Special Topics: Parallel Training

Lens supports batch-level parallel training with a client-server model. Each process must have its own copy of the network loaded and its own training set. In this model, the server is responsible for performing all weight updates and the clients are responsible for gathering link error derivatives over a batch of examples. The two training modes, synchronous and asynchronous, are described below.

The Server

The startServer command is used to make one process the server. The command will return immediately but will cause the process to begin listening for client connections on a certain port. The port may be specified or will be chosen automatically. startServer returns the port number that was used. The simplest option is to choose a port that you are pretty sure will be available (try the 2000 to 9000 range) and hard-code this number into the server and client scripts.

A more robust option is to allow an open port to be chosen automatically and to then store this value somewhere that the clients will be able to retrieve it, either by altering the client's script or by writing the port number into a known temporary file. A server will break the connections to all clients and stop listening on the server port when the stopServer command is issued.

The clientInfo command is used to find out how many clients are connected, some information about each one, and what each one is doing at the moment. A client that is idle is participating in the current training session but is waiting for the server before it will continue. A client that is halted has stopped participating in the current training session. This may be because the client was stopped explicitly by the user or because it has an incompatible network loaded or no training set. Otherwise, clients can be busy training or testing.

The Clients

A process becomes a client when the startClient command is invoked. This takes as arguments the host and port of the server and an optional address for the client. This address is useful if the client has multiple network connections. If the client successfully connects to the server, a new TCP stream will be opened which uses a new port on the server end. Like startServer, this command returns immediately. The connection to the server can be terminated with stopClient. A process may not be a client and server at the same time or a client for multiple servers.

However, because the server process often does a lot less work, you may find it advantageous to run both a server and a client on the same processor.

Client/Server Communication and Synchronization

The principal means by which clients and the server can communicate is the sendEval command. This sends a command to the server or to the clients. The receiver will then execute the command. This can be used by the server to tell one or all clients to run a particular script or to load a particular example file. It might also be used by a client to communicate information back to the server.

Lens provides a shortcut, sendObject, for setting an object field on the server and on all clients to the same value. This is used like setObject but the value will be set remotely as well. If called on a client, the value will only be set on that client and on the server, not on the other clients. If a client needs to cause an object value to be set on all machines, it can use sendEval to tell the server to execute sendObject.

When training in parallel non-interactively, synchronization controls are necessary to ensure that all participants are ready before training begins. Synchronization is provided by the waitForClients command. This has three modes, which are explained in the manual page for the command. In one mode, the server waits until a specified number of clients have connected before waitForClients returns. Alternately, it can arrange for a command to be executed in the server once a certain number of clients have connected. Finally, waitForClients can be used as a simple barrier. In this mode, the command is called by each client and the server when they are ready to proceed. Once all participants are ready and waiting, they will all be released. The examples below show how this command might be used.

Synchronous Training

All parallel training is controlled with the trainParallel command. Training can take one of two forms, synchronous and asynchronous. Synchronous training is functionally equivalent to standard single-processor training. In this mode, the server sends the network's weights to each client, the clients each process part of a batch of examples, accumulating the overall error and link derivatives and testing for the output unit accuracy stopping criterion, and then send this information back to the server. The server accumulates the error and derivatives over all clients, does a single weight update, and then sends the new weights to all clients and another batch begins.

The server's batchSize parameter is the effective batch size in synchronous mode. This determines the total number of examples processed on all clients between weight updates. Depending on the speed of each client, Lens will adjust the number of examples the client is fed to help balance the load.

The actual examples that will be processed in a batch are not chosen by the server. Each client selects its examples based on the example-set-mode of its training set. However, if ORDERED mode is being used, it would not be desirable to have each client process the exact same set of examples. Therefore, in this case clients will randomize the initial example on the very first batch. They will then proceed as they would in single processor training. When running in synchronous mode, you may prefer to simply give each client a different training set, although this could be problematic if a fast processor dominates the batch sharing and the network is over-exposed to its examples. This could also be problematic in asynchronous training for the same reason.

Just as in normal training, synchronous parallel training will stop either if the server is interrupted, enough updates occur, the overall network error on the last batch (summed over clients) reaches the network's criterion or the group criterion is met for all events in all examples in all clients. Once training is done, the server and the clients will print the total time elapsed and the time during which they were actually working, as opposed to waiting for the server or for a client. If the server is working close to 100% of the time, there will be little benefit in adding additional clients. If the server is rarely working, you might reduce the batch size or even start a client on the server machine. In future releases, Lens may also perform training in the server while the clients are working. The amount of training done would be scaled so that it would not interfere with the clients. Training directly in the server has the advantage that weights and derivatives need not be sent over the network.

During training, the clients will print a dot each time they perform a batch. The server will print a report, the frequency of which is given by its reportInterval value. The report is the same as that for normal training but it adds a field for the most recent client to have reported in. Testing may be performed periodically during training by specifying a test interval parameter for the trainParallel command. In synchronous mode, testing will be performed by the server once it has sent a set of batches off to the clients. Testing will always run through the entire test set. Therefore, the test set cannot be in PIPE mode if the pipe is indefinite as the testing will never end.

Asynchronous Training

In asynchronous training, a weight update is performed each time a client returns its derivatives. The server does not wait for all clients to be done, as in synchronous training. The client is immediately sent the new weights and begins the next batch. This has the advantage that it naturally performs load balancing as clients need not wait for each other, in principle. For a given batch size, there is less communication required in asynchronous training. However, it has the disadvantage that each client is always working with slightly out-of-date weights. If the weights are changing too rapidly this could be quite bad. When training asynchronously, it may be best to begin with a bit of standard single-processor training or synchronous parallel training until the network settles down.

Training will stop if the maximum number of updates are performed or the overall error or output units reach criterion. However, overall error and the unit criteria are just based on the last client to report in and are not combined across all clients as in synchronous mode.

Running Interactively

If the "Stop Parallel Training" button is pressed or an interrupt signal is received in the server during training, the training will stop. It is better to use the button because generating an interrupt signal while the server is processing a message from a client can interfere with that message. If the server detects such a problem, it will force the client to disconnect rather than risk trashing the network. This should not happen if the button is used.

If the stop button or Ctrl-C are used in a client while it is involved in training, the client will halt. This does not break the client's connection to the server, it just means that the client will no longer participate in the current round of training. Again, it may be better to use the button.

While a process is involved in parallel training, it is still able to receive GUI events. As long as it is not waiting, it will also interpret any commands that are given to the shell interface. Therefore, the user could run some testing or change parameters in the midst of training.

If Lens was started in batch mode, the interpreter will not automatically check for events other than keyboard input to the shell. Therefore, clients and servers will not detect when messages arrive. The wait and waitForClients commands can be used to remedy this problem. The former will loop indefinitely or until the parallel training ends or an interrupt occurs, repeatedly checking for messages. The latter is similar but acts as a barrier and will release all of the waiters once the barrier is broken.

Running Non-Interactively

When running non-interactively, it is important to use synchronization mechanisms to make sure the server and clients are coordinated before training. Additionally, if the processes are being run in the background, you cannot let them reach the end of the main script or the process may suspend itself because there is no standard input (this is true if the process is run in the background with & or bg but not if run remotely with rsh). Therefore, wait and waitForClients should be used in the clients after connecting to the server to prevent them from dying. There are many ways to actually structure the interactions. One example is provided below.

You could arrange for the network's weights to be saved every hour by using a command such as this:

    set allDone 0
    proc saveIt i {
        global allDone
        saveWeights log$i.wt.gz
        if {$allDone == 0} {after 3600000 "saveIt [incr i]"}
    saveIt 0
    set allDone 1

This saves the initial weights and arranges for weights to be saved every hour, starts the training, and then waits for training to finish. The automatic saving is cancelled once training completes.


Here are two scripts that can be run in the server and clients to control non-interactive parallel training. They are explained below. Here is the server script:
### Server Code
set workingDirectory ~/Lens
set networkScript  ~/Lens/Examples/filler.in
set clientScript   client.tcl
set clientMachines {sp_4 sp_6 sp_8 sp_9 sp_10 sp_11 sp_12 sp_13}
set synchronous    1
set batch          400
set updates        1000
set reportInterval 10
set rate           0.005
set weightFile     $workingDirectory/final.wt.gz

# Start the server
set port [startServer]

# Write the customized client script
set customClientScript $workingDirectory/client[getSeed].tcl
regsub -all / $networkScript \\/ netScript
sed "s/SCRIPT/$netScript/; s/SERVER/$env(HOST)/; s/PORT/$port/" \
	$clientScript > $customClientScript

# Here we define a command for launching clients using rsh
proc launchClients {customClientScript machines} {
    global env
    set i 0
    foreach client $machines {
        puts "  launching on $client..."
	exec /usr/bin/rsh $client -n \
            "lens -b $customClientScript > /dev/null" &
        incr i
   return $i

# Now we use the command
puts "Launching clients..."
set numClients [launchClients $customClientScript $clientMachines]

# Load the network and training set
source $networkScript

# Now wait for the clients to connect.
puts "Waiting for $numClients clients..."
waitForClients $numClients

# Set some parameters.
setObject batchSize    $batch
setObject learningRate $rate

# Start training and wait for it to finish,
# but don't wait if it didn't start correctly.
puts "Training..."
puts [trainParallel $synchronous $updates $reportInterval]

puts "Saving weights..."
saveWeights $weightFile

# Now break the barrier holding the clients so they can exit.
puts "Releasing clients..."
exec rm [glob $customClientScript]
puts "Ba-bye"

The basic client script is as follows. The upper case words are replaced with their values by the sed command in the server script to created a customized client script:

### Client Code

source SCRIPT
startClient SERVER PORT
puts "Ready to go..."
puts "Gee, that didn't hurt a bit."

If you plan to do parallel training, it will be worthwhile to read through these scripts and understand them fully. The first part of the server script contains the main definitions the user may want to customize. The working directory is a directory that the server has write privilege to and all clients can read from. Using a ~ in this case assumes that the server and clients are using the same file system. If this is not the case, there are a number of ways around this problem, the easiest of which may be using symbolic links on all client machines.

The next section makes the current process a server and stores the port number. It will then write a customized client script and place it in the working directory. The customized client script has the network building script and the server's name and port stored in it. This is done by using sed to replace the keywords in the generic client script with the desired values. Before this is done, all forward slashes in the network script name are protected so the sed command parses correctly.

Next we define a function that will launch the remote clients using rsh. In order to do this, the server machine must be listed in either the /etc/hosts.allow or ~/.rhosts files on the client machines. The output from each client is discarded in this example, but it could be logged in a file on the client if so desired. This command returns the number of clients launched.

Each client will source the network and training set building file, connect to the server and then wait. The client will not wake up from this wait until the training is over. Meanwhile, the server loads the same network, although it need not load the training set, and then waits for all clients to connect. Once the clients connect, training begins.

At the conclusion of training, the server saves the weights and then releases the clients by calling waitForClients. The clients exit on their own and the server removes the customized training file before exiting. The automatic weight saving code could be added to this if periodic backups were desired.

Douglas Rohde
Last modified: Mon Nov 13 13:45:52 EST 2000