Cplant Release 1.0

April 17, 2001

This first public release of Cplant software adds several new features and enhancements over the previous version. The Cplant parallel runtime environment software now uses the Portals 3.0 API for all message passing. Previous versions used the Portals 2.0 API, which has been completely removed from the source code. This release of Cplant software provides greater message passing reliablility through enhanced error detection and correction protocols. Previously, the Myrinet message passing software was unable to recover from severe network errors, such as misrouted or truncated packets. Enhancements to the RTS/CTS module now recover from a wider variety of Myrinet network errors. Because this is the first public release of Cplant software, the version has been named 1.0 for ease of reference.

Highlights

• Enhanced Myrinet error correctin/detection protocol
• Server library uses Portals 3.0
• RTS/CTS does more complete error correction
• PCT performs a node health check

Bug Fixes

• Fixed a bug where PCT would terminate with a SIGSEGV occasionally, right after an application process had terminated.
• Bug fix to the nodekill() library call.

Enhnancements

• yod computes a quick check sum on the executable file and sends that to the PCT. If the PCT computes a different sum, we abort the load. The purpose is to detect network errors early and give up.
• Debugging points added where application process will be held after fork and before user code begins.
• Added Cplant job ID to the user log file.
• Improved yod's reporting of load failures caused by PCT group communication failures.
• If a load fails due to PCT group communication errors, yod will retry the load up to three times.

Known Bugs

• There is a known bug in the RTS/CTS module that exhibits itself when a node is sending large messages (> ~256 kB) to many nodes (> ~400) simultaneously. The sending node will run out of internal send buffers and not be able to complete any of the messages for a long time (minutes to hours or never) We have a fix for this, but it is not tested enough to be part of this Release.
• There is a feature in the message passing system that has not been implemented fully yet. The service utilities (bebopd and pct) once in a while determine that a node is not alive and give up on messages that have not completed yet. The notification that a message should be canceled is not passed all the way to the RTS/CTS module, and it keeps trying to send the message. We are investigating what the effects of this could be, but during normal operation this should not be a problem. Programs that exit, such as yod and user applications, should not be affected, since the message queues are cleaned out when a process exits.