Cplant Release 0.5 Notes

Cplant Release 0.5

June 22, 2000

This release of Cplant software makes great strides in improving the reliability and stability of the runtime environment, especially in the area of message passing. The release contains the next generation of a scalable message passing layer designed specifically for the architecture of a PC cluster, Portals 3.0. The new design has led to increased robustness as well as increased message passing performance. The release also contains a port of the latest version of the MPICH implementation of MPI on top of Portals 3.0. This new implementation uses a standard two-level protocol that addresses some of the limitations of the previous MPI implementation. This release has undergone more stringent testing than any previous releases of Cplant system software. A four-phase approach to testing that covered basic MPI functionality, application micro-benchmarks, several real applications, and system stress tests (ported from ASCI Red) was used. The release was contingent upon satisfactory completion of this suite. More information and documentation on these features will be available at the Cplant web site:

Highlights

Portals 3.0 (kernel module implementation)
MPICH 1.2.0 over Portals 3.0
PBS 2.2

Bug Fixes

Fixed bug in Portals 2.0 buffer touching code
Fixed copy-on-write bug in bebopd and PCT
Fixed bug in RTS/CTS where send-side resources were not getting freed when a message was aborted
Enhanced RTS/CTS so that user process gets a SIGPIPE when a receive or send touches memory that the process does not own
Fixed a bug with using 'copy_to_user()'
Enhanced RTS/CTS by sending a SIGPIPE to a process whose buffer is corrupt due to a lost packet

MPICH MPI tests
Various developer MPI tests
NAS Parallel Benchmarks version 2.3
Applications: Granflow, MPSalsa, Zapotec (CTH and Pronto), Alegra, Ladera, LAMPS
Stress tests (from Intel EATS and SATS)

Known Problems

Alegra encounters message corruption after several hours of running. In heavily congested network situations, it is possible for packet loss to occur. This can result in an application being terminated due to MPI message truncation errors or floating point errors due to erroneous data. It is also possible for the application to hang in an MPI collective operation. The Cplant team has identified the cause of this problem and is working on a resolution.
Some applications on Alaska and Siberia have had trouble obtaining consistent runtimes. A problem with the Linux klogd and syslogd logging daemons has been identified. The Cplant team is presently testing a new version of these daemons and will update the machines if they resolve the problem. If not, these daemons will be turned off.