Thanks. More comments inline. On Mar 3, 2010, at 3:40 PM, Manojkumar Krishnan wrote:

Bill,

Please scroll for my comments.

On Wed, 3 Mar 2010, William Gropp wrote:

...

Thanks. I've added some comments inline.

On Mar 3, 2010, at 1:23 PM, Manojkumar Krishnan wrote:

...

Bill,

Here are some of the MPI RMA requirements for Global Arrays (GA) and ARMCI. ARMCI is GA's runtime system. GA/ARMCI exploits native network communication interfaces and system resources (such as shared memory) to achieve the best possible performance of the remote memory access/one-sided communication. GA/ARMCI relies *heavily* on optimized contiguous and non-contiguous RMA operations (get/put/acc).

For GA/ARMCI and its applications, below are some specfic examples of operations that are hard to achieve in MPI-2 RMA.

1. Memory Allocation: (This might be an implementation issue) The user or library implementors should be able to allocate memory (e.g. shared memory), and register with MPI. This is useful in case of Global Arrays/ARMCI, which use RMA across nodes and shared memory within nodes. ARMCI allocates shared memory segment, and pins/registers with the network.

MPI already provides MPI_Alloc_mem , though this is not necessarily shared memory usable by other processes. But the MPI implementation could allocate this from a shared memory pool and perform one-sided operations using load-store. Do you need something different?

The thing is, this involves additional copy via RMA_Get. Is it possible for other processes within the same SMP node access the memory directly (rather than get/put using load-store).

For example something like this. In GA, you get *direct* access (pointer) to memory regions allocated by processes within SMP node. If proc 0 and 1 are in the same node, proc 1 can access proc 0's region without explict get/put (i.e. proc 1 gets a pointer to this region).

What we may need is a way to provide shared memory (provided by some other service) to an MPI Window, and arrange the access/completion for it. This is something we should look at.

...

...

2. Locks: Should be made optional to keep the RMA programming model simple. If the user doesnot require concurrency, then locks are unnecessary. Enforcing to use locks as default might introduce unnecessary bugs if not carefully programmed (esp. at the level of of extreme scale systems, it is hard to debug).

You need something to provide for local and remote completion. Do you mean to have blocking RMA operations that (in MPI terms) essentially act as lock-RMA_operation-unlock?

Not really, I meant - performing RMA operations without locks, if locks are not needed. Instead of making it default, we can have locks as optional.

I still don't understand what the problem is with "locks" (note that the name is poorly chosen, it isn't really a lock in the usual sense). What lock/unlock does is simple indicate "begin RMA access at target" and "end RMA access at target", with the option of permitting exclusive access to the target window. In the below, I've put what I think are the equivalent MPI RMA calls:

With respect to completion and synchronization, here is an example of how it is done in ARMCI.

ARMCI_Put: Ensures local completion and source buffer is safe to reuse

Lock/Put/Unlock . Permits concurrent non-overlapping destination updates if shared mode is selected.

ARMCI_NbPut: Non-blocking put. User has to call wait() to reuse source buffer.

Lock/Put . Unlock is like Wait

ARMCI_Put(dst_proc)+ARMCI_Fence(dst_proc): Ensures local and remote completion.

Also Lock/Put/Unlock

ARMCI_Allfence. This is collective. Ensures all outstanding RMA operations are completed on the remote side. Mostly called with Barrier.

Win_fence (active target sync)

ARMCI_Put_Notify/Wait. This is like light-weight message passing. After a (large) put, the sender can send a notify message, and the destination process can wait on it. This is useful if send/recv is in redevous mode and send essentially blocks for a slow receiver. In case of Put/Put_notify, the sender can issue 2 seperate R(D)MA's and no need to handshake with the receiver. Thus it is relativly less synchronous. On the other hand, if the sender is slow, then the receiver has to wait!

Please note that, if such a synchronization/completion is required between source and destination, then user should avoid using RMA (and use Send/Recv), as this is truly 2-sided communication.

...

...

3. Support for non-overlapping concurrent operations in a window.

MPI RMA already does this, in all of the active target modes, and in the passive target mode with the shared lock mode. What do you need that isn't supported?

In the passive target mode, does this require some synchronization with the source and destination process if they are accessing non- overlapping operations in the same window. If so, then this might not be truly one-sided operation. Please correct me if I am wrong.

The only synchronization is that an exclusive_lock must block subsequent shared locks and must wait for executing shared locks to complete. Is the issue that you never want to check? If so, then this could be added as a Window characteristic.

...

...

4. RMW Operation - Useful for implementing dynamic load balancing algorithms (e.g. task queues/work stealing, group/global counters, etc).

Yes, this is a major gap. Is fetch-and-increment enough, or do you need compare-and-swap, or a general RMI?

For ARMCI, we need both: fetch-and-increment and compare-and-swap. A general RMI is certainly useful, as we have plans to extend the capability of ARMCI to perform general RMI.

-- Thanks, -Manoj. --------------------------------------------------------------- Manojkumar Krishnan High Performance Computing Group Pacific Northwest National Laboratory Ph: (509) 372-4206 Fax: (509) 372-4720 http://hpc.pnl.gov/people/manoj ---------------------------------------------------------------

...

Bill

...

The above are feature requirements rather than performance issues, which are implementation specific.

If you have any questions, I would be happy to explain the above in detail.

Thanks, -Manoj. --------------------------------------------------------------- Manojkumar Krishnan High Performance Computing Group Pacific Northwest National Laboratory Ph: (509) 372-4206 Fax: (509) 372-4720 http://hpc.pnl.gov/people/manoj ---------------------------------------------------------------

On Wed, 3 Mar 2010, William Gropp wrote:

...

I went through the mail that was sent out in response to our request for use cases, and I must say it was underwhelming. I've included a short summary below; based on this, we aren't looking at the correct needs. I don't think that these are representative of *all* of the needs of RMA, but without good use cases, I don't see how we can justify any but the most limited extensions/changes to the current design. Please (a) let me know if I overlooked something and (b) send me (and the list) additional use cases. For example, we haven't included any of the issues needed to implement PGAS languages, nor have we addressed the existing SHMEM codes. Do we simply say that a high-quality implementation will permit interoperation with whatever OpenSHMEM is? And what do we do about the RMI issue that one of the two use cases that we have raises?

Basically, I received two detailed notes for the area of Quantum Chemistry. In brief:

MPI-2 RMA is already adequate for most parts, as long as the implementation makes progress (as it is required to) on passive updates. Requires get or accumulate; rarely requires put. Dynamic load balancing (a related but separate issue) needs some sort of RMW. (Thanks to Jeff Hammond)

More complex algorithms and data structures appear to require remote method invocation (RMI). Sparse tree updates provide one example. (Thanks to Robert Harrison)

Bill

William Gropp Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

William Gropp Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

William Gropp Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign