PM/Myrinet Test Procedure

If your Myrinet network is Myrinet 2000 with serial or fiber link interface (M3S-PCI64B and M3F-PCI64B), please use myrinet2k network type and lanaiM2k.mcp firmware. If your Myrinet network is Myrinet XP (M3F-PCIXD), please use myrinetxp network type and lanaixp.mcp firmware. In order to change network type in rpmtest and scstest, you must modify your scorehosts.db on your server. Please refer to SCore System Installation on Server section.

When you don't use myrinet network name, please replace network name myrinet to myrinet2k in following commands. Ex. $ ./rpminit comp0 myrinet2k

1. Loopback test
   Issue the following commands:

   $ cd /opt/score/sbin
   $ ./rpminit comp0 myrinet
   $ ./rpmtest comp0 myrinet -dest 0 -ping
   	Success: Two numbers similar to "8     1.2269e-05" are printed
                    The first number is the data size (in bytes)
                    The second number is the latency (in seconds)
   	Failure: Error and dump messages are printed
   

   If this test fails, read the Troubleshooting section.

2. Point-to-Point test (Message)
   Make sure that the PM kernel driver is installed on all hosts. To test the PM communication function, use the rpmtest command. For example, if you want to test communication from comp0 (node 0) to comp1 (node 1), please follow the instruction below:

   Issue the following commands:

   $ cd /opt/score/sbin
   $ ./rpmtest comp1 myrinet -reply
   

   Issue the following commands in another window on the server host:

   $ cd /opt/score/sbin
   $ ./rpmtest comp0 myrinet -dest 1 -ping
   	Results should be the same as the loopback test
   

   When the test finishes, do not forget to kill the pmtest process on the server host because the process is in an infinite loop.
   If this test fails, read the Troubleshooting section.

3. Point-to-Point test (Zero Copy)
   To test communication from comp0 (node 0) to comp1 (node 1), please follow the instruction below:

   Issue the following commands:

   $ cd /opt/score/sbin
   $ ./rpmtest comp1 myrinet -vreply
   

   Issue the following commands in another window on the server host:

   $ cd /opt/score/sbin
   $ ./rpmtest comp0 myrinet -dest 1 -vwrite
   	Success: Two numbers similar to "8     913452" are printed
                    The first number is the data size (in bytes)
                    The second number is the bandwidth (in bytes/sec)
   	Failure: Error and dump messages are printed
   

   When the test finishes, do not forget to kill the pmtest process on the server host because the process is in an infinite loop.
   If this test fails, read the Troubleshooting section.

4. Total test
   This is the stress test. All hosts randomly send burst messages to other hosts. Issue the following commands:

   $ cd /opt/score/deploy
   $ scout -g pcc
   SCOUT: Spawn done.   
   SCOUT: session started
   $ ./scstest -network myrinet
   CSTEST: BURST on myrinet
   50 K messages.
   100 K messages.
   150 K messages.
   200 K messages.
   250 K messages.
   300 K messages.
           Success: number of messages to send are printed.
           Failure: Error messages are printed.
   

   To stop this test, press Ctrl-C or another interrupt command.
   If this test fails, read the Troubleshooting section.

Troubleshooting

Pminit causes an error.

• Myrinet cables are NOT connected.
• The Myrinet device driver is not installed.
• /dev/pmmyri/number has not been created by the mkpmmyridev.sh command.
• The permissions of /dev/pmmyri/number are not correct.
• There is no mcp file or no configuration file.
• The contents of the configuration file are not correct.
• No Myrinet card is installed.
• The Myrinet card is not tightly plugged in.
• The PCI bus clock on the motherboard is higher than 33MHz (ex: 37.5 MHz).
• The Myrinet card is damaged.

Loop back test fails.

• The Myrinet card is not connected to the Myrinet switch correctly.
• Some other processes using Myrinet exist on the host.
• The Myrinet card is not tightly plugged in.
• The Myrinet cable is damaged.
• The Myrinet card is not operating normally.
• The Myrinet switch is not operating normally.

Point-to-point test fails.

• The specification in the configuration file and the actual connection differ.
• Some other processes using Myrinet exist on some hosts.
• The trimer is set to 2000 on the Myrinet-2000 card. it should be set to 1280 because no Myrinet-2000 switches are available at that time (2000/10/25).
• The Myrinet card is not connected to the Myrinet switch correctly.
• The Myrinet cable is damaged.
• The Myrinet card is not operating normally.
• The Myrinet switch is not operating normally.
• Zero copy transfer is not supported on the machine.

Total test fails.

• The specification in the configuration file and the actual connection differ.
• Some other processes using Myrinet exist on some hosts.
• The trimer should be set to 2000 or 1280 on the Myrinet-2000 card. it should be set to 1280 for old type Myrinet switch, or 2000 for Myrinet-2000 switches.
• The Myrinet cable is damaged.
• The Myrinet card is not operating normally.
• The Myrinet switch is not operating normally.