[mpiwg-tools] Problem with API From 2020-10-15 Conversation

[Wesley Bland]

I’m not sure such a tool makes a lot of sense given that we’re not strongly specifying the mechanism for determining tool order. If we want to do that, then having a mapping from environment variable to tool ID would make more sense.

Also, even if you have tool ID, that’s probably not enough to go on since you’ll also need to know whether a particular tool intercepted a particular function. If you call the QMPI_Get_next_function(my_tool_id=1, …) and expect to get the function for tool_id=2, that doesn’t work if tool 2 didn’t register an interception for that function. You may get an interception function for tool 3, or none at all and you go all the way back to MPI.

> On Oct 16, 2020, at 10:09 AM, Joachim Protze <protze at itc.rwth-aachen.de> wrote:
> 
> Hi Wesley, Martin,
> 
> I still don't see the need for splitting the function, or what this
> would give us.
> I attached the slides from yesterday for illustration. Slide 3 also
> sketches, how I would emulate the PnMPI_Xy_NewStack function.
> 
> Whoever implements the function (Supertool?) will need to know the
> ordering of the tools and apply a mapping to the stacks. That's why I
> propose a function like:
> 
> QMPI_Get_tool_ids (int count, int * array_of_tool_ids);
> 
> to query the tool IDs according to their ordering in the QMPI_TOOLS env
> variable. The supertool can then build a map for the lookup of stacks:
> {"stack1": 12, "stack2": 13}
> A position map:
> pos = {0:1, 45:2, 12:3, 3:4, 5:5, ...}
> As well as a map for the end of the stacks:
> stacklimit = {12: 6, 13: 9}
> 
> 
> I also propose to have a predefined tool-ID (QMPI_MPI_ID) to allow for
> querying the actual MPI function.
> 
> 
> Best
> Joachim
> 
> 
> 
> Am 16.10.20 um 16:17 schrieb Wesley Bland via mpiwg-tools:
>> Hi all,
>> 
>> In yesterday’s call, we discussed separating the current QMPI_Get_function call (that takes the current tool_id and an enum value to get the function in the call stack) into two function calls. The first would determine the next tool in the stack and the second would get the desired function for that tool. So they would look something like this:
>> 
>> QMPI_Get_next_tool(int tool_id, int num_next_tools, int *next_tool_ids);
>> QMPI_Get_tool_fn(int tool_id, enum QMPI_Functions_enum function_enum, void (*function_ptr)(void), QMPI_Context *context);
>> 
>> However, as I went to implement that today, I realized that the reason this is a single function today is that not all tools implement all functions. So if you have a tool stack that looks like this:
>> 
>> A - B - C
>> 
>> and tools A and C intercept MPI_SEND and tool B does not, when tool A asks for the next tool to implement MPI_SEND, it should get back the function for tool C. Unfortunately, if we don’t ask for a specific function, MPI doesn’t have a way to figure out whether it should return B or C.
>> 
>> So there’s two options here.
>> 
>> 1) Keep the two functions combined. This makes it clear in the current, linear case which tool comes next. However, it doesn’t move us toward the potential future of having a mechanism (internal or external) to make dynamic decisions about the next tool.
>> 
>> 2) Add an enum value to the QMPI_Get_next_tool function so it is only searching for tools that implement the particular function. This does complicate the search functionality a little bit more, but not significantly. The main drawback I see here is just the confusion of passing that enum value into both functions. If we’re not providing this decision functionality in QMPI v1, it’s weird that we have this placeholder for it now. However, that’s already true with the array of next tools that will always be 1, so maybe it’s not a big deal.
>> 
>> I would imagine that 2 is the best option here to support the PnMPI use case, but I’m happy to hear feedback.
>> 
>> Thanks,
>> Wes
>> 
>> 
>> _______________________________________________
>> mpiwg-tools mailing list
>> mpiwg-tools at lists.mpi-forum.org
>> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-tools
>> 
> 
> 
> -- 
> Dipl.-Inf. Joachim Protze
> 
> IT Center
> Group: High Performance Computing
> Division: Computational Science and Engineering
> RWTH Aachen University
> Seffenter Weg 23
> D 52074  Aachen (Germany)
> Tel: +49 241 80- 24765
> Fax: +49 241 80-624765
> protze at itc.rwth-aachen.de
> www.itc.rwth-aachen.de
> <QMPI-stacks.pdf>