids(start,end) integer start, endreceives two integers. It places the p4-assigned id's of the first and last ids within the current cluster into the two arguments (including the remote master).
integer p4clock()returns a value in milliseconds. This is a wall-clock value, usually obtained from the system via gettimeofday. Also see p4ustimer below.
integer p4ustimer()returns a wall-clock time value in microseconds. The precision of this number depends on the timer installed on the individual machine. In some cases the resolution may be no greater than that of p4clock().
p4flush()flushes standard out. On some systems this needs to be done explicitly for prompts. This is just a convenience routine that has nothing to do with p4.
p4error(str,val) character*n str integer valprints string as an error message and then forcefully terminates all p4 processes.
p4softerrs(new,old) integer new, oldenables/disables soft errors, returning the previous setting in old. The default is ``disabled'', which means that certain p4 functions will call p4_error instead of returning -1.
integer p4version()returns a string containing the version number of p4 being run.
p4avlbufs()P4 maintains an array of buffer lists of various sizes, so that it can very rapidly allocate and deallocate buffers. You can see the contents of the buffer pools at any time by calling this procedure.
p4setavlbuf(idx,size) integer idx, sizeThis procedure is used to set the size of buffers in p4's buffer pools. The parameter bufidx specifies a particular buffer list, and should be a number from 0 to 7. The size parameter specifies that buffers up to that size will be managed by p4 in a particular list. It is important to maintain the buffer sizes in increasing order. The default list of buffer sizes is 64, 256, 1024, 4096, 16384, 65536, 262144, 1048576. This causes wasted space if you send only one large message, causing the allocation of a large buffer which is not reused. Savings in space can be achieved by adjusting these numbers to correspond with the message sizes of your application. If no large messages are sent at all, however, no space is wasted since the large buffers will never be allocated. If you send a message larger that the largest size in this array, p4 will allocate the buffer, and then free it back to the system as soon as it can.
p4globop(type,x,nelem,size,op,data_type,rc)where op is one of:
p4intsumop p4intabsmaxop p4intabsminop p4intmaxop p4intminop p4intmultop p4dblsumop p4dblabsmaxop p4dblabsminop p4dblmaxop p4dblminop p4dblmultop p4fltsumop p4fltabsmaxop p4fltabsminop p4fltmaxop p4fltminop p4fltmultopThe data_type parameter in the above operations should be one of
P4INT P4LNG P4FLT P4DBLThese symbolic constants are defined in the include file p4f.h. The size argument is the size in bytes of one element, and the type argument is the message type used in the global operation.
There are also Fortran routines for creating logfiles See section Creating Log Files in Fortran.
P4 processes on remote machines are ordinarily created by rsh. For this to work, the user must have permission to create processes on that machine. This permission is normally granted either globally by the system administrator, or locally by the use of .rhosts files. (See the normal unix man pages under rhosts).
Since rsh is relatively slow, p4 provides a way to get things started faster. This is accomplished by running the program serv_p4 in the background on the remote machine. When p4 is creating processes, it will automatically check for the existence of this server and use it if it is running. Remote processes typically start much faster when the server is running. When p4 uses rsh, the remote process's stdout is sent back to the stdout of the parent (the p4 master process). We have not yet tested this server on all of the machines that we support. Thus far, we have tested it somewhat on the SYMMETRY, SUN, DEC5000, and SGI. We believe that it will work on many other machines, but have not yet verified it on all machines.
An invocation of a set of servers is (currently) associated with a specific port number. This way multiple users can each be running multiple server networks without mutual interference, provided each network of servers is started with a different port number.
To start the secure server on a machine one can do
serv_p4 -d -p <num>where <num> is a port number to be associated with a network of servers. If the -p option is omitted, the server will pick an unused port number and report
Listening on <num>.Then p4 programs to use this network should be started with
-p4ssport <num>The p4 application must also be listed in the user's .p4apps file in his home directory. This file should be readable only by the user, and should contain the full path names of programs that the user wishes to be startable by the p4 server.
When a p4 master process tries to start a slave process on a remote machine, it will first attempt to do it via the server. If it cannot do so for any reason (no server running, port number mismatch, or program not found in .p4apps file), then it tries to do so with the remote shell command.
Note that the server is used only to start processes; it plays no role in a p4 computation once the slave processes have been initiated. Rather, a temporary process, called the listener, is spawned to manage connection requests that occur during the execution of a p4 program. Neither the server nor the listener consumes any significant amount of CPU time.
There is further discussion of installation options for the servers in the README file in the p4/servers subdirectory.
A number of useful utilities can be found in the bin subdirectory. These can be used to start and stop server processes based on the contents of a file of machines one regularly uses, to kill runaway p4 processes in the unlikely case that they cannot or do not terminate automatically when one processes ends abnormally or is interrupted from the keyboard, and to merge logfiles created for the use of upshot (See section Creating Logfiles for Upshot). Some of these scripts may have to be edited to reflect the installation directory of p4.
P4 is distributed with a set of routines for creating logfiles (see README in the p4/alog directory. The resulting logfiles can be examined by upshot, distributed separately. For details about upshot, see [(ref herrarte-lusk:upshot)].
The p4/alog directory contains a package (ALOG) for creating logs of time-stamped events, that is of general utility, outside of p4. The timestamps are obtained from various microsecond-level resolution timers on various machines. The portable microsecond timing package is contained in the usc subdirectory. It is used by the ALOG package as well as by the p4_ustimer function in p4. Similarly, the ALOG package can be used independently of p4 and upshot. Its logfiles were designed to be read and displayed by upshot, but other display packages can be used as well.
The ALOG package consists of a set of macros that can be used to instrument a C program and a set of functions that can be used to instrument a Fortran program. We will focus here primarily on the use of the C interface, which contains more functionality.
The macros that can be used to instrument a program are as follows (from the file README_ALOG in the alog directory):
ALOG_SETUP(pid,flag): pid - (integer) process id of callee flag - (integer) either ALOG_WRAP or ALOG_TRUNCATEThis macro initializes the tracing area for a slave process and must be called once before any event is logged. If the value of flag is set to ALOG_WRAP, then in the event of no more space for logging events the system will only report the latest n events. If flag is set to ALOG_TRUNCATE the system will stop logging events as soon as there is no more memory for the events to be logged.
ALOG_MASTER(pid,flag): pid - (integer) process id of the callee flag - (integer) either ALOG_WRAP or ALOG_TRUNCATEThis macro has the same effect over its parameters as ALOG_SETUP with the difference that this macro should be referenced by the master process only.
ALOG_DEFINE(event,strdef,format): event - (integer) id of event being defined strdef - (string) description of 'event' format - (string) control string in "printf" formatThis macro puts an event definition code into the logfile.
ALOG_LOG(pid,event,intdata,strdata): pid - (integer) process id of callee event - (integer) event id to be logged intdata - (integer) any integer data for this event strdata - (string) any string data (can be the null string)This macro provides the event logging service.
ALOG_OUTPUT no parametersThis macro dumps the events logged into a log file with the name alogfile.pxx where xx is the logical PID of the callee process. The log file is created in the current directory unless specified otherwise through the macro ALOG_SETDIR. All processes should execute this.
ALOG_SETDIR(dir) dir - (string) directory where log file is createdThis macro sets the output directory for the log file. The default directory for the creation of the log file is the current directory of the process. If used, then this macro MUST be invoked before ALOG_MASTER/ALOG_SETUP.
ALOG_STATUS(status): status - (integer) either ALOG_ON or ALOG_OFF
This macro controls the logging status of ALOG as follows. Setting status to ALOG_ON enables logging until it is turned off. Setting status to ALOG_OFF disables logging until it is turned on again. Logging is enabled at the outset by default.
ALOG_ENABLE no parameters
This macro enables event logging; same as calling ALOG_STATUS(ALOG_ON).
ALOG_DISABLE no parametersThis macro disables event logging; same as calling ALOG_STATUS(ALOG_OFF).
The sample program gridlog.shmem.c in the monitors subdirectory contains an example of a program instrumented with ALOG statements. The macro definitions for ALOG are included when you include #include "p4.h" in your program. If the line #define ALOG_TRACE is not included before the #include "p4.h", these macros will generate no code. Thus it is easy to effectively de-instrument the code by recompiling, and there is no need to protect each ALOG statement with an #ifdef.
When an ALOG-instrumented program is run, it will produce one logfile for each process. The files will be named alogfile.p0, alogfile.p1, .... These files need to be merged into a single file with the events sored by timestamp. This is accomplished with the program mergelogs, found in the bin subdirectory. To merge the logfiles, do
mergelogs alogfile.p* > myprog.log rm alogfile.p*The resulting logfile can be examined by upshot or some other logfile examination facility. See [(ref herrarte-lusk:upshot)] for details of the logfile format.
On networks of workstations and some distributed memory machisnes, the microsecond timers on the various processors are synchronized. To produce a usable merged logfile, the adjlogs program, also found in the bin directory, can be used to adjust the timestamps for offset and drift before they are merged. For this to work, synchronization events must be placed in the logfiles by an ALOG_LOG statement. The event type is then passed to adjlogs, which aligns the timestamps, based on the timestamps of the synchonization events. The call to adjlogs looks like this, where <n> is the type of the synchronization event. This program makes us of high-precision numeric libraries, and has been tested only on Sun's.
adjlogs -e <n>Both mergelogs and adjlogs are less portable than the other p4 code; you might want to run them on a workstation such as a Sun.
Log files can also be created by Fortran programs. The routines to do so are:
alogfsetup(pid,flag): pid - (integer) process id of callee flag - (integer) either ALOG_WRAP or ALOG_TRUNCATEThis function initializes the tracing area for a slave process and must be called once before any event is logged. If the value of flag is set to ALOG_WRAP, then in the event of no more space for logging events the system will only report the latest n events. If flag is set to ALOG_TRUNCATE the system will stop logging events as soon as there is no more memory for the events to be logged.
alogfmaster(pid,flag): pid - (integer) process id of the callee flag - (integer) either 0 or 1 (see above)This function has the same effect over its parameters as alogfsetup with the difference that this function should be referenced by the master process only.
alogfdefine(event,strdef,format): event - (integer) id of event being defined strdef - (string) description of 'event' format - (string) control string in "printf" formatThis function puts an event definition code into the logfile.
alogflog(pid,event,intdata,strdata): pid - (integer) process id of callee event - (integer) event id to be logged intdata - (integer) any integer data for this event strdata - (string) any string data (can be the null string)This function provides the event logging service.
alogfoutput()This function dumps the events logged into a log file with the name alogfile.pxx where xx is the logical PID of the callee process. The log file is created in the current directory unless specified otherwise through the function alogfsetdir.
alogfsetdir(dir) dir - (string) directory where log file is createdThis function sets the output directory for the log file. The default directory for the creation of the log file is the current directory of the process. If used, then this function MUST be invoked before alogfmaster/alogfsetup.
alogfstatus(status): status - (integer) either ALOG_ON or ALOG_OFF
This function controls the logging status of ALOG as follows. Setting status to ALOG_ON enables logging until it is turned off. Setting status to ALOG_OFF disables logging until it is turned on again. Logging is enabled at the outset by default.
alogfenable()
This function enables event logging; same as calling alogfstatus(ALOG_ON). It must be called first, even before alogfmaster or alogfsetup.
alogfdisable()This function disables event logging; same as calling alogfstatus(ALOG_OFF).
The sample program sr_log.f in the messages_f subdirectory contains an example of a Fortran program instrumented with logging statements.
Upshot is not part of the p4 distribution, but can be obtained from the same anonymous ftp location as p4. Take the file upshot.tar.Z from the directory pub/p4 on info.mcs.anl.gov. The distribution contains all necessary documentation on how to install and run upshot. It is an X-window program that runs on most workstations. There is no need for a parallel macchine to be involved, once the log files have been obtained.
Upshot produces the most interesting displays when certain events (not necessarily all) are defined to be the entry and exit events for certain states and then colors are associated with the states. This association is reflected in a statefile with a format like the following:
1 1 2 red asking 2 3 4 blue working 3 5 6 green updatingThis statefile describes three states. State 1 is defined to be between events 1 and 2. Upshot will color it red and label it ``asking''.
We have found that the most useful events to log and study are those identified by the user and specified in his program. That way he can control the number of events to be logged and the grain size of the states that are represented.
In some cases, however, one wants to study the details of the internal operation of a p4 application, or get some idea of the behavior on one's program without going to the trouble of instrumenting it himself. To this end, p4 itself is instrumented with ALOG statements, although by default they are inactive. To get automatic logging of p4 events (including sending and receiving of each message) one needs first to link to a version of the p4 library that has been compiled with the line #define ALOG_TRACE uncommented out in the OPTIONS file, and secondly, to run with -p4log on the command line.
Some important things to know about using the internal logging features of p4 are:
<progname> -pn .compute -sz <nprocs>
where nprocs is the total number of processes. On some Paragons,
one must first create a partition with
the mkpart command (mkpart -sz <size> <partname>),
specifying the size and name of the partition. At some installations, these
partitions are specicified ahead of time. (The lspart command says
which partitions currently exist, and the rmpart command is used to
remove a user-allocated partition. Once you have a partition, start your
program with:
myprog -pn <partition name>The procgroup file should contain one line:
<machinename> <nprocs-1> <pathname>
spnode1 0 <file>for the master. To use the socket interface to the switch on the SP-1, use swnodes for the nodes, including swnode1 for the master.
setenv MP_PROCS N (where N is the number of processes you want)ormyprog
myprog -procs NThe procgroup file should look like
local 15 <progname>to run with 16 processes (15 slaves). The program name in the procgroup file should be the full path name of the executable file. Such a program may be be interactive, but there currently are problems flushing lines to the terminal that do not end in a newline character (typically, interactive prompts).
The procgroup file should look like/usr/lpp/euih/eui/cotb0 -b <progname> <numprocs> <user args>
local 15 <progname>to run with 16 processes (15 slaves). The program name in the procgroup file should be the full path name of the executable file. Such a program should not be interactive.
rinfo -aand then start the job with
prun -n <nprocs> -lsv [-p <partition name>] <progname>The -lsv option causes stdout to be routed to files.
SUN
(1) P4 can be installed on this machine with or without SYSV IPC.
SUN_SOLARIS
(1) P4 is made with shared memory by default, and instead of SysV shared
memory, uses the more efficient mmap.
HP
(1) P4 can be installed on this machine with or without SYSV IPC.
(2) Fortran not tested (not avail on our test machine).
DEC5000
(1) P4 can be installed on this machine with or without SYSV IPC.
ALPHA
(1) P4 can be installed on this machine with or without SYSV IPC.
RS6000
(1) P4 can be installed on this machine with or without SYSV IPC.
(2) It is important to use the option -lbsd on the link step to
get sockets to support the NONBLOCKING option.
IBM3090
(1) P4 can be installed on this machine with or without SYSV IPC.
(2) Fortran not supported due to absence of iargc/getarg.
(3) There are multiply defined macros in include/rpc/rpc.h. IBM
is fixing this in a later OS release. Meanwhile make your own
copy and the fix the problem yourself.
TITAN
(1) P4 can be installed on this machine with or without SYSV IPC.
(2) Fortran not supported due to problems with getting args.
SGI
(1) P4 can be installed on this machine with or without SYSV IPC.
NEXT
(1) Fortran not supported due to absence of iargc/getarg.
FX2800/FX2800_SWITCH
(1) Alliant's switch code not yet ensuring messages
remain ordered. p4 currently discovers the switch port for
the machine it is running on by invoking the internal
procedure getswport. This procedure must be customized to
the installation. Alliant's switch is currently unsupported.
FX8
(1) You might need to add MFLAGS = -i to the Makefile
KSR
(1) The latest version of the OS produces a link-time error for
Fortran programs.
(2) Use of sockets fails because of a bug in socketpair.
IPSC860
(1) the script ``runcube'' (in the messages directory) may
be useful
DELTA
(1) the script ``rundelta'' (in the messages directory) may
be useful
BALANCE
(1) Fortran not supported.
SYMMETRY/SYMMETRY_PTX
(1) -Z compiler option may be changed to control the
shmalloc/malloc split. This is often needed when
creating logfiles on a symmetry.
(2) shared memory message passing not supported in Fortran
TC_2000/TC_2000_TCMP
(1) TCMP port not yet complete.
(2) For shared-memory execution, one must use cluster ... to
obtain a private cluster for execution
NCUBE
(1) Messages are limited to 32K in length.
CM5
(1) Logfiles are not supported.
SP1
(1) Using ``spnoden'' for node names causes p4 to use the TCP
interface to the Ethernet.
(2) Using ``swnoden'' for node names causes p4 to use the TCP
interface to the switch. In this case, replace the line
``local 0'' with ``swnode1 0'' in the procgroup file.
(3) It is important to use the option -lbsd on the link step to
get sockets to support the NONBLOCKING option when using the
TCP interface to either the switch or the Ethernet.
(4) EUIH programs may not be able to read from the keyboard.
Our attempt with this manual has been to prevent you from having difficulties. Experience shows that certain common problems recur, however. In this section we hope to address some of these problems.
rsh target.machine date
If you still get the ``Permission denied.'' message, then the problem has
nothing to do with p4. See hosts.equiv or .rhosts in the
system man pages.
local 0
machine1 1 pathname
machine2 1 pathname
machine3 1 pathname
.
.
The ``local'' means ``only the master on the startup machine; no local
slaves sharing memory''.
It is possible, at some cost in message-passing efficiency, to have a cluster of processes sharing memory on a workstation, but in this case p4 must have been installed with the SYSV_IPC option set in the OPTIONS file. The cost is that a process waiting for a message must spin between checking for a message arriving on a socket and a message arriving through shared memory.
This copy was produced on Wed Apr 6 14:29:28 1994.