<br><tt><font size=2>>So far, the argument I have heard for allflushall
is: BGP does not give remote completion information to the source.
Surely making it collective would be better. <br>
<br>
>When I challenged that and asked for an implementation sketch, the
implementation sketch provided is demonstrably worse for many scenarios
than calling flushall and a barrier. It would >be a lot easier
for the IBM people to do the math to show where the crossover point is,
but so far, they haven't. <br>
</font></tt>
<br><tt><font size=2>My point was, the way Jeff is doing synchronization
in NWChem is via a fenceall(); barrier(); on the equivalent of MPI_COMM_WORLD.
If I knew he was going to be primarily doing this (ie, that he wanted to
know that all nodes were synched), I would do something like maintain counts
of sent and received messages on each node. I could then do something like
an allreduce of those 2 ints over the tree to determine if everyone is
synched. There are probably some technical details that would have to be
worked out to ensure this works but it seems good from 10000 feet.</font></tt>
<br>
<br><tt><font size=2>Right now we do numprocs 0-byte get operations to
make sure the torus is flushed on each node. A torus operation is ~3us
on a 512-way. It grows slowly with number of midplanes. I'm sure a 72 rack
longest Manhattan distance noncongested pingpong is <10us, but I don't
have the data in front of me.</font></tt>
<br>
<br><tt><font size=2>A tree int/sum is roughly 5us on a 512-way and grows
similarly. I would postulate that a 72 rack MPI allreduce int/sum is on
the order of 10us. </font></tt>
<br>
<br><tt><font size=2>So you generate np*np messages vs 1 tree message.
Contention and all the overhead of that many messages will be significantly
worse than even several tree messages.</font></tt>
<br>
<br><tt><font size=2>I think you really summarized it for me on BGP at
least:</font></tt>
<br><tt><font size=2>>MPI_Bcast/(insert: "collective synchronization")
can ALWAYS be made faster than a naïve implementation over p2p. That
is the point of a collective. </font></tt>
<br>
<br><tt><font size=2>Anytime I know that an operation is collective, I
can almost guarantee I can do it better than even a good pt2pt algorithm
if I am utilizing our collective network. I think on machines that have
remote completion notification an allfenceall() is just a barrier(), and
since fenceall(); barrier(); is going to be replaced by allfenceall(),
it doesn't seem to me like it is any extra overhead if allfenceall() is
just a barrier() for you.</font></tt>
<br>
<br><tt><font size=2>Just my $.02.</font></tt>
<br>
<br>
<br><font size=2 face="sans-serif"><br>
Brian Smith (smithbr@us.ibm.com)<br>
BlueGene MPI Development/<br>
Communications Team Lead<br>
IBM Rochester<br>
Phone: 507 253 4717</font>
<br>
<br>
<br>
<table width=100%>
<tr valign=top>
<td><font size=1 color=#5f5f5f face="sans-serif">From:</font>
<td><font size=1 face="sans-serif">"Underwood, Keith D" <keith.d.underwood@intel.com></font>
<tr valign=top>
<td><font size=1 color=#5f5f5f face="sans-serif">To:</font>
<td><font size=1 face="sans-serif">"MPI 3.0 Remote Memory Access working
group" <mpi3-rma@lists.mpi-forum.org></font>
<tr valign=top>
<td><font size=1 color=#5f5f5f face="sans-serif">Date:</font>
<td><font size=1 face="sans-serif">05/20/2010 09:19 AM</font>
<tr valign=top>
<td><font size=1 color=#5f5f5f face="sans-serif">Subject:</font>
<td><font size=1 face="sans-serif">Re: [Mpi3-rma] RMA proposal 1 update</font>
<tr valign=top>
<td><font size=1 color=#5f5f5f face="sans-serif">Sent by:</font>
<td><font size=1 face="sans-serif">mpi3-rma-bounces@lists.mpi-forum.org</font></table>
<br>
<hr noshade>
<br>
<br>
<br><tt><font size=2>> What is available in GA itself isn't really relevant
to the Forum. We<br>
> need the functionality that enables someone to implement GA<br>
> ~~~efficiently~~~ on current and future platforms. We know ARMCI
is<br>
> ~~~necessary~~~ to implement GA efficiently on some platforms, but<br>
> Vinod and I can provide very important cases where it is ~~~not<br>
> sufficient~~~.<br>
<br>
Then let's enumerate those and work on a solution.<br>
<br>
> The reason I want allfenceall is because a GA sync requires every<br>
> process to fence all remote targets. This is combined with a
barrier,<br>
> hence it might as well be a collective operation for everyone to fence<br>
> all remote targets. On BGP, implementing GA sync with fenceall
from<br>
> every node is hideous compared to what I can imagine can be done with<br>
> active-message collectives. I would bet a kidney it is hideous
on<br>
> Jaguar. Vinod can sell my kidney in Singapore if I'm wrong.<br>
> <br>
> The argument for allfenceall is the same as for sparse collectives.<br>
> If there is an operation which could be done with multiple p2p calls,<br>
> but has a collective character, it is guaranteed to be no worse to<br>
> allow an MPI runtime to do it collectively. I know that many<br>
> applications will generate a sufficiently dense one-sided<br>
> communication matrix to justify allfenceall.<br>
<br>
So far, the argument I have heard for allflushall is: BGP does not
give remote completion information to the source. Surely making it
collective would be better. <br>
<br>
When I challenged that and asked for an implementation sketch, the implementation
sketch provided is demonstrably worse for many scenarios than calling flushall
and a barrier. It would be a lot easier for the IBM people to do
the math to show where the crossover point is, but so far, they haven't.
<br>
<br>
> If you reject allfenceall, then I expect, and for intellectual<br>
> consistency demand, that you vigorously protest against sparse<br>
> collectives when they are proposed on the basis that they can<br>
> obviously be done with p2p efficiently already. Heck, why not
also<br>
> deprecate all MPI_Bcast etc. since some on some networks it might
not<br>
> be faster than p2p?<br>
<br>
MPI_Bcast can ALWAYS be made faster than a naïve implementation over p2p.
That is the point of a collective. <br>
<br>
Ask Torsten how much flak I gave him over some of the things he has proposed
for this reason. Torsten made a rational argument for sparse collectives
that they convey information that the system can use successfully for optimization.
I'm not 100% convinced, but he had to make that argument. <br>
<br>
> It is really annoying that you are such an obstructionist. It
is<br>
> extremely counter-productive to the Forum and I know of no one<br>
<br>
I am attempting to hold all things to the standards set for MPI-3:<br>
<br>
1) you need a use case.<br>
2) you need an implementation<br>
<br>
Now, I tend to think that means you need an implementation that helps your
use case. In this particular case, you are asking to add collective
completion to a one-sided completion model. This is fundamentally
inconsistent with the design of MPI RMA, which separates active target
(collective completion) from passive target (one-sided completion). This
maps well to much of the known world of PGAS-like models: CoArray
Fortran uses collective completion and UPC uses one-sided completion (admittedly,
a call to barrier will give collective completion in UPC, but that is because
a barrier without completion is meaningless). This mixture of the
two models puts us at risk of always getting poor one-sided completion
implementations, since there is the "out" of telling people to
call the collective completion routine. This would effectively gut
the advantages of passive target. <br>
<br>
So far, we have proposed adding:<br>
<br>
1) Completion independent of synchronization<br>
2) Some key remote operations<br>
3) an ability to operate on the full window in one epoch<br>
<br>
In my opinion, adding collective communication to passive target is a much
bigger deal.<br>
<br>
> deriving intellectual benefit from the endless stream of protests
and<br>
> demands for OpenSHMEM-like behavior. As the ability to implement
GA<br>
> on top of MPI-3 RMA is a stated goal of the working group, I feel
no<br>
> shame in proposing function calls which are motivated entirely by
this<br>
> purpose.<br>
<br>
Endless stream of demands for OpenSHMEM-like behavior? I have asked
(at times vigorously) for a memory model that would support the UPC memory
model. The ability to support UPC is also in that stated goal along
with implementing GA. I have used SHMEM as an example of that memory
model being done in an API and having hardware support from vendors. I
have also argued that the memory model that supports UPC would be attractive
to SHMEM users and that OpenSHMEM is likely to be a competitor for mind
share for RMA-like programming models. I have lost that argument
to the relatively vague "that might make performance worse in some
cases". I find that frustrating, but I don't think I have raised
it since the last meeting.<br>
<br>
Keith<br>
<br>
_______________________________________________<br>
mpi3-rma mailing list<br>
mpi3-rma@lists.mpi-forum.org<br>
</font></tt><a href="http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma"><tt><font size=2>http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma</font></tt></a><tt><font size=2><br>
</font></tt>
<br>
<br>