> Yes I want to pass different arguments to different invocation. For
> example I can have different sets of channels that some localities will
> need to use to communicate.
> 
> I am able to use the parallel_executor as follow:
> 
> -----
> hpx::parallel::execution::parallel_executor exec;
> 
> hpx::when_all(hpx::parallel::execution::bulk_async_execute(
>       exec, &spmd_bulk, args)).get();
> 
> ——
> 
> Now I am trying to understand where those tasks are launched.
> 
> When I run this using -hpx::threads 1 with 8 MPI ranks the
> hpx::find_here() is telling me that all my spmd_bulk are running on the
> same locality (even if I have 8 localities), as a consequence the
> execution is serialized.
> 
> I thought that this executor was supposed to spawn the tasks over
> different localities, but probably I am missing something (or some
> parameters).

No. This executor does local tree-style spawning of tasks. As said, for remote 
operation you could use broadcast, or write your own executor. We have a lot of 
work to do to integrate remote operation with the current executor framework.

Hold on - there is also distribution_policy_executor which can be used to send 
work to remote localities (see 
https://github.com/STEllAR-GROUP/hpx/blob/master/hpx/parallel/executors/distribution_policy_executor.hpp).
 I need to have a look what state this is in, though. I don't think it's doing 
anything fancy like tree-style spawning, though.

> How can I spawn my tasks over multiple localities?
> 
> Should I use something like this instead? https://stellar-
> group.github.io/hpx/docs/html/hpx/manual/containers/partitioned_vector_vie
> ws/spmd_block.html

Well, let's not go there (yet).

Regards Hartmut
---------------
http://boost-spirit.com
http://stellar.cct.lsu.edu


> Thank you,
> Steve
> 
> 
> 
> On 13 Oct 2017, at 06:08, Hartmut Kaiser <[email protected]> wrote:
> 
> Steve,
> 
> 
> Yes, I got the things to work distributing the channels from the main
> task, but I guess simply using hpx:async to spawn the tasks in an SPMD
> fashion is not the best to scale, but it’s the only way I have now to pass
> all these channels to each locality.
> 
> Using the other executors that you mentioned how can I pass different
> arguments to different localities (e.g., the channels)?
> 
> Do you want to pass different arguments to different invocations and still
> use tree-spawning?
> 
> 
> Is there any example of using tree-like spawning or bulk_async_execute
> other then the tests?
> 
> For efficient _local_ tree-style spawning you can use the
> parallel_executor. See here for an example: https://github.com/STEllAR-
> GROUP/hpx/blob/master/tests/unit/parallel/executors/parallel_executor.cpp#
> L94-L113. Please note that the 3rd parameter to bulk_async_execute could
> be any (input) range, for instance a boost::irange(0, num) or similar, it
> doesn't have to be a real container.
> 
> Remote tree-style spawning can be done using broadcast (or
> broadcast_with_index), see here for an example:
> https://github.com/STEllAR-
> GROUP/hpx/blob/master/tests/unit/lcos/broadcast.cpp
> 
> HTH
> Regards Hartmut
> ---------------
> http://boost-spirit.com
> http://stellar.cct.lsu.edu
> 
> 
> 
> 
> Thanks,
> Steve
> 
> 
> 
> 
> 
> 
> On 12 Oct 2017, at 09:53, Hartmut Kaiser <[email protected]> wrote:
> 
> Steve,
> 
> 
> I see your point, but the problem is that I cannot know when the receiver
> is going to be executes and so when it will receive the data.
> 
> Is there a way to wait on this channel until some task connects to and
> make a get? This way I can keep the channel (and the task) alive until the
> data has been consumed. I think there should be some mechanism to say
> “wait for N tasks to connect”.
> 
> That's part of the deeper problem I hinted at. Let me think about this
> some more before I respond.
> 
> 
> I am reporting my other question:
> More in general, when you have many channels at scale, do you
> think is better to use a register_as/connect_to mechanism or to pass alle
> the necessary channels to each locality at the beginning when I do the
> initial SPMD spawn?
> 
> We usually use the register_as/connect_to mechanism as it separates
> concerns nicely. But YMMV, so please try it out for yourself what works
> best in your case. Sending a channel over the wire might overcome the
> lifetime issues we're discussing as this would keep the channel alive no
> matter what.
> 
> Regards Hartmut
> ---------------
> http://boost-spirit.com
> http://stellar.cct.lsu.edu
> 
> 
> 
> 
> Thank you,
> Steve
> 
> 
> 
> 
> 
> On 12 Oct 2017, at 09:28, Hartmut Kaiser <[email protected]> wrote:
> 
> Steve,
> 
> 
> Going back to the channels discussion, how do you know if a channel has
> been already registered or not?
> 
> Look at the following code below (running with 2 localities):
> - they create or connect to the same channel name
> - if I introduce a delay in the receiver locality I get the following
> error:
> 
> Assertion failed: (buffer_map_.empty()), function ~receive_buffer,
> file hpx_install/include/hpx/lcos/local/receive_buffer.hpp, line 101.
> 
> I think I understand what is going on. This assert says that you're trying
> to destroy a channel with data sitting in the pipeline. This is caused by
> your code:
> 
>  {
>       hpx::lcos::channel<int> c(hpx::find_here());
>       c.register_as(“name");
>       c.set(3);
>  }
> 
> Which makes the channel go out of scope before the other locality had a
> chance to connect and to extract the value.
> 
> There is also a deeper problem lurking behind the scenes, but that's
> something I need to think about more before doing anything about it.
> 
> Regards Hartmut
> ---------------
> http://boost-spirit.com
> http://stellar.cct.lsu.edu
> 
> 
> 
> Here is the sample code:
> —————————
> 
> 
> void spmd_int(DataFlow::ShardId cid){
> if(cid%2==0){
>  hpx::lcos::channel<int> c(hpx::find_here());
>  c.register_as(“name");
>  c.set(3);
> }
> else{
>  usleep(1000000); // delay
>  hpx::lcos::channel<int> c;
>  c.connect_to("name");
>  hpx::future<int> p = c.get();
>  p.get();
> }
> }
> 
> 
> int hpx_main(boost::program_options::variables_map& vm){
> std::vector<hpx::naming::id_type> localities =
>          hpx::find_all_localities();
> 
> int loc_count = 0;
> for(auto loc: localities){
> 
>  spmd_int_action act;
>  hpx::async(act, loc, loc_count);
> 
>  loc_count++;
> }
> }
> 
> —————————
> 
> 
> What is happening here?
> If I add a
> c.connect_to("name");
> 
> to the the same task that does the registration (after
> c.register_as(“name");) it works, but I don’t like it (and in my actual
> application I still get this error).
> 
> More in general, when you have many of this channels at scale, do you
> think is better to use a register_as/connect_to mechanism or to pass alle
> the necessary channels to each locality at the beginning when I do the
> initial SPMD spawn?
> 
> Thanks,
> Steve
> 
> 
> On 11 Oct 2017, at 14:53, Steve Petruzza <[email protected]> wrote:
> 
> Thanks Hartmut, it all makes sense now.
> 
> 
> 
> On 11 Oct 2017, at 14:51, Hartmut Kaiser <[email protected]> wrote:
> 
> 
> 
> I think I’ve found a workaround.
> 
> If I use a typedef as following:
> 
> typedef std::vector<char> vec_char;
> 
> HPX_REGISTER_CHANNEL(vec_char);
> 
> It works, but if I try to use directly:
> HPX_REGISTER_CHANNEL(std::vector<char>)
> 
> this gives me the error I reported before.
> The issue might be in the expansion of the macro HPX_REGISTER_CHANNEL.
> 
> Yes, that confirms my suspicion. I will have a looks what's going on.
> 
> Doh! The problem is that the (literal) parameter you give to the macro has
> to conform to the rules of a valid symbol name, i.e. no special characters
> are allowed (no '<', '>', etc.). Sorry, this has to be documented properly
> somewhere, and I forgot to mention it in the first place.
> 
> The 'workaround' you propose is the only way to circumvent problems. There
> is nothing we can do about it.
> 
> Also, wrt your comment that everything is working if you use
> hpx::lcos::local::channel instead - this is not surprising. The local
> channel type is representing a channel which can be used inside a given
> locality only (no remote operation, just inter-thread/inter-task
> communication), hence its name. Those channels don't require the use of
> the ugly macros, thus there is no problem.
> 
> HTH
> Regards Hartmut
> ---------------
> http://boost-spirit.com
> http://stellar.cct.lsu.edu
> 
> 
> 
> 
> Thanks!
> Regards Hartmut
> ---------------
> http://boost-spirit.com
> http://stellar.cct.lsu.edu
> 
> 
> 
> 
> Steve
> 
> 
> 
> 
> On 10 Oct 2017, at 18:38, Steve Petruzza <[email protected]> wrote:
> 
> Sorry, regarding the version that I am using it is the commit of your
> split_future for vector:
> 
> Adding split_future for std::vector
> 
> - this fixes #2940
> 
> commit 8ecf8197f9fc9d1cd45a7f9ee61a7be07ba26f46
> 
> Steve
> 
> 
> 
> On 10 Oct 2017, at 18:33, Steve Petruzza <[email protected]> wrote:
> 
> hpx::find_here()
> 
> 


_______________________________________________
hpx-users mailing list
[email protected]
https://mail.cct.lsu.edu/mailman/listinfo/hpx-users

Reply via email to