Meeting - 14th December 2016
Protocol of the GASPI Forum meeting on December, 14th, 2016, Frankfurt/Main
Date: 2016.12.14 Location: Frankfurt/Main, Wilhelm-Fay Straße 54 Start: 11:00h End: 16:00h
- T-System (Christian Simmendinger)
- LRZ (Christoph Bernauer)
- ITWM (Rui Machado, Daniel Grünewald)
- ZIH Dresden (Olaf Krzikalla)
- DLR (Georg Geiser)
- IMEC (Tom van der Aa)
11.00-11.30 Attendees, Quorum, next Forum meeting, Actions Items
11.30-12.00 Errata Proposals
12.45-13.15 Errata Proposals cont.
13.45-14.30 General text proposal
15.00-16.00 GASPI Context, GASPI Standard Library
Frankfurt, June 21
- T-Systems: Put a hyperref package to the specification in order to allow easy browsing
- ZIH: Errata proposal for local completion in the gaspi notify specificication. (Remove double “the” typo)
- All: What are the functions that could be removed or extracted to a standard library in order to shrink the specification
- All: Is GASPI ready for multithreaded communication on ~1000 Threads? Start thinking about possible limitations / solutions.
- T-Systems: Errata proposal for gaspi_write_list_notify: segment id for notification is missing in signature
- T-Systems: Render the version number instead of ‘\today’ into the document.
- ZIH Dresden: Prepare proposal for clear (or no) definitions of “blocking”, “non-blocking”
- ITWM: Make a proposal for clarification that GASPI_ERROR shall not be used in equality comparison. Thinking about the concecpt of a GASPI_ERROR itself
- T-Systems: Prepare a proposal for a change of statute that future proposals shall include a merge request. github voting technology could be used for accepting merge requests after modifications to a proposal have been discussed and accepted during a meeting.
- T-Systems: Present good example of how to use multiple queues (why should we have queues) -> not yet done but should be about local completion. Present an example how to use double buffered queues in order to have a no-op.
- T-Systems: Write prototype for a library/wrapper that demonstrates how to optimize bandwidth
- T-Systems: Talk to KTH to stop running the second/old/outdated forum site has been done
- T-Systems: Tutorial at LRZ: Christian has contacted Mr. Bode. It will probably happen during the next extreme scale workshop held at second quarter 2017
- T-Systems: Why do we need queues? -> Queues are needed in order to be able to provide a check for local completion
- T-Systems: Provided a non-GPL header, the pure GASPI header, generated from a script which extracts the header from the specification tex.
Errata proposal: GASPI_ERROR_END_OF_QUEUE
CS: There is a potential race condition in multithreaded communication when the queue has
less then #threads entries left -> GASPI_ERROR_END_OF_QUEUE proposal
Georg: In the proposal there is a typo: gaspi_wait_queue needs to be replaced by gaspi_wait
Tom: What is the relation to the timeout? The timeout only .
Daniel: is there still a reason for the gaspi_queue_size? No, it is not.
Olaf: it still makes sense to have gaspi_queue_size_max in order to be able to estimate the amount of required queues for a given problem.
Christian: Use gaspi_config_get in order to request max parameters. Throw away explicit get functions. One could put them in a standard library.
Rui: ITWM is missing an adaption of the paragraphs describing the behaviour in case of a full queue in all of the write, read calls containing such a section.
Olaf: Is GASPI_ERROR_END_OF_QUEUE really an error? No. It should be an additional return value. In case of an error we do not know what to do. However, for end of queue, we perfectly know that. So, it is not an error. Rename it to GASPI_QUEUE_FULL and treat it as a special return value.
Vote for rejection of original proposal: unanimously rejected
Vote for the changed and clarified proposal: unanimously accepted
Errata proposal: Read / Write ordering
CS: gaspi_write_notify, gaspi_write_list_notify, gaspi_read_notify and gaspi_read_list_notify can be considered as unordered relative to each other and also as unordered relative ot the function calls gaspi_read and gaspi_write. -> Replace “preceding” communication requests with “other” communciation requests accordingly.
Vote: unanimously accepted
Errata proposal: clarification of global lock example
Olaf: global_try_lock returns and error if you don’t get the lock -> it is not an error. It also returns an error if a real error occured. One cannot decide whether one could not get a lock or whether an error occured.
- Substitute SUCCESS_OR_RETURN by SUCCES_OR_DIE
- Remove timeout from signature and use GASPI_BLOCK internally
- Introduce specific return values: LOCK_ACQUIRED, NO_LOCK_ACQUIRED
Voting: Unanimously rejected
ITWM: global_unlock not symmetric anymore
- remove timeout also here
- use SUCCESS_OR_DIE instead of SUCCESS_OR_RETURN
- return value void
Olaf: add assertion of lock value.
Voting: unanimously accepted
General text proposal: gaspi_read_list_notify (First reading)
CS: Extend gaspi_read_list with a notification on the local side
Rui: The specification of the target segment for the notification is missing. Actually, there are also inconsistencies in the fortran / c / interface for write_list_notify where you don’t have a target segment specification for the notification in the parameter section / c interface
- The second reading of the proposal must include the target segment for notification.
- Need an errata proposal to fix write_list_notify
CS: We need more than one implementation to get more reception: people get uneasy on depending on a single implementation from a single institution. In order to achieve that, we need to shrink the interface to a real minimum of basis functions with which all other functions can be implemented on top of that. All convenience functions can be outsourced to a standard library which is using these basis functions
Rui: The goal always has been to keep the interface as small as possible. Except for getter/setter functions there isn’t much we can remove. We should be specific: check and make a list of potential candidates to be removed. A standard library has been in discussion for long time and again we need to be more specific of what functionality should go in there. Convenience functions yes (although I only see gaspi_segment_create) and the often requested collectives. Instead of increasing the reception on the implementor side, we should increase the reception in the user space.
Tom: Eric Petit has implemented a tree based reduction on top of Gaspi. Perhaps it could be published on the website and used as a first piece of a standard library. CS: is there anything which spoils GASPI scalability of communication on up to 5000 cores per process? -> think about potentiel bottlenecks and their solution.
Olaf: For instance, on KNL locks are deadly -> Try to avoid locks as much as possible.