-
Notifications
You must be signed in to change notification settings - Fork 861
WeeklyTelcon_20191112
- Dialup Info: (Do not post to public mailing list or public wiki)
- Geoffrey Paulsen (IBM)
- Thomas Naughton (ORNL)
- Todd Kordenbrock (Sandia)
- Noah Evans (Sandia)
- Artem Polyakov (Mellanox)
- Austen Lauria (IBM)
- Howard Pritchard (LANL)
- Edgar Gabriel (UH)
- Josh Hursey (IBM)
- Matthew Dosanjh (Sandia)
- William Zhang (AWS)
- Jeff Squyres (Cisco)
- Brendan Cunningham (Intel)
- Brian Barrett (AWS)
- Akshay Venkatesh (NVIDIA)
- George Bosilca (UTK)
- Michael Heinz (Intel)
- David Bernhold (ORNL)
- Harumi Kuno (HPE)
- Brandon Yates (Intel)
- Charles Shereda (LLNL)
- Erik Zeiske
- Joshua Ladd (Mellanox)
- Mark Allen (IBM)
- Matias Cabral (Intel)
- Nathan Hjelm (Google)
- Ralph Castain (Intel)
- Tom Naughton
- Xin Zhao (Mellanox)
- mohan (AWS)
- Open Call - if you want 1 or 2 slides in BOF for your organization, need them by Thursday evening.
- No CALL next week (11/19/2019)
- A comment about stability / testing.
- There are not explicit testing for ORTE, but it gets tested in Open MPI CI / MTT.
- PRRTE has less testing, because it's not directly testing.
- PRRTE will be needed for PMIX community.
- Probably same community, and a shame to double the effort.
- Binding options are the same and are there in PRRTE.
- Singleton and ??? frameworks not in PRRTE, because it's not needed in PMIX
- A ticket open on singletons / PMIx_Spawn()
- Georges code for reliable connections will get pushed upstream soon when it's ready.
- DVM has switched over to PRRTE.
- Concerns about supporting another project.
- It adds another level of overhead coordinating / synchronizing with PRRTE/PMIX community.
- It is valuable to have a runtime system that's divorced from MPI.
- Don't know how to balance these two, because they're both not wrong.
- We need a launcher, but don't want to support more than one.
- But we support many launchers like slurm, lsf, flux, etc.
- Those other launchers have companies and organizations behind them, and they support them through Open MPI.
- A compromise between the two, would be to create an ORTE with all MPI removed.
- easy to make a dist tarball
- would skirt many political issues
- Shouldn't support both ORTE and PRRTE.
- Need to use the high level interfaces for PMIx, so we can move from version to version.
- So this gets back to making PMIx the first class citizen.
- Has to happen if we are working away from ORTE.
-
All of this in context in v5.0
-
Intel is no longer driving PRRTE work, and Ralph won't be available for PRRTE much either.
-
PRRTE will be a good PMIX developement environment, but no longer a focus to be a scale and robust launcher.
-
OMPI community could come into PRRTE, and put in production / scalability testing, features, etc.
-
Given that we have not been good at contributing to PRRTE (other than Ralph), there's another proposal
- There's been a drift from ORTE / PRRTE, so transitioning is risky.
-
Step 1. Make PMIX a first class citizen
- Still good to keep PMIX as a static framework (no more glue, but still under
orte/mca/pmix, but basicly just passes through, and callPMIX_calls directly. - Allows us to still have internal backup PMIx if no external PMIX is found.
- Still good to keep PMIX as a static framework (no more glue, but still under
-
Step 2. We can whittle down orte, since PMIX does much of this.
-
Two things PRRTE won't care about, is scale and all binding patterns.
-
Only recent versions of SLURM have PMIx
-
Need to continue to support ssh.
- Not just core PMIx, still need daemons for SSH to work, but they're not part of PMIx.
- Part of ORTE that we wouldn't be deleting.
-
What does Altair PbsPro and open source PbsPro do?
- Torque is different than PbsPro
-
Are there OLD systems that we currently support that we still don't care, and could discontinue support in v5.x
- Who supports PMIx, and who doesn't
-
If PMIx becomes a first class citizen and rest of code base just makes PMIx calls, how do we support these things?
- mpirun would still have to launch orteds via plm.
- srun wouldn't need
- But this is how it works today. Torque doesn't support PMIx at all, but TM just launches ORTEDs
- ALPS - aprun ./a.out - requires a.out to connect up to ALPS daemons.
- Cray still supports PMI - someone would need to write a PMI -> PMIX adapter.
- ORTE does not have the concept of persistant daemons
-
Is there a situation where we might have a launcher launching ortes and we'd need to relay pmix calls to the correct pmix server layer?
- Generally we won't have that situation, since the launcher won't launch ORTEds.
-
George's work currently depends on PRRTE
- If ORTEDs provides PMIx_Events, would that be enough?
- No George needs PRRTE's fault-tollerant overlay network.
- George will scope the effort to port that feature from PRRTE to ORTE.
- If ORTEDs provides PMIx_Events, would that be enough?
-
ACTION - Please gather list of resource managers, and Tools that we care about supporting in Open-MPI v5.0.x
-
Today - Howard
- Summary - make PMIx a first class citizen.
- Then whittle away ORTE as much as possible.
- We think the only one who uses PMI1, and PMI2 might be cray.
- Howard doesn't think Cray's even going to go that direction, might be adopting pmix for future direciton. Good super computing question.
- Most places will be whatever SLURM does.
- What will MPICH do? suspect PMIx
- Howard thinks that by the time Open-MPI v5 gets out
- Is SLURM + PMIx dead? No, it's supported, just not all of the
-
George looked into scoping the amount of work to bring reliable overlay network from
- PRRTE frameworks not in
-
Howard also brought up that Sessions only works with PRRTE right now, so would need to backport this as well.
-
Only thing that depends on PRRTE is Sessions, Reliable connections, and Resource allocation support. Thing Geoffry Valle was working on before. Howard will investigate.
- William Zhang has not yet committed some graph code for reachability similar to usnic.
- Brian/William will get with Josh Hursey to potentially test some more.
- Not sure what wanted the behavior of the netlinks reachability component.
- Wasn't detecting mark's
- Linux is always going to give you local route before localhost. This is one place where using reachability framework changes behavior.
- Options, where do we want to fix this?
- Truthfully (even if you use 192.168...., linux will route over localhost device).
- can say if these two addresses are the same it's always "reachable".
- could put this down in the frameworks itself.
- This is what netmasks does today, they're the same, so have high reachability.
- Can handle it in the higher layers
- Users could always specify localhost, or we could specify it for them.
- You don't use a device, you give the OS a hint of what device it should use.
- What does usnic do?
- usnic ONLY uses reachability for remote hosts.
- We encourage customers to say ifinclude localhost.
- This path has worked for years, and should probably keep it working.
- What happens if I have 3 devices, loopback, and 2 eth not wired together
- What if I say my source is 192, but dest is the 10. path.
- regardless of what reachability tells us, will this actually work?
- netlink will say it WONT work, but the OS will just make it work by routing over loopback device.
- Probably right answer is to special case in the netlink module, to return a
- It's official! Portland Oregon, Feb 17, 2020.
- Safe to begin booking travel now.
- Please register on Wiki page, since Jeff has to register you.
- Date looks good. Feb 17th right before MPI Forum
- 2pm monday, and maybe most of Tuesday
- Cisco has a portland facility and is happy to host.
- But willing to step asside if others want to host.
- about 20-30 min drive from MPI Forum, will probably need a car.
-
No update 11/12
-
Can we just turn on locbot / probot until we can get AWS bot online? *
-
OMPI has been waiting for some git submodule work in Jenkins on AWS.
- Need someone to have someone to figure out why Jenkins doesn't like Jeff's PR.
- Anyone with github account for ompi team should have access.
- PR 6821
- Apparently Jenkin's isn't behaving as it should.
- Three pieces: Jenkins, CI, bot.
- AWS has a libfabirc setup like this for testing.
- Issue is that they're reworking the design, and will rollout for both libfabric and open-mpi.
- William Zhang talked to Brian
- Not something AWS team will work on, but Brian will work on it.
- Jeff will talk to Brian as well.
- Need someone to have someone to figure out why Jenkins doesn't like Jeff's PR.
-
Howard and Jeff have access to Jenkins on AWS. Part of the problem is that we don't have much expertise on Jenkins/AWS.
- William will probably be admining the Jenkins/AWS or communicating with those who will.
-
Merged
--recurse-submodulesupdate intoompi-scriptsJenkins script as first step. Let's see if that works. -
Modular thread re-write (noah)
- UGNI and Vader BTLs were getting better performance, not sure why.
- For modular threading library, might be interesting to decide at compile time or runtime.
- Previously similar things seemed to be related to ICACHE.
- Howard will lok at.
Blockers All Open Blockers
Review v3.0.x Milestones v3.0.4
Review v3.1.x Milestones v3.1.4
- release v3.0.5 and v3.1.5 tomorrow.
Review v4.0.x Milestones v4.0.2
- v4.0.3 in the works.
- Schedule: Originally end of january.
- PR 1752 may drive an earlier release in case if UCX will be released sooner.
- Schedule: Originally end of january.
- PR 7151 - enhancement -
- UCX 1.7 release schedule - was an RC 1.
- Artem can check.
- There's a problem in Open MPI v4.0.2, that packagers will hit in UCX 1.7
- PR 1752 may drive an earlier release in case if UCX will be released sooner.
- Schedule: April 2020?
- Wiki - go look at items, and we should discuss a bit in weekly calls.
- Some items:
- MPI1 removed stuff.
Review Master Master Pull Requests
- IBM's PGI test has NEVER worked. Is it a real issue or local to IBM.
- Austen is looking into
- Absoft 32bit fortran failures.
- No discussion this week.
- See older weekday notes for prior items.
- No discussion this week.