NERSC FY 1999 User Survey Results: Software

Comments about NERSC's software resources, suggested improvements, future needs - 24 responses

•   8:     Comments on compilers
  • f90 slow, unforgiving, I/O is slow, data structures cryptic
  • f90 needs to optimize for the SV1
  • KCC slow and doesn't work with totalview
  • wants gcc
  • wants FORTRAN 77
•   5:     Comments on graphics
  • NCAR on the T3E (2 requests)
  • more graphics on the T3E in general
  • wants DISSPLA, PGPLOT
  • wants graphics other than NCAR
  • uses own version of gnuplot
•   5:     Comments on debuggers, analysis tools
  • Problems using totalview on the T3E, doesn't work with KCC.
  • Provide better performance profiling tool on the T3E.
  • Needs easier ways to debug (as easy as on a workstation).
  • Needs a non-graphics based debugging utility.
•   4:     Comments on utilities, editors, shells, user environment
  • doesn't like modules
  • GNU tools and emacs need to be kept up-to-date
  • system doesn't recognize user's terminal type
  • wants easier editors
•   3:     Comments on setcub
  • it's cryptic
  • wants job level accounting
  • doesn't have updated account info
•   5:     Other comments
  • prompt fix of MPI bug
  • need good man pages
  • as a center, should help with software performance improvements

• Comments on compilers:   8 responses

  [...] The f90 compiler [on mcurie] is very slow.

  I am somewhat dissatisfied with KCC's compiler on the T3E for 2 reasons.
  1. It takes a long time to compile, and a longer time to link. This make the debugging cycle extremely painful. I do as little code development as possible on the T3E.
  2. The totalview debugger does not work with KCC. This makes debugging especially difficult.

  It would be nice if the latest gcc were ported to the T3E, but that may be wishful thinking.

  It is very hard to figure out exactly how data structures are laid out. I eventually had to rewrite all my IO in C because the F90 IO is too slow and the layout of F90 structures by the compiler is not helpful. [T3E user]

  The currently installed Cray C++ compiler has a runtime environment that is so archaic and so at odds with the ANSI/ISO C++ standard as to make it exceedingly difficult to use with even the most modest degree of code portability. That might be a Cray problem rather than a NERSC problem, but it is certainly a problem. The availability of KAI's "KCC" C++ compiler ameliorates this somewhat. [...]

  It would be great if NERSC had a standard FORTRAN 77 compiler.

  The FORTRAN compiler on NERSC is very "non-forgiving." I compile my programs on my NT-station (ABSOFT) with no problem, however, upon transferring it to NERSC it will not compile at all. Sometimes, it looks like I am better off rewriting the program on NERSC than to debug it. [...] [PVP user]

  The SV1 needs fortran compilers better able to optimize for its particular architecture. Documentation on how to optimise C90 codes to perform well on the SV1 is needed. The multitasking (OpenMP) software and fortran compilers need a safe mode, i.e. one where the code is guaranteed to produce reproducible results (to the last bit).

• Comments on graphics:   5 responses

  [...] Please load NCAR on the T3E. [...]

  I would like to see graphics packages on the T3E, especially NCAR.

  As mentioned above, having DISSPLA would help. I also think you should have PGPLOT available. NCAR Graphics is a pile of crap.

  I would like additional graphics libraries. I do not like using NCAR graphics because it is difficult to use and the contour plots look bad. When I move codes from other computers or run old codes from NERSC, I always have to first strip off the graphics calls to DISSPLA, PLPLOT, TV80, Quickdraw, etc.

  I use gnuplot, installed in my bin directory.

• Comments on debuggers, analysis tools:   5 responses

  [...] Totalview is usually good [on mcurie], but barfs on complicated (but standard) Fortran-90 constructs, requiring re-writes for some debugging purposes. Totalview does not handle multiple PE debugging with total success very often. The problems can be worked around, but a new version would be nice. Managing memory, checking memory usage (for dynamically allocated arrays), and finding memory errors in my codes is not very convenient. If there are tools available to help with this, please consider getting them.

  [...] The totalview debugger does not work with KCC. This makes debugging especially difficult.

  Better performance profile tool on T3E.

  [...] Also the debugging of FORTRAN does not come close to that of ABSOFT (NT-version). [PVP user]

  As a remote user, I miss having a non-graphics based debugging utility.

• Comments on utilities, editors, shells, user environment:   4 responses

  I am not a fan of the Modules package.....Just thought I'd let you know there is at least one of us.

  [...] Some of the software could stand to be updated more frequently, particularly GNU tools. Emacs, for example, is a pretty fundamental tool for any software developer, yet the T3E version is rather ancient. That said, the system administrators are quite responsive to specific requests for installations and upgrades. I've been holding back on some requests because I don't want to make a pest of myself.
  More shells are needed! "zsh" and "tcsh" should be installed, and it should be possible to set "zsh", "tcsh", or "bash" as one's login shell.

  I have to specify my terminal type at each login.

  Use more friendly environment, easy editors, such as pico.

• Comments on setcub:   3 responses

  Accounting tools are rather cryptic.

  A quick way to account how much repo was used by each job.

  I find that setcub doesn't have updated account information. Twice I've been fooled into thinking I have a larger account balance than I really do, and then I unexpectedly used up our time. Now I'm keeping track on my own (or trying to) of how much time I've used.

• Other comments:   5 responses

  [...] An MPI bug I reported was fixed promptly. Thank you.

  The best one among all I have used.

  I need to learn the basics of how to use a piece of software very quickly and get on with my work. If I can't do this for whatever reason, I probably don't use it unless I absolutely have to. I get the bulk of my info from the on-line man pages, not the web documentation.

  Others in my group have beat through the configuration aspects, so that my code essentially works now out of the box. I gather however that there was a bit of pain getting it there....a long winded way of saying no opinion, I s'pose.

  I think in the next decade, demands for an order of magnitude improvement in the speed of hardware and performance of software ,etc. would be mandatory for any supercomputing facility like NERSC. NERSC should start plans NOW how it would tackle this mandate. Dr. Horst Simon is most knowledgeable in this matter and I am sure he would cross the bridge when he would be close to it. Dr. Simon at Boeing ( 25 years ago) and NERSC is doing a most magnificent job! He deserves our sincerest thanks for making NERSC what it is today. NERSC is undoubtedly a state-of the art supercomputing facility second to none in the western world. Congratulations to Dr. Horst Simon.

Next: Training