This chapter may be safely skipped on a first reading. If you still want to read this chapter, it should mean that you are familiar with the basics of the TOP-C model, and are looking for advice on how to use the model more effectively. The first piece of advice is that the choice of task and shared data interact strongly with the choice of parallel algorithm. We review those concepts more precisely here, in light of the overall context of the TOP-C model.
()application routine. In the TOP-C model, this rule is bent to accommodate caching of private data to efficiently handle a
REDO_ACTION(see Section Caching slave task outputs (ParSemiEchelonMat revisited)), or to accommodate a
CONTINUATION_ACTION()(see Section The GOTO statement of the TOP-C model).
MasterSlave(). The shared data is never explicitly declared. However, it is best for the application programmer to include a comment specifying the shared data for his or her application. The TOP-C model poses certain restrictions on what legally constitutes shared data. The shared data must include enough of the global data (variables that occur free in the
()procedure) so that the task output of
()is uniquely determined by the task input and the shared data. However, the shared data must not include any variables whose values are modified outside of the application routine
(). Also, the shared data is updated non-preemptively, in the sense that a slave process will always complete its current task before reading a newly arrived message that invokes
(). If a slave privately caches data for purposes of a
CONTINUATION_ACTION(), such data is explicitly not part of the shared data.
In testing a program using
MasterSlave(), a hierarchy of testing is
suggested. The principle is to test the simplest example first, and then
iterate to more complex examples. When a stable portion of the program is
ready for testing, the following sequence of tests is suggested:
SeqMasterSlave()(see definition below) and see if the program performs correctly.
SeqMasterSlave()will run only on the master, without sending any messages, and so the full range of sequential debugging tools is available.
MasterSlave()and set up the
procgroupfile to have only one slave process (one line,
local 0, and one line
localhost ...). Initially test with no
MasterSlave(), and then test with the full set of parameters.
A second easy testing strategy is to set
true. (This is
the default value.) This causes all taskInputs, taskOutputs, and
non-trivial actions (actions other than
NO_ACTION) to be displayed at
the terminal. The information is printed in the same sequence as seen by
the master process.
Another ``cheap'' debugging trick is to inspect the values of global
variables on the slave after it has been thrown out of the
MasterSlave() procedure. The following code demonstrates by
interrogating the sum of the variables
y on slave number 2.
gap> SendRecvMsg("x+y;\n", 2);
This is useful to inspect cached data on a slave used for a
CONTINUATION_ACTION(). It may also be useful to verify if the
shared data on the slave is the same as on the master. If the slave
process is still inside the procedure
MasterSlave(), then from within a
break loop on the master, you may also want to interactively call
) to determine if the expected taskOutput is
If the master process is still within
MasterSlave(), then it is useful
() locally on the master process, and debug this
There is also the time-honored practice of inserting print statements.
Print statements ``work'' both on the master and on the slaves. If
ParTrace produces too much output, or not the right kind of information,
one can add print statements exactly where one needs them. As with any
UNIX debugging, it is sometimes useful to include a call to
fflush(stdout) to force any pending output. ParGAP binds this to:
This has the same effect as the UNIX
fflush(stdout). There may be
pending output in a buffer, that UNIX delays printing for efficiency.
Printing any remaining output in the buffer is forced by this command. A
common sequence is:
Print("information"); UNIX_FflushStdout();. Note
also that when the slave prints, there are ``two'' standard outputs
involved. You may also want to include a call to
the master to force any pending output that originated on a slave.
Finally, you should be conscious of network delays, and so a print
statement in a slave process will typically take longer to appear than a
print statement in the master process.
]]] ) F
If a bug is exhibited even in the context of a single slave, then the
code is ``almost'' sequential. In this case, one can test further by
replacing the call to
MasterSlave() by a call to
and debug in a context that involves zero messages and no interaction
with any slave. It can also be helpful to carry out initial debugging in
this context. Note that in the case of a single slave, which is what
IsUpToDate() will always return true, and so
most applications will not call for a
There are two common reasons for loss of efficiency in parallel applications. One is a lack of enough tasks, so that some slaves are starverd for work while waiting for the next task input. A second reson is that the ratio of communication time to compilation time is too large. The second case, poor communication efficiency, is the more common one.
The communication efficiency can be formally defined as the ratio of the
time to execute a task by the time taken for the master to send an
initial task message to a slave plus the time for the slave to send back
a result message. A good way to diagnose your efficiency is to execute
MasterSlaveStats() after executing
This function currently returns statistics in the form of a list of two records. The first record provides the global information:
UPDATEwas returned, and
The second record provides per-slave information:
QuoInt(1000*total,num)in GAP, and
Note that for purposes of the per-slave statistics, separate time
intervals are recorded for each initial task,
REDO action, and
CONTINUATION() action. The time for
UpdateSharedData() is not
included in these statistics. This is because after an
the slave does not reply to the master to acknowledge when the update was
Poor communication efficiency is typically caused either by too small a task execution time (which would be the case in the example of section or too large a message (in which case the communication time is too long). We first consider execution times that are too small.
On many Ethernet installations, the communication time is about 0.01
seconds to send and receive small messages (less than 1 Kb). Hence the
task should be adjusted to consume at least this much CPU time. If the
naturally defined task requires less than 0.01 seconds, the user can
often group together several consecutive tasks, and send them as a single
larger task. For example, in the factorization problem of section, one
() to test the next 1000 numbers as factors and
() to increment
counter by 1000.
There is another easy trick that often improves communication efficiency. This is to set up more than one slave process on each processor. This improves the communication efficiency because during much of the typical 0.01 seconds of communication time the CPU has off-loaded the job onto a coprocessor. Hence, having a second slave process running its own task on the CPU while a first process is concerned with communication allows one to overlap communication with computation.
We next consider the case of messages that are too large. In this case, it is important to structure the problem appropriately. The task architecture is intended to be especially adaptable to this case. The philosophy is to minimize communication time by duplicating much of the execution time on each processor.
After the initial data structure has been built, it will usually be
modified as a result of the computation. In order to again minimize
communication, the result of a task, which is typically passed to
(), should consist of the minimum information needed
to update the global data structure. Each process can then perform this
update in parallel.
Any long-running computation must be concerned with checkpointing. The
TOP-C model also provides a simple model for checkpointing. The key
observation is that the master process always has the latest state of the
computation, and the information in the master process is sufficient to
reconstruct any ongoing computation. Any application may take advantage
of this by checkpointing the necessary information either in the
() or in
A simple way to checkpoint is to record:
This model for checkpointing assumes that your program has no
CONTINUATION() actions. If you use
CONTINUATION() actions, then you
may require a more complex model for checkpointing.
This function returns a GAP list (with holes) of all pending task
inputs. If slave i is currently working on a task, index i of the
list will record that task. If slave i is currently idle or executing
(), then there will be a hole at index i. This
function is available for use within either the application routine
(), as specified in the
parameters of your call to
MasterSlave(). (Of course, your application
may be using a name other than
in the parameters of
An important question for long-running computations, is when to decide that a slave process is dead. For our purposes, dead is not a well-defined concept. If a user on the remote machine decides to re-boot, it is clear that any slave processes residing on that machine should be declared dead. However, suppose there is temporary congestion on the network making the slave unavailable. Suppose that another user on the remote machine has started up many processes consuming many resources, and the TOP-C slave process is being starved for CPU time or for RAM. Perhaps the most difficult case of all to decide is if one particular TOP-C task requires ten times as much time as all other tasks. This last example is conceivable if, for example, each task consists of factoring a different large integer.
Hence, our implementation of the TOP-C model will employ the following
heuristic in a future version, to decide if a task is dead. You may
wish to employ this heuristic now, if you have a demanding application.
We use the ParGAP function,
UNIX_Realtime(), to keep track of how
much time has been spent on a task (based on ``wall clock time'', and not
on CPU time). If a task has taken
slaveTaskTimeFactor times as much
time as the longest task so far, then it becomes a candidate for being
declared dead. The GAP variable
slaveTaskTimeFactor is initially set
to the default value of 2.
Once a slave process becomes a candidate for being declared dead,
MasterSlave() will create a second version of the same task, with the
same task input as the original task.
MasterSlave() will then record
which task finishes first. If the original version finishes first, then
the second version of the task is ignored, and the slave process
executing the original task is no longer considered a candidate for
If, however, the second version of the task finishes before the original
version, then the time for the second task is recorded. Further, the
output from the second task will be used, and any output resulting from
the original task will be ignored.
MasterSlave() then periodically
checks until the ration of the time spent so far on the original version
of the task is at least
slaveTaskTimeFactor times greater than the time
spent on the second version of the task, then the process executing the
original version of the task is then declared dead. No further messages
from the process executing original task will be recognized and no
further messages will be sent to that slave process.
A future version of this distribution will include direct support for
this heuristic. A customized version of it may be used now, by taking
advantage of the ParGAP routine,
UNIX_Realtime(). In addition, a
future version of this distribution may include the ability to start new
slave processes in an ongoing computation. The reference CG98
describes how this was done in a C implementation, and why this concept
fits naturally with the TOP-C model.
[Up] [Previous] [Next] [Index]