Troubleshoot Common Problems

This section offers advice on solving problems you might encounter with MATLAB® Distributed Computing Server™ software.

License Errors

When starting a MATLAB worker, a licensing problem might result in the message

License checkout failed. No such FEATURE exists. 
License Manager Error -5

There are many reasons why you might receive this error:

  • This message usually indicates that you are trying to use a product for which you are not licensed. Look at your license.dat file located within your MATLAB installation to see if you are licensed to use this product.

  • If you are licensed for this product, this error may be the result of having extra carriage returns or tabs in your license file. To avoid this, ensure that each line begins with either #, SERVER, DAEMON, or INCREMENT.

    After fixing your license.dat file, restart your license manager and MATLAB should work properly.

  • This error may also be the result of an incorrect system date. If your system date is before the date that your license was made, you will get this error.

  • If you receive this error when starting a worker with MATLAB Distributed Computing Server software:

    • You may be calling the startworker command from an installation that does not have access to a worker license. For example, starting a worker from a client installation of the Parallel Computing Toolbox™ product causes the following error:

      The mdce service on the host hostname 
      returned the following error:
      
      Problem starting the MATLAB worker.
      
      The cause of this problem is:
      
      ==============================================================
         Most likely, the MATLAB worker failed to start due to a
         licensing problem, or MATLAB crashed during startup.  Check 
         the worker log file
         /tmp/mdce_user/node_node_worker_05-11-01_16-52-03_953.log
         for more detailed information.  The mdce log file
         /tmp/mdce_user/mdce-service.log
         may also contain some additional information.
      ===============================================================
      

      In the worker log files, you see the following information:

      License checkout failed.
      License Manager Error -15
      MATLAB is unable to connect to the license server. 
      Check that the license manager has been started, and that the 
      MATLAB client machine can communicate with the license server.
      
      Troubleshoot this issue by visiting: 
      http://www.mathworks.com/support/lme/R2009a/15
      
      Diagnostic Information:
      Feature: MATLAB_Distrib_Comp_Engine 
      License path: /apps/matlab/etc/license.dat 
      FLEXnet Licensing error: -15,570. System Error: 115
      

    • If you installed only the Parallel Computing Toolbox product, and you are attempting to run a worker on the same machine, you will receive this error because the MATLAB Distributed Computing Server product is not installed, and therefore the worker cannot obtain a license.

Memory Errors on UNIX Operating Systems

If the number of threads created by the server services on a machine running a UNIX® operating system (Linux® or Macintosh) exceeds the limitation set by the maxproc value, the services fail and generate an out-of-memory error. Check your maxproc value on a UNIX operating system with the limit command. (Different versions of UNIX software might have different names for this property.)

Run Server Processes on Windows Network Installation

Many networks are configured not to allow LocalSystem to have access to UNC or mapped network shares. In this case, run the mdce process under a different user with rights to log on as a service. See Set the User.

Required Ports

With Job Manager

BASE_PORT.  The mdce_def file specifies and describes the ports required by the job manager and all workers. See the following file in the MATLAB installation used for each cluster process:

  • matlabroot/toolbox/distcomp/bin/mdce_def.sh (on UNIX operating systems)

  • matlabroot\toolbox\distcomp\bin\mdce_def.bat (on Windows® operating systems)

Communicating Jobs.  On worker machines running a UNIX operating system, the number of ports required by MPICH for the running of communicating jobs ranges from BASE_PORT + 1000 to BASE_PORT + 2000.

With Third-Party Scheduler

Before the worker processes start, you can control the range of ports used by the workers for communicating jobs by defining the environment variable MPICH_PORT_RANGE with the value minport:maxport.

Client Ports

With the pctconfig function, you specify the ports used by the client. If the default ports cannot be used, this function allows you to configure ports separately for communication with the job scheduler and communication with pmode or a parallel pool.

Ephemeral TCP Ports with Job Manager

If you use the job manager on a cluster of nodes running Windows operating systems, you must make sure that a large number of ephemeral TCP ports are available on the job manager machine. By default, the maximum valid ephemeral TCP port number on a Windows operating system is 5000, but transfers of large data sets might fail if this setting is not increased. In particular, if your cluster has 32 or more workers, you should increase the maximum valid ephemeral TCP port number using the following procedure:

  1. Start the Registry Editor.

  2. Locate the following subkey in the registry, and click Parameters:

    HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Services\Tcpip\Parameters
  3. On the Registry Editor window, select Edit > New > DWORD Value.

  4. In the list of entries on the right, change the new value name to MaxUserPort and press Enter.

  5. Right-click on the MaxUserPort entry name and select Modify.

  6. In the Edit DWORD Value dialog, enter 65534 in the Value data field. Select Decimal for the Base value. Click OK.

    This parameter controls the maximum port number that is used when a program requests any available user port from the system. Typically, ephemeral (short-lived) ports are allocated between the values of 1024 and 5000 inclusive. This action allows allocation for port numbers up to 65534.

  7. Quit the Registry Editor.

  8. Reboot your machine.

Host Communications Problems

If a worker is not able to make a connection with its MATLAB job scheduler (MJS, or job manager), or if a client session cannot validate a profile that uses that scheduler, this might indicate communications problems between nodes.

With Command-Line Interface

First, be sure that the machines in question agree on their IP resolutions. The IP address for a particular host should be the same for itself as it is from the perspective of another host. For example, if a process on hostB cannot connect to one on hostA, find out the hostA IP address for itself, then see what the IP address for hostA is from hostB. They should be the same.

If the machines can identify each other, the nodestatus command can be useful for diagnosing problems between their processes. Use the function to determine what MATLAB Distributed Computing Server processes are running on the local host, and which are accessible from remote hosts. If a worker on hostA cannot register with its job manager on hostB, run nodestatus on both hosts to see what each can see on hostB.

On hostB, execute:

nodestatus -remotehost hostB

Then on hostA, run exactly the same command:

nodestatus -remotehost hostB

The results should be the same, showing the same listing of job managers and workers.

If the output indicates problems, run the command again with a higher information level to receive more detailed information:

nodestatus -remotehost hostB -infolevel 3

With Admin Center GUI

You can diagnose some communications problems using Admin Center.

If you cannot successfully add hosts to the listing by specifying host name, you can use their IP addresses instead (see Add Hosts). If you suspect any communications problems, in the Admin Center GUI click Test Connectivity (see Test Connectivity). This testing verifies that the nodes can identify each other and allow their processes to communicate with each other.

Verify Network Communications for Cluster Discovery

If you want to use the discover cluster capabilities in Parallel Computing Toolbox, your network must be configured with at least one of the following:

DNS SRV Record

Using DNS for cluster discovery requires that you have a DNS SRV record of the following general form:

_mdcs._tcp.domainname.com SSSS IN SRV PPPP WWWW MJS_PORT MJS_FQDN_HOSTNAME

The parts of this record are:

  • _mdcs._tcp. The record must start with this text, followed by your domain name (like company.com or university.edu) that the client machine searches.

  • SSSS indicates how long (in seconds) the DNS record can be cached; 3600 is recommended.

  • IN SRV is required as shown, indicating that this is a service record.

  • PPPP and WWWW indicate priority and wait values. These are not used, so 0 is recommended for each.

  • MJS_PORT is the port on which you connect to the MJS server. The default is 27350, but if you change it for the server you must change it here accordingly.

  • MJS_FQDN_HOSTNAME is the fully qualified domain name for the host serving the MJS. For example, mjs-1.company.com.

A valid DNS SRV record for the company.com network running an MJS on machine mjs-1 might look like this:

_mdcs._tcp.company.com 3600 IN SRV 0 0 27350 mjs-1.company.com

For your network, create the appropriate DNS SRV record using the standard procedure for your DNS system. Then you can verify that your network is configured with the necessary DNS SRV records by using standard utilities, such as the nslookup command. For example, this system command indicates the existence of the applicable DNS SRV records:

nslookup -type=SRV _mdcs._tcp.company.com 

Multicast

To use multicast, it is required on the head node running the MATLAB job scheduler (MJS) and on the client system.

Multicast, unlike TCP/IP or UDP, is a subscription-based protocol where a number of machines on a network indicate to the network their interest in particular packets originating somewhere on that network. By contrast, both UDP and TCP packets are always bound for a single machine, usually indicated by its IP address.

The main tools for investigating this type of packet are:

  • tcpdump for UNIX operating systems

  • winpcap and ethereal for Microsoft® Windows operating systems

  • A Java® class included with the parallel computing products.

The Java class is called com.mathworks.toolbox.distcomp.test.MulticastTester. Both its static main method and its constructor take two input arguments: the multicast group to join and the port number to use.

This Java class has a number of simple methods to attempt to join a specified multicast group. Once the class has successfully joined the group, it has methods to send messages to the group, listen for messages from the group, and display what it receives. You can use this class both from a command-line call to Java software and inside MATLAB.

From a shell prompt (assuming that java is on your path), type

java -cp distcomp.jar com.mathworks.toolbox.distcomp.test.MulticastTester

You should see an output something like this:

0 : host1name : 0
1 : host2name : 0

The following example shows how to use the Java class inside MATLAB.

Start MATLAB on two machines (e.g., host1name and host2name) for which you want to test multicast. In each MATLAB session, enter the following commands:

m = com.mathworks.toolbox.distcomp.test.MulticastTester('239.1.1.1', 9999);
m.startSendingThread;
m.startListeningThread;

These instructions cause each MATLAB session to issue a stream of multicast test packets, and to listen for test packets. If multicast is working between the machines, you see a stream of lines like the following:

 0 : host1name : 0
 1 : host2name : 0
 2 : host2name : 1
 3 : host2name : 2

The number on the left in each string is the line number for the received packet. The text in the center is the host from which the packet is received. The number on the right is the packet number sent by the sending host. It is normal for a host to report a test packet from itself.

If either machine does not receive a stream of test packets, or if the remote host is not included in either stream, then multicast communication is not operating properly.

To terminate the test stream, execute the following in both MATLAB sessions:

m.stopSendingThread;
m.stopListeningThread;
Was this topic helpful?