Re: [OMPI devel] RFC: job size info in OPAL
Fair enough - yeah, that is an issue I've been avoiding :-) On Jul 31, 2014, at 9:14 AM, Nathan Hjelm <hje...@lanl.gov> wrote: > > This approach will work now but we need to start thinking about how we > want to support multiple simultaneous btl users. Does each user call > add_procs with a single module (or set of modules) or does each user > call btl_component_init and get their own module? If we do the latter > then it might make sense to add a max_procs argument to the > btl_component_init. Keep in mind we need to change the > btl_component_init interface anyway because the threading arguments no > longer make sense in their current form. > > -Nathan > > On Thu, Jul 31, 2014 at 09:04:09AM -0700, Ralph Castain wrote: >> Like I said, why don't we just do the following: >> >>> I'd like to suggest an alternative solution. A BTL can exploit whatever >>> data it wants, but should first test if the data is available. If the data >>> is *required*, then the BTL gracefully disqualifies itself. If the data is >>> *desirable* for optimization, then the BTL writer (if they choose) can >>> include an alternate path that doesn't do the optimization if the data >>> isn't available. >> >> Seems like this should resolve the disagreement in a way that meets >> everyone's need. It basically is an attribute approach, but not requiring >> modification of the BTL interface. >> >> >> On Jul 31, 2014, at 8:26 AM, Pritchard Jr., Howard <howa...@lanl.gov> wrote: >> >>> Hi George, >>> >>> The ompi_process_info.num_procs thing that seems to have been an object >>> of some contention yesterday. >>> >>> The ugni use of this is cloned off of the way I designed the mpich netmod. >>> Leveraging off size of the job was an easy way to scale the mailbox size. >>> >>> If I'd been asked to have the netmod work in a context like it appears we >>> may want to be eventually using BTLs - not just within ompi but for other >>> things, I'd have worked with Darius (if still in mpich world) on changing >>> the netmod initialization >>> method to allow for an optional attributes struct to be passed into the >>> init >>> method to give hints about how many connections may need to be established, >>> etc. >>> >>> For the GNI BTL - the way its currently designed - if you are only expecting >>> to use it for a limited number of connections, then you want to initialize >>> for big mailboxes (IBer's can think many large buffers posted as RX WQEs). >>> But for very large jobs, with possibly highly connected communication >>> pattern, >>> you want very small mailboxes. >>> >>> Howard >>> >>> >>> -Original Message- >>> From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of George Bosilca >>> Sent: Thursday, July 31, 2014 9:09 AM >>> To: Open MPI Developers >>> Subject: Re: [OMPI devel] RFC: job size info in OPAL >>> >>> What is your definition of "global job size"? >>> >>> George. >>> >>> On Jul 31, 2014, at 11:06 , Pritchard Jr., Howard <howa...@lanl.gov> wrote: >>> >>>> Hi Folks, >>>> >>>> I think given the way we want to use the btl's in lower levels like >>>> opal, it is pretty disgusting for a btl to need to figure out on its >>>> own something like a "global job size". That's not its business. >>>> Can't we add some attributes to the component's initialization method >>>> that provides hints for how to allocate resources it needs to provide its >>>> functionality? >>>> >>>> I'll see if there's something clever that can be done in ugni for now. >>>> I can always add in a hack to probe the apps placement info file and >>>> scale the smsg blocks by number of nodes rather than number of ranks. >>>> >>>> Howard >>>> >>>> >>>> -Original Message- >>>> From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan >>>> Hjelm >>>> Sent: Thursday, July 31, 2014 8:58 AM >>>> To: Open MPI Developers >>>> Subject: Re: [OMPI devel] RFC: job size info in OPAL >>>> >>>> >>>> +2^1000 >>>> >>>> This information is absolutely necessary at this point. If someone has a >>>> better solution they can provide it as an altern
Re: [OMPI devel] RFC: job size info in OPAL
This approach will work now but we need to start thinking about how we want to support multiple simultaneous btl users. Does each user call add_procs with a single module (or set of modules) or does each user call btl_component_init and get their own module? If we do the latter then it might make sense to add a max_procs argument to the btl_component_init. Keep in mind we need to change the btl_component_init interface anyway because the threading arguments no longer make sense in their current form. -Nathan On Thu, Jul 31, 2014 at 09:04:09AM -0700, Ralph Castain wrote: > Like I said, why don't we just do the following: > > > I'd like to suggest an alternative solution. A BTL can exploit whatever > > data it wants, but should first test if the data is available. If the data > > is *required*, then the BTL gracefully disqualifies itself. If the data is > > *desirable* for optimization, then the BTL writer (if they choose) can > > include an alternate path that doesn't do the optimization if the data > > isn't available. > > Seems like this should resolve the disagreement in a way that meets > everyone's need. It basically is an attribute approach, but not requiring > modification of the BTL interface. > > > On Jul 31, 2014, at 8:26 AM, Pritchard Jr., Howard <howa...@lanl.gov> wrote: > > > Hi George, > > > > The ompi_process_info.num_procs thing that seems to have been an object > > of some contention yesterday. > > > > The ugni use of this is cloned off of the way I designed the mpich netmod. > > Leveraging off size of the job was an easy way to scale the mailbox size. > > > > If I'd been asked to have the netmod work in a context like it appears we > > may want to be eventually using BTLs - not just within ompi but for other > > things, I'd have worked with Darius (if still in mpich world) on changing > > the netmod initialization > > method to allow for an optional attributes struct to be passed into the > > init > > method to give hints about how many connections may need to be established, > > etc. > > > > For the GNI BTL - the way its currently designed - if you are only expecting > > to use it for a limited number of connections, then you want to initialize > > for big mailboxes (IBer's can think many large buffers posted as RX WQEs). > > But for very large jobs, with possibly highly connected communication > > pattern, > > you want very small mailboxes. > > > > Howard > > > > > > -Original Message- > > From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of George Bosilca > > Sent: Thursday, July 31, 2014 9:09 AM > > To: Open MPI Developers > > Subject: Re: [OMPI devel] RFC: job size info in OPAL > > > > What is your definition of "global job size"? > > > > George. > > > > On Jul 31, 2014, at 11:06 , Pritchard Jr., Howard <howa...@lanl.gov> wrote: > > > >> Hi Folks, > >> > >> I think given the way we want to use the btl's in lower levels like > >> opal, it is pretty disgusting for a btl to need to figure out on its > >> own something like a "global job size". That's not its business. > >> Can't we add some attributes to the component's initialization method > >> that provides hints for how to allocate resources it needs to provide its > >> functionality? > >> > >> I'll see if there's something clever that can be done in ugni for now. > >> I can always add in a hack to probe the apps placement info file and > >> scale the smsg blocks by number of nodes rather than number of ranks. > >> > >> Howard > >> > >> > >> -Original Message- > >> From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan > >> Hjelm > >> Sent: Thursday, July 31, 2014 8:58 AM > >> To: Open MPI Developers > >> Subject: Re: [OMPI devel] RFC: job size info in OPAL > >> > >> > >> +2^1000 > >> > >> This information is absolutely necessary at this point. If someone has a > >> better solution they can provide it as an alternative RFC. Until then this > >> is how it should be done... Otherwise we loose uGNI support on the trunk. > >> Because we ARE NOT going to remove the mailbox size optimization. > >> > >> -Nathan > >> > >> On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: > >>> WHAT: Should we make the job size (i.e., initial number of procs) > >>> available
Re: [OMPI devel] RFC: job size info in OPAL
Like I said, why don't we just do the following: > I'd like to suggest an alternative solution. A BTL can exploit whatever data > it wants, but should first test if the data is available. If the data is > *required*, then the BTL gracefully disqualifies itself. If the data is > *desirable* for optimization, then the BTL writer (if they choose) can > include an alternate path that doesn't do the optimization if the data isn't > available. Seems like this should resolve the disagreement in a way that meets everyone's need. It basically is an attribute approach, but not requiring modification of the BTL interface. On Jul 31, 2014, at 8:26 AM, Pritchard Jr., Howard <howa...@lanl.gov> wrote: > Hi George, > > The ompi_process_info.num_procs thing that seems to have been an object > of some contention yesterday. > > The ugni use of this is cloned off of the way I designed the mpich netmod. > Leveraging off size of the job was an easy way to scale the mailbox size. > > If I'd been asked to have the netmod work in a context like it appears we > may want to be eventually using BTLs - not just within ompi but for other > things, I'd have worked with Darius (if still in mpich world) on changing the > netmod initialization > method to allow for an optional attributes struct to be passed into the init > method to give hints about how many connections may need to be established, > etc. > > For the GNI BTL - the way its currently designed - if you are only expecting > to use it for a limited number of connections, then you want to initialize > for big mailboxes (IBer's can think many large buffers posted as RX WQEs). > But for very large jobs, with possibly highly connected communication pattern, > you want very small mailboxes. > > Howard > > > -Original Message- > From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of George Bosilca > Sent: Thursday, July 31, 2014 9:09 AM > To: Open MPI Developers > Subject: Re: [OMPI devel] RFC: job size info in OPAL > > What is your definition of "global job size"? > > George. > > On Jul 31, 2014, at 11:06 , Pritchard Jr., Howard <howa...@lanl.gov> wrote: > >> Hi Folks, >> >> I think given the way we want to use the btl's in lower levels like >> opal, it is pretty disgusting for a btl to need to figure out on its >> own something like a "global job size". That's not its business. >> Can't we add some attributes to the component's initialization method >> that provides hints for how to allocate resources it needs to provide its >> functionality? >> >> I'll see if there's something clever that can be done in ugni for now. >> I can always add in a hack to probe the apps placement info file and >> scale the smsg blocks by number of nodes rather than number of ranks. >> >> Howard >> >> >> -Original Message- >> From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan >> Hjelm >> Sent: Thursday, July 31, 2014 8:58 AM >> To: Open MPI Developers >> Subject: Re: [OMPI devel] RFC: job size info in OPAL >> >> >> +2^1000 >> >> This information is absolutely necessary at this point. If someone has a >> better solution they can provide it as an alternative RFC. Until then this >> is how it should be done... Otherwise we loose uGNI support on the trunk. >> Because we ARE NOT going to remove the mailbox size optimization. >> >> -Nathan >> >> On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: >>> WHAT: Should we make the job size (i.e., initial number of procs) available >>> in OPAL? >>> >>> WHY: At least 2 BTLs are using this info (*more below) >>> >>> WHERE: usnic and ugni >>> >>> TIMEOUT: there's already been some inflammatory emails about this; >>> let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 >>> >>> MORE DETAIL: >>> >>> This is an open question. We *have* the information at the time that the >>> BTLs are initialized: do we allow that information to go down to OPAL? >>> >>> Ralph added this info down in OPAL in r32355, but George reverted it in >>> r32361. >>> >>> Points for: YES, WE SHOULD >>> +++ 2 BTLs were using it (usinc, ugni) Other RTE job-related info are >>> +++ already in OPAL (num local ranks, local rank) >>> >>> Points for: NO, WE SHOULD NOT >>> --- What exactly is this number (e.g., num currently-connected procs?), and >>> when is it updated? &
Re: [OMPI devel] RFC: job size info in OPAL
I do not like the fact that add_procs is called with every proc in the MPI_COMM_WORLD. That needs to change, so, I will not rely on the number of procs being added being the same as the world or universe size. -Nathan On Thu, Jul 31, 2014 at 09:22:00AM -0600, George Bosilca wrote: >I definitively think you misunderstood this scope of this RFC. The >information that is so important to you to configure the mailbox size is >available to you when you need it. This information is made available by >the PML through the call to add_procs, which comes with all the procs in >the MPI_COMM_WORLD. So, ugni doesn't need anything more than it is >available today. [This is of course under the assumption that someone >clean the BTL and remove the usage of MPI_COMM_WORLD.] > >The real scope of this RFC is to move this information before that in >order to allow the BTLs to have access to some possible number of >processes between the call to btl_open and the call to btl_all_proc (in >other words during btl_init). > > George. > >PS: here is the patch that fixes all issues in ugni. > >On Jul 31, 2014, at 10:58 , Nathan Hjelmwrote: > >> >> +2^1000 >> >> This information is absolutely necessary at this point. If someone has a >> better solution they can provide it as an alternative RFC. Until then >> this is how it should be done... Otherwise we loose uGNI support on the >> trunk. Because we ARE NOT going to remove the mailbox size optimization. >> >> -Nathan >> >> On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: >>> WHAT: Should we make the job size (i.e., initial number of procs) >available in OPAL? >>> >>> WHY: At least 2 BTLs are using this info (*more below) >>> >>> WHERE: usnic and ugni >>> >>> TIMEOUT: there's already been some inflammatory emails about this; >let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 >>> >>> MORE DETAIL: >>> >>> This is an open question. We *have* the information at the time that >the BTLs are initialized: do we allow that information to go down to OPAL? >>> >>> Ralph added this info down in OPAL in r32355, but George reverted it in >r32361. >>> >>> Points for: YES, WE SHOULD >>> +++ 2 BTLs were using it (usinc, ugni) >>> +++ Other RTE job-related info are already in OPAL (num local ranks, >local rank) >>> >>> Points for: NO, WE SHOULD NOT >>> --- What exactly is this number (e.g., num currently-connected procs?), >and when is it updated? >>> --- We need to precisely delineate what belongs in OPAL vs. above-OPAL >>> >>> FWIW: here's how ompi_process_info.num_procs was used before the BTL >move down to OPAL: >>> >>> - usnic: for a minor latency optimization / sizing of a shared receive >buffer queue length, and for the initial size of a peer lookup hash >>> - ugni: to determine the size of the per-peer buffers used for >send/recv communication >>> >>> -- >>> Jeff Squyres >>> jsquy...@cisco.com >>> For corporate legal information go to: >http://www.cisco.com/web/about/doing_business/legal/cri/ >>> >>> ___ >>> devel mailing list >>> de...@open-mpi.org >>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >>> Link to this post: >http://www.open-mpi.org/community/lists/devel/2014/07/15373.php >> ___ >> devel mailing list >> de...@open-mpi.org >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >> Link to this post: >http://www.open-mpi.org/community/lists/devel/2014/07/15394.php > >___ >devel mailing list >de...@open-mpi.org >Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >Link to this post: >http://www.open-mpi.org/community/lists/devel/2014/07/15399.php pgpo6WjkLZPnT.pgp Description: PGP signature
Re: [OMPI devel] RFC: job size info in OPAL
Hi George, The ompi_process_info.num_procs thing that seems to have been an object of some contention yesterday. The ugni use of this is cloned off of the way I designed the mpich netmod. Leveraging off size of the job was an easy way to scale the mailbox size. If I'd been asked to have the netmod work in a context like it appears we may want to be eventually using BTLs - not just within ompi but for other things, I'd have worked with Darius (if still in mpich world) on changing the netmod initialization method to allow for an optional attributes struct to be passed into the init method to give hints about how many connections may need to be established, etc. For the GNI BTL - the way its currently designed - if you are only expecting to use it for a limited number of connections, then you want to initialize for big mailboxes (IBer's can think many large buffers posted as RX WQEs). But for very large jobs, with possibly highly connected communication pattern, you want very small mailboxes. Howard -Original Message- From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of George Bosilca Sent: Thursday, July 31, 2014 9:09 AM To: Open MPI Developers Subject: Re: [OMPI devel] RFC: job size info in OPAL What is your definition of "global job size"? George. On Jul 31, 2014, at 11:06 , Pritchard Jr., Howard <howa...@lanl.gov> wrote: > Hi Folks, > > I think given the way we want to use the btl's in lower levels like > opal, it is pretty disgusting for a btl to need to figure out on its > own something like a "global job size". That's not its business. > Can't we add some attributes to the component's initialization method > that provides hints for how to allocate resources it needs to provide its > functionality? > > I'll see if there's something clever that can be done in ugni for now. > I can always add in a hack to probe the apps placement info file and > scale the smsg blocks by number of nodes rather than number of ranks. > > Howard > > > -Original Message- > From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan > Hjelm > Sent: Thursday, July 31, 2014 8:58 AM > To: Open MPI Developers > Subject: Re: [OMPI devel] RFC: job size info in OPAL > > > +2^1000 > > This information is absolutely necessary at this point. If someone has a > better solution they can provide it as an alternative RFC. Until then this is > how it should be done... Otherwise we loose uGNI support on the trunk. > Because we ARE NOT going to remove the mailbox size optimization. > > -Nathan > > On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: >> WHAT: Should we make the job size (i.e., initial number of procs) available >> in OPAL? >> >> WHY: At least 2 BTLs are using this info (*more below) >> >> WHERE: usnic and ugni >> >> TIMEOUT: there's already been some inflammatory emails about this; >> let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 >> >> MORE DETAIL: >> >> This is an open question. We *have* the information at the time that the >> BTLs are initialized: do we allow that information to go down to OPAL? >> >> Ralph added this info down in OPAL in r32355, but George reverted it in >> r32361. >> >> Points for: YES, WE SHOULD >> +++ 2 BTLs were using it (usinc, ugni) Other RTE job-related info are >> +++ already in OPAL (num local ranks, local rank) >> >> Points for: NO, WE SHOULD NOT >> --- What exactly is this number (e.g., num currently-connected procs?), and >> when is it updated? >> --- We need to precisely delineate what belongs in OPAL vs. >> above-OPAL >> >> FWIW: here's how ompi_process_info.num_procs was used before the BTL move >> down to OPAL: >> >> - usnic: for a minor latency optimization / sizing of a shared >> receive buffer queue length, and for the initial size of a peer >> lookup hash >> - ugni: to determine the size of the per-peer buffers used for >> send/recv communication >> >> -- >> Jeff Squyres >> jsquy...@cisco.com >> For corporate legal information go to: >> http://www.cisco.com/web/about/doing_business/legal/cri/ >> >> ___ >> devel mailing list >> de...@open-mpi.org >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >> Link to this post: >> http://www.open-mpi.org/community/lists/devel/2014/07/15373.php > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15395.php ___ devel mailing list de...@open-mpi.org Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel Link to this post: http://www.open-mpi.org/community/lists/devel/2014/07/15396.php
Re: [OMPI devel] RFC: job size info in OPAL
I definitively think you misunderstood this scope of this RFC. The information that is so important to you to configure the mailbox size is available to you when you need it. This information is made available by the PML through the call to add_procs, which comes with all the procs in the MPI_COMM_WORLD. So, ugni doesn’t need anything more than it is available today. [This is of course under the assumption that someone clean the BTL and remove the usage of MPI_COMM_WORLD.] The real scope of this RFC is to move this information before that in order to allow the BTLs to have access to some possible number of processes between the call to btl_open and the call to btl_all_proc (in other words during btl_init). George. PS: here is the patch that fixes all issues in ugni. ugni.patch Description: Binary data On Jul 31, 2014, at 10:58 , Nathan Hjelmwrote: > > +2^1000 > > This information is absolutely necessary at this point. If someone has a > better solution they can provide it as an alternative RFC. Until then > this is how it should be done... Otherwise we loose uGNI support on the > trunk. Because we ARE NOT going to remove the mailbox size optimization. > > -Nathan > > On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: >> WHAT: Should we make the job size (i.e., initial number of procs) available >> in OPAL? >> >> WHY: At least 2 BTLs are using this info (*more below) >> >> WHERE: usnic and ugni >> >> TIMEOUT: there's already been some inflammatory emails about this; let's >> discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 >> >> MORE DETAIL: >> >> This is an open question. We *have* the information at the time that the >> BTLs are initialized: do we allow that information to go down to OPAL? >> >> Ralph added this info down in OPAL in r32355, but George reverted it in >> r32361. >> >> Points for: YES, WE SHOULD >> +++ 2 BTLs were using it (usinc, ugni) >> +++ Other RTE job-related info are already in OPAL (num local ranks, local >> rank) >> >> Points for: NO, WE SHOULD NOT >> --- What exactly is this number (e.g., num currently-connected procs?), and >> when is it updated? >> --- We need to precisely delineate what belongs in OPAL vs. above-OPAL >> >> FWIW: here's how ompi_process_info.num_procs was used before the BTL move >> down to OPAL: >> >> - usnic: for a minor latency optimization / sizing of a shared receive >> buffer queue length, and for the initial size of a peer lookup hash >> - ugni: to determine the size of the per-peer buffers used for send/recv >> communication >> >> -- >> Jeff Squyres >> jsquy...@cisco.com >> For corporate legal information go to: >> http://www.cisco.com/web/about/doing_business/legal/cri/ >> >> ___ >> devel mailing list >> de...@open-mpi.org >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >> Link to this post: >> http://www.open-mpi.org/community/lists/devel/2014/07/15373.php > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15394.php
Re: [OMPI devel] RFC: job size info in OPAL
The maximum number of peer processes that may be added over the course of the job will suffice. So either the world or universe size. This is a reasonable piece of information to expect the upper layers to provide to the communication layer. And the impact of providing this information is no less intrusive than providing information like the number of local ranks. -Nathan On Thu, Jul 31, 2014 at 11:09:24AM -0400, George Bosilca wrote: > What is your definition of “global job size”? > > George. > > On Jul 31, 2014, at 11:06 , Pritchard Jr., Howard <howa...@lanl.gov> wrote: > > > Hi Folks, > > > > I think given the way we want to use the btl's in lower levels like opal, > > it is pretty disgusting for a btl to need to figure out on its own something > > like a "global job size". That's not its business. Can't we add some > > attributes > > to the component's initialization method that provides hints for how to > > allocate resources it needs to provide its functionality? > > > > I'll see if there's something clever that can be done in ugni for now. > > I can always add in a hack to probe the apps placement info file and > > scale the smsg blocks by number of nodes rather than number of ranks. > > > > Howard > > > > > > -Original Message- > > From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan Hjelm > > Sent: Thursday, July 31, 2014 8:58 AM > > To: Open MPI Developers > > Subject: Re: [OMPI devel] RFC: job size info in OPAL > > > > > > +2^1000 > > > > This information is absolutely necessary at this point. If someone has a > > better solution they can provide it as an alternative RFC. Until then this > > is how it should be done... Otherwise we loose uGNI support on the trunk. > > Because we ARE NOT going to remove the mailbox size optimization. > > > > -Nathan > > > > On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: > >> WHAT: Should we make the job size (i.e., initial number of procs) > >> available in OPAL? > >> > >> WHY: At least 2 BTLs are using this info (*more below) > >> > >> WHERE: usnic and ugni > >> > >> TIMEOUT: there's already been some inflammatory emails about this; > >> let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 > >> > >> MORE DETAIL: > >> > >> This is an open question. We *have* the information at the time that the > >> BTLs are initialized: do we allow that information to go down to OPAL? > >> > >> Ralph added this info down in OPAL in r32355, but George reverted it in > >> r32361. > >> > >> Points for: YES, WE SHOULD > >> +++ 2 BTLs were using it (usinc, ugni) Other RTE job-related info are > >> +++ already in OPAL (num local ranks, local rank) > >> > >> Points for: NO, WE SHOULD NOT > >> --- What exactly is this number (e.g., num currently-connected procs?), > >> and when is it updated? > >> --- We need to precisely delineate what belongs in OPAL vs. above-OPAL > >> > >> FWIW: here's how ompi_process_info.num_procs was used before the BTL move > >> down to OPAL: > >> > >> - usnic: for a minor latency optimization / sizing of a shared receive > >> buffer queue length, and for the initial size of a peer lookup hash > >> - ugni: to determine the size of the per-peer buffers used for > >> send/recv communication > >> > >> -- > >> Jeff Squyres > >> jsquy...@cisco.com > >> For corporate legal information go to: > >> http://www.cisco.com/web/about/doing_business/legal/cri/ > >> > >> ___ > >> devel mailing list > >> de...@open-mpi.org > >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > >> Link to this post: > >> http://www.open-mpi.org/community/lists/devel/2014/07/15373.php > > ___ > > devel mailing list > > de...@open-mpi.org > > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > > Link to this post: > > http://www.open-mpi.org/community/lists/devel/2014/07/15395.php > > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15396.php pgpvqqekN2qM3.pgp Description: PGP signature
Re: [OMPI devel] RFC: job size info in OPAL
I'd like to suggest an alternative solution. A BTL can exploit whatever data it wants, but should first test if the data is available. If the data is *required*, then the BTL gracefully disqualifies itself. If the data is *desirable* for optimization, then the BTL writer (if they choose) can include an alternate path that doesn't do the optimization if the data isn't available. This would seem to meet everyone's needs, yes? On Jul 31, 2014, at 8:09 AM, George Bosilca <bosi...@icl.utk.edu> wrote: > What is your definition of “global job size”? > > George. > > On Jul 31, 2014, at 11:06 , Pritchard Jr., Howard <howa...@lanl.gov> wrote: > >> Hi Folks, >> >> I think given the way we want to use the btl's in lower levels like opal, >> it is pretty disgusting for a btl to need to figure out on its own something >> like a "global job size". That's not its business. Can't we add some >> attributes >> to the component's initialization method that provides hints for how to >> allocate resources it needs to provide its functionality? >> >> I'll see if there's something clever that can be done in ugni for now. >> I can always add in a hack to probe the apps placement info file and >> scale the smsg blocks by number of nodes rather than number of ranks. >> >> Howard >> >> >> -Original Message- >> From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan Hjelm >> Sent: Thursday, July 31, 2014 8:58 AM >> To: Open MPI Developers >> Subject: Re: [OMPI devel] RFC: job size info in OPAL >> >> >> +2^1000 >> >> This information is absolutely necessary at this point. If someone has a >> better solution they can provide it as an alternative RFC. Until then this >> is how it should be done... Otherwise we loose uGNI support on the trunk. >> Because we ARE NOT going to remove the mailbox size optimization. >> >> -Nathan >> >> On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: >>> WHAT: Should we make the job size (i.e., initial number of procs) available >>> in OPAL? >>> >>> WHY: At least 2 BTLs are using this info (*more below) >>> >>> WHERE: usnic and ugni >>> >>> TIMEOUT: there's already been some inflammatory emails about this; >>> let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 >>> >>> MORE DETAIL: >>> >>> This is an open question. We *have* the information at the time that the >>> BTLs are initialized: do we allow that information to go down to OPAL? >>> >>> Ralph added this info down in OPAL in r32355, but George reverted it in >>> r32361. >>> >>> Points for: YES, WE SHOULD >>> +++ 2 BTLs were using it (usinc, ugni) Other RTE job-related info are >>> +++ already in OPAL (num local ranks, local rank) >>> >>> Points for: NO, WE SHOULD NOT >>> --- What exactly is this number (e.g., num currently-connected procs?), and >>> when is it updated? >>> --- We need to precisely delineate what belongs in OPAL vs. above-OPAL >>> >>> FWIW: here's how ompi_process_info.num_procs was used before the BTL move >>> down to OPAL: >>> >>> - usnic: for a minor latency optimization / sizing of a shared receive >>> buffer queue length, and for the initial size of a peer lookup hash >>> - ugni: to determine the size of the per-peer buffers used for >>> send/recv communication >>> >>> -- >>> Jeff Squyres >>> jsquy...@cisco.com >>> For corporate legal information go to: >>> http://www.cisco.com/web/about/doing_business/legal/cri/ >>> >>> ___ >>> devel mailing list >>> de...@open-mpi.org >>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >>> Link to this post: >>> http://www.open-mpi.org/community/lists/devel/2014/07/15373.php >> ___ >> devel mailing list >> de...@open-mpi.org >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >> Link to this post: >> http://www.open-mpi.org/community/lists/devel/2014/07/15395.php > > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15396.php
Re: [OMPI devel] RFC: job size info in OPAL
What is your definition of “global job size”? George. On Jul 31, 2014, at 11:06 , Pritchard Jr., Howard <howa...@lanl.gov> wrote: > Hi Folks, > > I think given the way we want to use the btl's in lower levels like opal, > it is pretty disgusting for a btl to need to figure out on its own something > like a "global job size". That's not its business. Can't we add some > attributes > to the component's initialization method that provides hints for how to > allocate resources it needs to provide its functionality? > > I'll see if there's something clever that can be done in ugni for now. > I can always add in a hack to probe the apps placement info file and > scale the smsg blocks by number of nodes rather than number of ranks. > > Howard > > > -Original Message- > From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan Hjelm > Sent: Thursday, July 31, 2014 8:58 AM > To: Open MPI Developers > Subject: Re: [OMPI devel] RFC: job size info in OPAL > > > +2^1000 > > This information is absolutely necessary at this point. If someone has a > better solution they can provide it as an alternative RFC. Until then this is > how it should be done... Otherwise we loose uGNI support on the trunk. > Because we ARE NOT going to remove the mailbox size optimization. > > -Nathan > > On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: >> WHAT: Should we make the job size (i.e., initial number of procs) available >> in OPAL? >> >> WHY: At least 2 BTLs are using this info (*more below) >> >> WHERE: usnic and ugni >> >> TIMEOUT: there's already been some inflammatory emails about this; >> let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 >> >> MORE DETAIL: >> >> This is an open question. We *have* the information at the time that the >> BTLs are initialized: do we allow that information to go down to OPAL? >> >> Ralph added this info down in OPAL in r32355, but George reverted it in >> r32361. >> >> Points for: YES, WE SHOULD >> +++ 2 BTLs were using it (usinc, ugni) Other RTE job-related info are >> +++ already in OPAL (num local ranks, local rank) >> >> Points for: NO, WE SHOULD NOT >> --- What exactly is this number (e.g., num currently-connected procs?), and >> when is it updated? >> --- We need to precisely delineate what belongs in OPAL vs. above-OPAL >> >> FWIW: here's how ompi_process_info.num_procs was used before the BTL move >> down to OPAL: >> >> - usnic: for a minor latency optimization / sizing of a shared receive >> buffer queue length, and for the initial size of a peer lookup hash >> - ugni: to determine the size of the per-peer buffers used for >> send/recv communication >> >> -- >> Jeff Squyres >> jsquy...@cisco.com >> For corporate legal information go to: >> http://www.cisco.com/web/about/doing_business/legal/cri/ >> >> ___ >> devel mailing list >> de...@open-mpi.org >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >> Link to this post: >> http://www.open-mpi.org/community/lists/devel/2014/07/15373.php > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15395.php
Re: [OMPI devel] RFC: job size info in OPAL
Hi Folks, I think given the way we want to use the btl's in lower levels like opal, it is pretty disgusting for a btl to need to figure out on its own something like a "global job size". That's not its business. Can't we add some attributes to the component's initialization method that provides hints for how to allocate resources it needs to provide its functionality? I'll see if there's something clever that can be done in ugni for now. I can always add in a hack to probe the apps placement info file and scale the smsg blocks by number of nodes rather than number of ranks. Howard -Original Message- From: devel [mailto:devel-boun...@open-mpi.org] On Behalf Of Nathan Hjelm Sent: Thursday, July 31, 2014 8:58 AM To: Open MPI Developers Subject: Re: [OMPI devel] RFC: job size info in OPAL +2^1000 This information is absolutely necessary at this point. If someone has a better solution they can provide it as an alternative RFC. Until then this is how it should be done... Otherwise we loose uGNI support on the trunk. Because we ARE NOT going to remove the mailbox size optimization. -Nathan On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: > WHAT: Should we make the job size (i.e., initial number of procs) available > in OPAL? > > WHY: At least 2 BTLs are using this info (*more below) > > WHERE: usnic and ugni > > TIMEOUT: there's already been some inflammatory emails about this; > let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 > > MORE DETAIL: > > This is an open question. We *have* the information at the time that the > BTLs are initialized: do we allow that information to go down to OPAL? > > Ralph added this info down in OPAL in r32355, but George reverted it in > r32361. > > Points for: YES, WE SHOULD > +++ 2 BTLs were using it (usinc, ugni) Other RTE job-related info are > +++ already in OPAL (num local ranks, local rank) > > Points for: NO, WE SHOULD NOT > --- What exactly is this number (e.g., num currently-connected procs?), and > when is it updated? > --- We need to precisely delineate what belongs in OPAL vs. above-OPAL > > FWIW: here's how ompi_process_info.num_procs was used before the BTL move > down to OPAL: > > - usnic: for a minor latency optimization / sizing of a shared receive > buffer queue length, and for the initial size of a peer lookup hash > - ugni: to determine the size of the per-peer buffers used for > send/recv communication > > -- > Jeff Squyres > jsquy...@cisco.com > For corporate legal information go to: > http://www.cisco.com/web/about/doing_business/legal/cri/ > > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15373.php
Re: [OMPI devel] RFC: job size info in OPAL
+2^1000 This information is absolutely necessary at this point. If someone has a better solution they can provide it as an alternative RFC. Until then this is how it should be done... Otherwise we loose uGNI support on the trunk. Because we ARE NOT going to remove the mailbox size optimization. -Nathan On Wed, Jul 30, 2014 at 10:00:18PM +, Jeff Squyres (jsquyres) wrote: > WHAT: Should we make the job size (i.e., initial number of procs) available > in OPAL? > > WHY: At least 2 BTLs are using this info (*more below) > > WHERE: usnic and ugni > > TIMEOUT: there's already been some inflammatory emails about this; let's > discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 > > MORE DETAIL: > > This is an open question. We *have* the information at the time that the > BTLs are initialized: do we allow that information to go down to OPAL? > > Ralph added this info down in OPAL in r32355, but George reverted it in > r32361. > > Points for: YES, WE SHOULD > +++ 2 BTLs were using it (usinc, ugni) > +++ Other RTE job-related info are already in OPAL (num local ranks, local > rank) > > Points for: NO, WE SHOULD NOT > --- What exactly is this number (e.g., num currently-connected procs?), and > when is it updated? > --- We need to precisely delineate what belongs in OPAL vs. above-OPAL > > FWIW: here's how ompi_process_info.num_procs was used before the BTL move > down to OPAL: > > - usnic: for a minor latency optimization / sizing of a shared receive buffer > queue length, and for the initial size of a peer lookup hash > - ugni: to determine the size of the per-peer buffers used for send/recv > communication > > -- > Jeff Squyres > jsquy...@cisco.com > For corporate legal information go to: > http://www.cisco.com/web/about/doing_business/legal/cri/ > > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15373.php pgpRELGUpwpHm.pgp Description: PGP signature
Re: [OMPI devel] RFC: job size info in OPAL
On Jul 30, 2014, at 5:49 PM, George Bosilcawrote: > > On Jul 30, 2014, at 20:37 , Ralph Castain wrote: > >> >> On Jul 30, 2014, at 5:25 PM, George Bosilca wrote: >> >>> >>> On Jul 30, 2014, at 18:00 , Jeff Squyres (jsquyres) >>> wrote: >>> WHAT: Should we make the job size (i.e., initial number of procs) available in OPAL? WHY: At least 2 BTLs are using this info (*more below) WHERE: usnic and ugni TIMEOUT: there's already been some inflammatory emails about this; let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 MORE DETAIL: This is an open question. We *have* the information at the time that the BTLs are initialized: do we allow that information to go down to OPAL? Ralph added this info down in OPAL in r32355, but George reverted it in r32361. Points for: YES, WE SHOULD +++ 2 BTLs were using it (usinc, ugni) +++ Other RTE job-related info are already in OPAL (num local ranks, local rank) Points for: NO, WE SHOULD NOT --- What exactly is this number (e.g., num currently-connected procs?), and when is it updated? --- We need to precisely delineate what belongs in OPAL vs. above-OPAL >>> --- Using this information to configure the communication environment >>> limits the scope of communication substrate to a static application (in >>> number of participants). Under this assumption, one can simply wait until >>> the first add_proc to compute the number of processes, solution as >>> “correct” as the current one. >> >> Not necessarily - it depends on how it is used, and how it is communicated. >> Some of us have explored other options for using this that aren’t static, >> but where the info is of use. > > This is a little bit too much hand waving to be constructive. Some other > folks in the field have developed many communications libraries, and none of > them needed a random number of potential processes to initialize themselves > correctly. That's fine - everyone innovates and does something new. I'm not about to divulge proprietary, competitive info to you in advance just to justify our needs. I'll only note that notification of change isn't the sole jurisdiction of the FT group, and some of us have other uses for it. > >>> The other “global” information that were made available in OPAL >>> (num_local_peers and my_local_rank) are only used by local BTL (SM, SMCUDA >>> and VADER). Moreover, my_local_rank is only used to decide who initialize >>> the backend file, thing that can easily be done using an atomic operation. >>> The number of local processes is used to prevent SM from activating itself >>> if we don’t have at least 2 processes per node. So, their usage is >>> minimally invasive, and can eventually be phased out with a little effort. >> >> FWIW: the new PMI abstraction is in OPAL because it is RTE-agnostic. So all >> the info being discussed will actually be captured originally in the OPAL >> layer, and stored in the OPAL dstore framework. In the current code, the >> RTE grabs the data and exposes it to the OMPI layer, which then pushes it >> back down to the OPAL proc.h struct. >> >> since anyone can freely query the info from opal/pmix or >> opal/dstore, it is really irrelevant in some ways. The info is there, in the >> OPAL layer, prior to BTL's being initialized. If you don't want it in a >> global storage, people can just get it from the appropriate OPAL API. >> >> So what are we actually debating here? Global storage vs API call? > > Our goals in this project are clearly orthogonal. I put a lot of effort into > this move because I need to use the BTLs without PMI, without RTE. And you are certainly free to do so. Nobody is putting a gun to your head and demanding that your BTLs use it > In fact the question boils down to: Do you want to be able to use the BTL to > bootstrap the RTE or not? If yes, then the number of processes is out of the > picture, either as an API or as a global storage. Yes, I do - and no, it isn't a black/white question. I can use the BTLs to bootstrap just fine, even when someone uses that info for an initial optimization. I can always notify them later when things change, and they can make adjustments if necessary. Again, nobody is forcing you to use any of the data in the opal dstore. It is just there if someone *wants* to use it. I fail to understand why you want to tell everyone else what they can do in their BTL. If you don't like how they wrote it, you are always free to write your own version of it. Nobody will stop you. So what is the issue here? > > George. > > >> >>> >>> George. >>> >>> FWIW: here's how ompi_process_info.num_procs was used before the BTL move down to OPAL: - usnic: for a minor latency optimization / sizing
Re: [OMPI devel] RFC: job size info in OPAL
On Jul 30, 2014, at 5:25 PM, George Bosilcawrote: > > On Jul 30, 2014, at 18:00 , Jeff Squyres (jsquyres) > wrote: > >> WHAT: Should we make the job size (i.e., initial number of procs) available >> in OPAL? >> >> WHY: At least 2 BTLs are using this info (*more below) >> >> WHERE: usnic and ugni >> >> TIMEOUT: there's already been some inflammatory emails about this; let's >> discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 >> >> MORE DETAIL: >> >> This is an open question. We *have* the information at the time that the >> BTLs are initialized: do we allow that information to go down to OPAL? >> >> Ralph added this info down in OPAL in r32355, but George reverted it in >> r32361. >> >> Points for: YES, WE SHOULD >> +++ 2 BTLs were using it (usinc, ugni) >> +++ Other RTE job-related info are already in OPAL (num local ranks, local >> rank) >> >> Points for: NO, WE SHOULD NOT >> --- What exactly is this number (e.g., num currently-connected procs?), and >> when is it updated? >> --- We need to precisely delineate what belongs in OPAL vs. above-OPAL > --- Using this information to configure the communication environment limits > the scope of communication substrate to a static application (in number of > participants). Under this assumption, one can simply wait until the first > add_proc to compute the number of processes, solution as “correct” as the > current one. Not necessarily - it depends on how it is used, and how it is communicated. Some of us have explored other options for using this that aren't static, but where the info is of use. > >> > The other “global” information that were made available in OPAL > (num_local_peers and my_local_rank) are only used by local BTL (SM, SMCUDA > and VADER). Moreover, my_local_rank is only used to decide who initialize the > backend file, thing that can easily be done using an atomic operation. The > number of local processes is used to prevent SM from activating itself if we > don’t have at least 2 processes per node. So, their usage is minimally > invasive, and can eventually be phased out with a little effort. FWIW: the new PMI abstraction is in OPAL because it is RTE-agnostic. So all the info being discussed will actually be captured originally in the OPAL layer, and stored in the OPAL dstore framework. In the current code, the RTE grabs the data and exposes it to the OMPI layer, which then pushes it back down to the OPAL proc.h struct. since anyone can freely query the info from opal/pmix or opal/dstore, it is really irrelevant in some ways. The info is there, in the OPAL layer, prior to BTL's being initialized. If you don't want it in a global storage, people can just get it from the appropriate OPAL API. So what are we actually debating here? Global storage vs API call? > > George. > > >> FWIW: here's how ompi_process_info.num_procs was used before the BTL move >> down to OPAL: >> >> - usnic: for a minor latency optimization / sizing of a shared receive >> buffer queue length, and for the initial size of a peer lookup hash >> - ugni: to determine the size of the per-peer buffers used for send/recv >> communication >> >> -- >> Jeff Squyres >> jsquy...@cisco.com >> For corporate legal information go to: >> http://www.cisco.com/web/about/doing_business/legal/cri/ >> >> ___ >> devel mailing list >> de...@open-mpi.org >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel >> Link to this post: >> http://www.open-mpi.org/community/lists/devel/2014/07/15373.php > > ___ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/07/15378.php
[OMPI devel] RFC: job size info in OPAL
WHAT: Should we make the job size (i.e., initial number of procs) available in OPAL? WHY: At least 2 BTLs are using this info (*more below) WHERE: usnic and ugni TIMEOUT: there's already been some inflammatory emails about this; let's discuss next Tuesday on the teleconf: Tue, 5 Aug 2014 MORE DETAIL: This is an open question. We *have* the information at the time that the BTLs are initialized: do we allow that information to go down to OPAL? Ralph added this info down in OPAL in r32355, but George reverted it in r32361. Points for: YES, WE SHOULD +++ 2 BTLs were using it (usinc, ugni) +++ Other RTE job-related info are already in OPAL (num local ranks, local rank) Points for: NO, WE SHOULD NOT --- What exactly is this number (e.g., num currently-connected procs?), and when is it updated? --- We need to precisely delineate what belongs in OPAL vs. above-OPAL FWIW: here's how ompi_process_info.num_procs was used before the BTL move down to OPAL: - usnic: for a minor latency optimization / sizing of a shared receive buffer queue length, and for the initial size of a peer lookup hash - ugni: to determine the size of the per-peer buffers used for send/recv communication -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/