PathForward Project Teleconference 5/4/05

Attendees:
Matt, Roland, Sean, Bill, Woody, Hal


Agenda Bashing

Administrivia
	OLS paper almost complete
	Next meeting 5/11

AR Review

Kernel update
	nothing upstream yet
	2.6.12 taking time to stabilize
	memory pinning still under discussion
		converged (except for Caitlin)
		implementing to that

SuperComputing 05
	2 activities proposed for November
		infrastructure for SCInet (IB interconnect)
			customers rather than vendor managed
			optical technology issue (4x to 4 parallel optic dongle)
				switches provide power
					(unused ground now need to be driven by switches)
					Some Topspin switches and Mellanox HCAs support this
						(software transparent)
		Tutorial
			for customers
			virtual networks
				Xen
					currently not supported
					not started on
					started investigating architecture
					substantial risk
					target end of next week strawman to openib and xen lists
					November tight
			plan to do without virtualization

2.6.9 patch
	svn 2245 building
	user mode verbs works, IPoIB
	one SDP issue
	still IA64 issue (firmware ?)
	EM64T works
	2.6.9 backports whenever OpenIB goes upstream
	missing some get_sb_psuedo so need custom kernel
		IPoIB needs this
		that could be pulled so user veerbs as binary could be made to work
		but not IPoIB
	add patches to source and build new kernel RPM ?

	patch based on 2.6.12-mm3
	split into kernel, base drivers, fixup
	need to be put into subversion somewhere
		waiting for 2.6.12 to be released
	latest code to 2.6.9
		user mode support
		IA64 issue ?
		SDP also
	RedFlag request
		RedHat also 2.6.9 based (RHEL4)
			would they pick this up ?
			TriLabs will be using RHEL4
		latest non SLES 2.6.11
		RedHat Fedora
	support issue for this ?
		leave to distro

MPI
	nothing new to report here
	No news from Tim
	DK released MVAPICH 0.9.5 lastweek
		starting to look at gen2
	Tim (LANL) 1-2 months OpenMPI on OpenIB
		RC with Steve Poole
	Matt started on LA MPI (UD)
	Intel MPI over uDAPL is running

Status of Current Tasks
	Sean: RMPP
		responding to emails and retesting
		already changed API to remove userspace parameter
		working with InfiniCon on CM (Bill Jordan)
			REJ on bad request
	Libor: user CM
		Arlin not yet started on this

		posted last week with simple example
		working on more sophisticated example
			uses IB verbs
		need to get path record		
	Roland: user verbs
		changes to memory pinning being implemented
		kernel patch for DONT COPY flag on mprotect/mmap (fork support)
			reference counting of overlapping regions
		ultimate ALWAYS COPY for partial pages
		is this usable for 2.6.9 ?
			RedHat can't change kernel ABI
			does this do that ? change symbol versions
		IA 64 details
		general cleanup
		remaining: large number of minor pieces 
			attach/detach mc QPs, some query verbs (GID table)
			fair bit of loose ends still to do
			documentation
		2 more weeks for initial release ?
			tar balls and RPMs
	Shahar/Hal:
		Worked on user_mad.c, libibumad, and OpenSM to support RMPP
			on the send side
			Next week work on receive side
		bad port handling algorithm finally converged
			needs be implemented
		OpenSM initialization times
		RARP
		Solaris snoop real fix still pending
		cluster diag testing completed by Josh

Bugzilla (overhead)
	revisit later (~ 1 month)

Should Mellanox be part of this call ?
	at stage 2 when have a piece of PF
	OpenSM ? verification ? other ?
	
Other/Associated Project Issues
	Work split between Roland and Libor
	additional features (priority order)
		user space events (port state, LID/SMLID changes)
 		user space snoop MAD
	IPoIB partial connectivity
	osmtest to be supported ?
	Plugfest/IL list may be possible in April 05 as DTA is likely to be dropped as requirement
		vote this week at CIWG
	SM client reregister/IPoIB reregistration for multicast groups
	EUI-64 and IPoIB
	HCA capability mask