Re: [Mpi3-rma] non-contiguous support in RMA & one-sided pack/unpack (?)
This is exactly why I think it is critical to think of the dichotomy as fast/slow instead of raw/complex. What you are proposing is a raw/complex dichotomy that goes faster when you use the raw interface than when you use the complex interface. But, the point of the raw mode is to offer significantly more speed to those who don't need the complex features. Thus, anything that doesn't exist in the raw mode interface will inherently be slower. If hardware adds a feature that isn't part of the raw mode interface, you have to either change MPI (not something we want to do a lot) or use the complex (aka slow) interface. As I pointed out earlier, the flip side of that coin is that if you put something "too hard" in the raw/fast interface, you risk making the fast interface slow. Keith
-----Original Message----- From: [email protected] [mailto:mpi3-rma- [email protected]] On Behalf Of Jeff Hammond Sent: Wednesday, September 16, 2009 3:42 PM To: MPI 3.0 Remote Memory Access working group Subject: Re: [Mpi3-rma] non-contiguous support in RMA & one-sided pack/unpack (?)
Keith,
I don't really think in terms of the fast/slow dichotomy, rather the raw/complex one. This is really meant to align with the philosophical dichotomy of MPI users - one group is using less than 10 functions and are perfectly happy with the results, while another group's needs push the limits of the standard, or at least require a non-trivial subset of the MPI standard's calls.
Raw_xfer should be simple and fast when that is all that is needed, but should not be the only available option. I contend, as someone who has not implemented nor will most likely not implement the MPI standard, but only as a rather heavy user of GA/ARMCI, that excluding Raw_xfer from the standard will prevent optimal performance on some systems, particularly the ones at the bleeding-edge.
I would love it if the standard branch-predicted future hardware and defined RMA accordingly, but I think that the raw/complex dichotomy is appropriate for what we know to be true now and that it in no way handicaps the standard from responding to future hardware as new capability is realized. The only way I would modify the Raw_xfer based upon my optimistic vision of the future of hardware would be to add a stride argument such that if someone develops hardware support for strided transfer over primitive types, then they can seamlessly become part of Raw_xfer.
Dick,
If you believe that the MPI implementations can do all the datatype decoding required in the general purpose with ~very~ little overhead, then Raw_xfer is not needed. However, I'm pessimistic that this will be possible. Local completion of contiguous transfers on BlueGene/P occurs in less than 10000 clock cycles. If the datatype cache is large and is not prefetched to L1 cache, its a challenge to decode the general case on that time-scale. I echo Keith's most recent post along these lines.
I appreciate Vinod's comments that BlueGene/P is a special case, but few vendors are talking about clock speeds more than 5 times what PPC450 runs at (0.85 GHz) and neither have I heard that we're going to get to exascale with slower latency interconnects, so I hope 2-3 microsecond latency becomes available on more machines than just BlueGene/X varieties. Given that intra-node memory pipelines aren't getting any faster, any MPI function call that has to go to a database, i.e. for cached datatypes, is going to take a performance hit now and into the future.
Best,
Jeff
On Wed, Sep 16, 2009 at 3:43 PM, Underwood, Keith D <[email protected]> wrote:
It is a very fine line to walk. If you have "fast" and "slow" functions that map to what is possible today, then you have a standard which is nominally fast out of the box; however, you are then stuck having to implement EVERYTHING in the "slow" function before you can move that into the "fast" category for users.
In contrast, if you branch predict (a little) what networks will be able to do in a couple of years and push on the vendors a bit in the definition of the "fast" interface, then the day the standard ships, it may not be of optimal speed, but you have hope of getting a slightly broader set of functionality in the fast category.
Since I don't really expect that the new RMA ops will be available the same day as the first printed standard, I tend to favor the latter category. Let's define a subset that is a good subset, rather than simply the subset that exists in DCMF and Verbs.
Keith
-----Original Message----- From: [email protected] [mailto:mpi3-rma- [email protected]] On Behalf Of Jeff Hammond Sent: Wednesday, September 16, 2009 2:37 PM To: MPI 3.0 Remote Memory Access working group Subject: Re: [Mpi3-rma] non-contiguous support in RMA & one-sided pack/unpack (?)
I was having the same thoughts during my hideous commute today. Clearly, future systems may have very smart NICs or OSs which seamlessly manage communication co-processors to support, among other things, on-the-fly compression of non-contiguous buffers. However, until some vendor starts taking seriously about this, I don't think it is good idea to ignore the present hardware support of RMA and create a standard which can perform well when the first MPI-3 libraries become available.
I'm not sure I understand all this talk about assertions upon initialization, since I was hoping that Raw_xfer and general-purpose xfer (Gen_xfer) would both always be available but that the former would be much faster in certain contexts, since implementing the Raw version close to hardware doesn't require much work and would not interfere with how Gen_xfer operates. Am I missing something?
When the day finally comes that NICs can do Gen_xfer for a complex non-contiguous datatype as fast as Raw_xfer, would not the standard still be sufficient, since implementers would just produce a faster Gen_xfer and Raw_xfer would fade into the background?
Raw_xfer is a pragmatic response to the present and near-future state of HPC systems rather than some sort of marquee feature to tape to the metaphorical refrigerator door.
Best,
Jeff
On Wed, Sep 16, 2009 at 3:14 PM, Underwood, Keith D <[email protected]> wrote:
But we have to be very careful here. We don't want to overly constrain what can be thought of as "fast". For example, I think it is perfectly reasonable to implement accumulate on a NIC. Just because it doesn't exist today doesn't mean that it shouldn't be part of the "fast" MPI call.
Now, datatype conversion. it is nominally possible that a NIC could do datatype conversion - just like it is nominally possible to for a NIC to be hooked to a Rube Goldberg device to implement MPI_Make_Breakfast ;-)
Anyway, the point is that we need to be forward looking in defining "fast" and "slow", not backward looking.
Keith
From: [email protected] [mailto:[email protected]] On Behalf Of Richard Treumann Sent: Wednesday, September 16, 2009 2:06 PM To: MPI 3.0 Remote Memory Access working group Subject: Re: [Mpi3-rma] non-contiguous support in RMA & one-sided pack/unpack (?)
BINGO Jeff
We might also remove the datatype argument and twin count arguments from MPI_RMA_Raw_xfer just to eliminate the expectation that basic put/get do datatype conversions when origin and target are on heterogeneous nodes. There would be a single "count" argument and it represents the number of contiguous bytes to be transferred.
The assertion would be that there is no use of complex RMA. It would give the implementation the option to leave its software agent dormant. Note that having this assertion as an option for MPI_Init_asserted does not allow an MPI implementation to avoid having an agent available. An application that does not use the assertion can count on the agent being ready for any call to "full baked" RMA.
Dick
Dick Treumann - MPI Team IBM Systems & Technology Group Dept X2ZA / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601 Tele (845) 433-7846 Fax (845) 433-8363
[email protected] wrote on 09/16/2009 03:43:15 PM:
[image removed]
Re: [Mpi3-rma] non-contiguous support in RMA & one-sided pack/unpack (?)
Jeff Hammond
to:
MPI 3.0 Remote Memory Access working group
09/16/2009 03:44 PM
Sent by:
Please respond to "MPI 3.0 Remote Memory Access working group"
I think that there is a need for two interfaces; one which is a portable interface to the low-level truly one-sided bulk transfer operation and another which is completely general and is permitted to do operations which require remote agency.
For example, I am aware of no NIC which can do accumulate on its own, hence RMA_ACC_SUM and related operations require remote agency, and thus this category of RMA operations are not truly one-sided.
Thus the standard might support two xfer calls:
MPI_RMA_Raw_xfer(origin_addr, origin_count, origin_datatype, target_mem, target_disp, target_count , target_rank, request)
which is exclusively for transferring contiguous bytes from one place to another, i.e. does raw put/get only, and the second, which has been described already, which handles the general case, including accumulation, non-contiguous and other complex operations.
The distinction over remote agency is extremely important from a implementation perspective since contiguous put/get operations can be performed in a fully asynchronous non-interrupting way with a variety of interconnects, and thus exposing this procedure in the MPI standard will allow for very efficient implementations on some systems. It should also encourage MPI users to think about their RMA needs and how they might restructure their code to take advantage of the faster flavor of xfer when doing so requires little modification.
Jeff
On Wed, Sep 16, 2009 at 1:49 PM, Vinod tipparaju <[email protected]> wrote:
My argument is that any RMA depends on a call at the origin being able to trigger activity at the target. Modern RMA hardware has the hooksto do the remote side of MPI_Fast_RMA_xfer() efficiently based on a call at the origin. Because these hooks are in the hardware they are simply there. They do not use the CPU or hurt performance of things that do use the CPU.
I read this as an argument that says two interfaces are not necessary. Having application author promise (during init) it will not do anything that needs an agent is certainly useful. Particularly when, as you state, "having this agent standing by hurts general performance". The things that potentially cannot be done without an agent (technically, everything but atomics could be done with out need for any agents)are users choice through explicit usage. Users choses these attributes being aware of their cost hence they can indicate that they will not use them ahead of time when they don't use them. I have repeatedly considered dropping the atomicity attribute, I am unable to because it makes programming (and thinking) so much easier for many applications. Vinod.
________________________________ To: [email protected] From: [email protected] Date: Wed, 16 Sep 2009 14:18:15 -0400 Subject: Re: [Mpi3-rma] non-contiguous support in RMA & one- sided pack/unpack (?)
The assertion could then be: MPI_NO_SLOW_RMA (also a bit tongue in cheek)
My argument is that any RMA depends on a call at the origin being able to trigger activity at the target. Modern RMA hardware has the hooks to do the remote side of MPI_Fast_RMA_xfer() efficiently based on a call at the origin. Because these hooks are in the hardware they are simply there. They do not use the CPU or hurt performance of things that do use the CPU.
RMA hardware may not have the hooks to do the target side of any arbitrary MPI_Slow_RMA_xfer(). As a result, support for the more complex RMA_xfer may require a wake-able software agent (thread maybe) to be standing by at all tasks just because they may become target of a Slow_RMA_xfer.
If having this agent standing by hurts general performance of MPI applications that will never make a call to Slow_RMA_xfer, why not let the applications author promise up front "I have no need of this agent."
An MPI implementation that can support Slow_RMA_xfer with no extra costs (send/recv latency, memory, packet interrupts, CPU contention) will simply ignore the assertion.
BTW - I just took a look at the broad proposal and it may contain several things that cannot be done without a wake-able remote software agent. That argues for Keith's idea of an RMA operation which closely matches what RMA hardware does and a second one that brings along all the bells andwhistles. Maybe the assertion for an application that only uses the basic RMA call or uses no RMA at all could be MPI_NO_KITCHEN_SINK (even more tongue in cheek).
Dick
Dick Treumann - MPI Team IBM Systems & Technology Group Dept X2ZA / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601 Tele (845) 433-7846 Fax (845) 433-8363
[email protected] wrote on 09/16/2009 01:08:51 PM:
[image removed]
Re: [Mpi3-rma] non-contiguous support in RMA & one-sided pack/unpack (?)
Underwood, Keith D
to:
MPI 3.0 Remote Memory Access working group
09/16/2009 01:09 PM
Sent by:
Please respond to "MPI 3.0 Remote Memory Access working group"
But, going back to Bill's point: performance across a range of platforms is key. While you can't have a function for every usage (well, you can, but it would get cumbersome at some point), it may be important to have a few levels of specialization in the API. E.g. you could have two variants:
MPI_Fast_RMA_xfer(): no data types, no communicators, etc. MPI_Slow_RMA_xfer(): include the kitchen sink.
Yes, the naming is a little tongue in cheek ;-)
Keith
<snip>
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
-- Jeff Hammond Argonne Leadership Computing Facility [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond http://home.uchicago.edu/~jhammond/
_______________________________________________ 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
-- Jeff Hammond Argonne Leadership Computing Facility [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond http://home.uchicago.edu/~jhammond/
_______________________________________________ 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
-- Jeff Hammond Argonne Leadership Computing Facility [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond http://home.uchicago.edu/~jhammond/
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
participants (1)
-
Underwood, Keith D