Why MPI Makes You Scream!  And how can we simplify parallel debugging? Session: BoF-41 Event Type: BOF Time: 12:15pm - 1:15pm Speaker(s): Jeff Squyres, Jayant DeSouza Location: 6A Abstract: The BOF will be in the form of a "directed discussion," as opposed to an open "what do you think?" forum.  Suggested readings: http://cw.squyres.com/columns/2004-07-CW-MPI-Mechanic.pdf, http://cw.squyres.com/columns/2004-08-CW-MPI-Mechanic.pdf, http://cw.squyres.com/columns/2004-12-CW-MPI-Mechanic.pdf, http://csdl.ics.hawaii.edu/se-hpcs/papers/11.pdf Discussion Points: Pre-running problems: rsh/ssh sillyness (e.g., passwords, passphrases, "dot" files), other authentication issues, running in heterogeneous situations. Once running, current state of parallel debugging: printf and its problems (e.g., nondeterministic ordering, influencing timing, etc.), start up (gdb/multiple xterms), real parallel debuggers (vendors, etc.), parallel tools (open source and vendors), automated MPI message checkers (Marmot, Umpire, MPI Check, Intel Message Checker). Open questions: What should MPI implementations do to make life easier for end users and sys/cluster admins? What makes you scream with parallel debugging? What would make things better for parallel debugging? Chair/Speaker Details: Jeff Squyres Open Systems Laboratory, Indiana University Jayant DeSouza Advanced Computing Center, Intel Americas, Inc.