as with lustre, pread/pwrtie need a feature level newer than
--enable-strict requests.  autoconf already checked for the function at
configure time, so we know it's there.  TODO: more robust autoconf
checks and provide a "my_pwrite" that wraps lseek/write for a fallback.

P2Pcontig needed a prototype and clang complaiend about several shadowed
declarations.

until we figure out what's up with aio routines on blue gene, let's
just disable it.  the romio aio tests would just hang in aio_suspend

Paul Coffman authored Mar 10, 2014 and Rob Latham committed Mar 26, 2014

remove gpfs-specific shared fp call
further lockless removal
documentation fixups

Paul Coffman authored Mar 21, 2014 and Rob Latham committed Mar 26, 2014

reconfiguration changes from bg to gpfs with platformspec; removal of
lockless

astonishigly, the blue gene L(!) condition lacked a closing 'fi' but we
never noticed since async I/O never worked on blue gene.  Use the AS_IF
macro to make this less likely to recur in the future.

In order to accomodate deferred open, we can't do *any* collective
operations in ad_bg_open.  Any collectives have to happen one level up
at ADIOI_GEN_Opencoll.

We already promoted fs blksize in a prior patch, and simplified
"scalable sync" in another patch, so when we remove the collective call
(bcast of blocksize and fs type), we can also remove the "is it ok to
scalalbe sync"? (because it will always be ok) and the "are we an
fsync-aggregator" logic becuase now only the first io aggregator will be
such an aggregator.

If deferred open is enabled, the logic that says if we should do a
scalable flush and which processes should do the flush won't propagate
to the non-aggregator processes.  Replace old way of doing things with a
simpler stat-from-first-aggregator approach.

some file systems (e.g. bluegene) might stat the file and wish to inform
all processes about some bit of underlying file system information (e.g.
blocksize).  In the deferred open case, not all processes participate in
the lowest, fs-specific open, so let's broadcast here in common code.

"file system blocksize" seems like one of those generic-enough values we
should keep track of in the ADIOI_FileD structure.  This promotion will
make some deferred-open fixes easier, too.

Useful for situations like evaluating various collective I/O approaches.
Reading/writing /dev/null eliminates file system variablity.

Paul Coffman authored Feb 04, 2014 and Rob Latham committed Mar 26, 2014

Two features in this change:
- selection of file domains can result in some i/o nodes with more work
  than others (or some with no work at all), so distribute file domains
  with some awareness of i/o nodes

- since we have some awareness of I/O nodes, select processes that are
  closes to those i/o nodes.

robl's got a one-off logger.  can pass extra information to it with an
environment variable.  probably not useful in general.

Paul Coffman authored Mar 04, 2014 and Rob Latham committed Mar 26, 2014

For certain workloads, MPI processes will only speak to one aggregator.
In those cases, we will restrict communication to just point-to-point
among those processes and their aggregator.  Sometimes called
"p2pcontig" optimization.

in deferred open case, we will have created an "aggregator communicator"
consisting of i/o aggregators.  the 'ranklist' enumerates ranks in
fd->comm, but is not meaningful in the aggregator communicator.
likewise, we do not simply broadcast from '0' in the no-deferred-open
case because rank 0 might not be an aggregator.

Experimental async-with-pthread I/O approach to hiding some of the I/O
latency/variability from the two-phase collectives.

heavily modified from Paul Coffman's (pkcoffman@us.ibm.com) original work

crank up the size of coll_perf to something not laughably small

The only reason these locks exist is becuse way back in BGL days someone
at IBM thought it might be a good idea to have one driver that could
access both NFS and GPFS.  There was also some concern about a large
write call getting split up by the i/o forwarder.  fortunately, MPI-IO
semantics mean applications that would be harmed by such a split already
face "undefined" behavior.

There are many memory allocations in the write path.  Allocating the
two-phase intermediate buffer outside of the write path might on some
systems make a small difference, especially if there are many collective
I/O calls, or if the system (like Blue Gene) has a small amount of
memory.  Modified from Paul Coffman <pkcoff@us.ibm.com>'s original idea.

For quite some time the barrier here has had the comment 'Why?'.  Since
no one knows, and there are plenty of other syncronization points in
this path, remove it.

bluegene timer code had two "levels" of timing.  that seemed kind of
pointless so lump it all into one level.

this "new" system call (part of POSIX-2001) saves us a system call on
Blue Gene.  Seems to get us back 5 seconds for one workload at small
(half rack) scales.

bglockles uses the common read/write routines for contig read/wrties, so
bluegene timing infrastrucutre wasn't actually timing anything.  Since
this introduces blue gene bits into common code, please do not merge to
master.  Instead, we should rework all the timing bits so that it no
longer times "bluegene" but rather all of ROMIO.  Furthermore, the
locky bits of 'bg:' driver should be yanked anyway, obviating the need
for bglockless.

Protected by an 'ifdef', this BGL-era code bitrotted a bit.  clean it up
and see if it does anything useful today.
- Removes preprocessor guards: the counters and timers do nothing
  expensive unless environment variables are set
- remove the idea of a "level"
- remove barrier from timing collection.
- bugfix: MPI_Wtime() does not necessarily start at zero, so properly initialze
  timers for collective read/write
- report only from I/O aggregators.  when reporting "time spent in i/o"
  vs "time spent communicating" it makes more sense to look only at the
  aggregators.  The non-aggregators are going to skew the results
  because they are spending some communication time actually
  communicating, but some of that time blocked, waiting for aggregators
  to finish.

Kenneth Raffenetti authored Mar 19, 2014 and Pavan Balaji committed Mar 26, 2014

De-dupes executable and host names in the MPIR_proctable by pointing
to an existing copy. Closes #1821

Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

The following commits are reverted.

1. "Better checks for VA_ARGS."; commit
de80ec87.

2. "Warning squash for clang."; commit
ccf7f70c

.

The clang warning this was originally trying to solve has been fixed
by the newer versions of clang, AFAICT.

Signed-off-by: Huiwei Lu <huiweilu@mcs.anl.gov>

No reviewer.

ROMIO code assumes all processes will use the same ROMIO driver.  we
were not reaching the "find a common file system" logic when NFS was
enabled, everyone stat-ed the file system without errors, but some
processees found a different file system (like if some processes are
writing to NFS and others to UFS)

See discussion beginning here:
http://lists.mpich.org/pipermail/discuss/2014-March/002403.html



Tested-by: Jeff Squyres <jsquyres@cisco.com>

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

  powerpc64-bgq-linux-gcc v4.7.2