[ewg] RE: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09)
Sorry Jeff... I am playing middle man with another engineer here for this information... The query does not have to be per QP, but it does need to be per IB HCA port to IB HCA port communication path. For example, if a node has 64 CPUs, it could do the query once per each other node on behalf of the 64 processes. Its still an N^2 set of queries, but at least N can be reduced to be the number of end node IB ports as opposed to the number of processes. -Original Message- From: Jeff Squyres [mailto:jsquy...@cisco.com] Sent: Monday, January 05, 2009 12:07 PM To: John Russo Cc: Tziporet Koren; ewg@lists.openfabrics.org; gene...@lists.openfabrics.org Subject: Re: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09) Hmm. Perhaps I'm not grokking your answer -- did you answer my question? I'm indirectly asking about scalability of the SM to have hundreds/thousands of MPI processes simultaneously querying the SM. On Jan 5, 2009, at 11:47 AM, John Russo wrote: When using routing algorithms such as lash, the SL used per end to end connection will vary based on the route. lash uses multiple VLs to avoid credit loops. As such the SL reported will vary based on fabric topology and which pair of end nodes the path is being requested on behalf of. -Original Message- From: Jeff Squyres [mailto:jsquy...@cisco.com] Sent: Monday, January 05, 2009 11:39 AM To: John Russo Cc: Tziporet Koren; ewg@lists.openfabrics.org; gene...@lists.openfabrics.org Subject: Re: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09) Would all MPI processes need to query the SM for each path that they want to use in a QP? On Jan 5, 2009, at 11:31 AM, John Russo wrote: Another suggestion for 1.5 Implementation of SA queries for Path Records (using IBTA 1.2.1 ServiceId field) in all OFED ULPs, especially for MPI The IBTA standard defines that the proper way to establish a connection is to get a PathRecord from the SM/SA and use it to define all the attributes of the communication path. Ideally the IBTA CM should then be used to establish the connection and QPs as well. At present, openmpi, mvapich1 and mvapich2 do not use PathRecords, but instead hard code attributes like the PKey, SL, etc. In some cases these hardcoded values can be overridden by configurable values such as PKey and SL, but such values must be uniform across all connections and must be provided per job (which can be error prone/tedious). At present opensm supports PKeys and SLs, however MPI cannot easily use these features. Other features, such as lash routing, in opensm do not work properly with MPI because the SL must be uniform across all connections, but for lash it will vary per route. Additionally, applications which do not use PathRecords will have difficulties with advanced features like IB routing, partitioning, etc. All of which are available or being worked on in opensm. From: ewg-boun...@lists.openfabrics.org [mailto:ewg-boun...@lists.openfabrics.org ] On Behalf Of Tziporet Koren Sent: Monday, January 05, 2009 1:00 AM To: ewg@lists.openfabrics.org Cc: gene...@lists.openfabrics.org Subject: [ewg] Agenda for the OFED meeting today (Jan 5, 09) Hello all, I hope we all had nice holidays and vacations, and now it's the time to get back to business. Agenda for OFED meeting today: 1. Conclusions from OFED 1.4 release: Open discussion 2. Do we wish to have OFED 1.4.: Please send pros cons before the meeting 3. OFED 1.5: Schedule and features. This is what we presented in SC08 about 1.5: Preliminary Schedule: * Feature Freeze: 3/20/09 * Alpha Release: 3/20/09 * Beta Release: 4/20/09 * RC1: 5/5/09 * RC2-RCx: About every 2 weeks as needed * Release: June 2009 Features: * Kernel.org: 2.6.28 and 2.6.29 * Multiple Event Queues to support Multi-core CPUs * NFS/RDMA - GA * RDS support for iWARP * OpenMPI 1.3 * Add support/backports for RedHat EL 5.3 and EL 4.8, SLES 11 * Support for Mellanox vNIC (EoIB) and FCoIB with BridgeX device * more TBD... We also presented the OS matrix but I suggest we will close this in the next meeting. My proposal: * Have the release in July and not June - so we will have more time for development * Stick to one kernel version base and not change in the middle since we saw that changing the kernel base caused a delay. We need to decide in the meeting if it is 2.6.29 or we should wait for 2.6.30. * Add IB over Eth - this is similar to iWARP but more like IB (e.g. including UD), and can work over ConnectX. Please send your suggestions to the list before the meeting if possible Tziporet ___ general mailing list gene...@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please
[ewg] OFED teleconference
BEGIN:VCALENDAR METHOD:REQUEST PRODID:Microsoft CDO for Microsoft Exchange VERSION:2.0 BEGIN:VTIMEZONE TZID:GMT -0500 (Standard) / GMT -0400 (Daylight) BEGIN:STANDARD DTSTART:16010101T02 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;WKST=MO;INTERVAL=1;BYMONTH=11;BYDAY=1SU END:STANDARD BEGIN:DAYLIGHT DTSTART:16010101T02 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;WKST=MO;INTERVAL=1;BYMONTH=3;BYDAY=2SU END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20090105T112645Z DTSTART;TZID=GMT -0500 (Standard) / GMT -0400 (Daylight):20090105T12 SUMMARY:OFED teleconference UID:04008200E00074C5B7101A82E00820EF8587FE6EC901000 01A58D9270244CD4980609EC7FDFD1B5F ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN=e...@lists .openfabrics.org:MAILTO:ewg@lists.openfabrics.org ORGANIZER;CN=Jeff Squyres (jsquyres):MAILTO:jsquy...@cisco.com LOCATION:ID: 210020028 DTEND;TZID=GMT -0500 (Standard) / GMT -0400 (Daylight):20090105T13 RRULE:FREQ=WEEKLY;UNTIL=20090119T17Z;INTERVAL=2;BYDAY=MO;WKST=MO DESCRIPTION:\N_ _\NJeffrey Squyres has invited you to a Cisco Unified Meet ingPlace Conference\N\NDate/Time: JAN 5\, 2009 at 12:00PM Am erica/New_York\NLength: 60\NFrequency: 2\NM eeting ID: 210020028\NMeeting Password:\N\NGlobal Acc ess Numbers:\Nhttp://cisco.com/en/US/about/doing_business/conferencing/ind ex.html\N\NSan Jose\, CA: +1.408.525.6800 RTP: +1.919.39 2.3330 \NUS/Canada: +1.866.432.9903United Kingdom: +44.20. 8824.0117\NIndia: +91.80.4103.3979 Germany: +49.619.67 73.9002\NJapan: +81.3.5763.9394China:+86.10.8515. 5666\N\NTO ATTEND A WEB AND VOICE CONFERENCE:\N\NCISCO INTRANET ATTENDEES\ NJoin the Web Voice Conference*\N1. Go to http://meetingplaceinternal.ci sco.com/join.asp?210020028\N2. Enter your CEC User ID Password then clic k OK\N- Accept any security warnings you receive and wait for the Meeting Room to initialize\N3. Click on CONNECT from the Meeting Room to join the Voice Conference portion of the meeting\N\NEXTERNAL ATTENDEES - Outside th e Cisco Intranet**\NJoin the Web Voice Conference*\N1. Go to http://meet ingplace.cisco.com/join.asp?210020028\N2. Fill in the My Name is field the n click Attend Meeting\N- If you have a CEC User ID\, click on the Cisco i con\N- Accept any security warnings you receive and wait for the Meeting R oom to initialize\N3. Click on CONNECT from the Meeting Room to join the V oice Conference portion of the meeting\N- Note: Guest users will see a lin k to the Global Access Numbers.\N\N*If this is your first time attending a Web Conference\, disable any pop-up blockers and visit http://meetingplac e.cisco.com/mpweb/scripts/browsertestupper.asp to test your web browser fo r compatibility with the Web Conference.\N**Not all meetings are scheduled to allow external attendees into the Web Conference portion of the meetin g\, if the URL does not work\, please follow the Voice only Conference ins tructions below to attend.\N\NTO ATTEND A VOICE ONLY CONFERENCE\N1. Dial i nto Cisco Unified MeetingPlace (view the Access Numbers and link above)\N2 . Press 1 to attend the meeting\N3. Follow the prompts to enter the Meetin g ID 210020028 and join the meeting\N\NSUPPORT\NInformation about this Con ference: Contact Jeffrey Squyres\, 914085250971\NCisco IT Support Center: Attend the Voice Conference and then press #0 on your phone keypad\N\N __ _\N\N SEQUENCE:0 PRIORITY:5 CLASS: CREATED:20090105T112648Z LAST-MODIFIED:20090105T112649Z STATUS:CONFIRMED TRANSP:OPAQUE X-MICROSOFT-CDO-BUSYSTATUS:BUSY X-MICROSOFT-CDO-INSTTYPE:1 X-MICROSOFT-CDO-INTENDEDSTATUS:BUSY X-MICROSOFT-CDO-ALLDAYEVENT:FALSE X-MICROSOFT-CDO-IMPORTANCE:1 X-MICROSOFT-CDO-OWNERAPPTID:-62580775 X-MICROSOFT-CDO-APPT-SEQUENCE:0 X-MICROSOFT-CDO-ATTENDEE-CRITICAL-CHANGE:20090105T112645Z X-MICROSOFT-CDO-OWNER-CRITICAL-CHANGE:20090105T112645Z BEGIN:VALARM ACTION:DISPLAY DESCRIPTION:REMINDER TRIGGER;RELATED=START:-PT00H15M00S END:VALARM END:VEVENT END:VCALENDAR ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
[ewg] OFED teleconf outlook invites
An outlook invite is coming shortly for an OFED teleconference today and Jan 19th, both at the usual times. *** Please do not send me the replies (accept, tentative, decline). Thanks. -- Jeff Squyres Cisco Systems ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
[ewg] RE: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09)
When using routing algorithms such as lash, the SL used per end to end connection will vary based on the route. lash uses multiple VLs to avoid credit loops. As such the SL reported will vary based on fabric topology and which pair of end nodes the path is being requested on behalf of. -Original Message- From: Jeff Squyres [mailto:jsquy...@cisco.com] Sent: Monday, January 05, 2009 11:39 AM To: John Russo Cc: Tziporet Koren; ewg@lists.openfabrics.org; gene...@lists.openfabrics.org Subject: Re: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09) Would all MPI processes need to query the SM for each path that they want to use in a QP? On Jan 5, 2009, at 11:31 AM, John Russo wrote: Another suggestion for 1.5 Implementation of SA queries for Path Records (using IBTA 1.2.1 ServiceId field) in all OFED ULPs, especially for MPI The IBTA standard defines that the proper way to establish a connection is to get a PathRecord from the SM/SA and use it to define all the attributes of the communication path. Ideally the IBTA CM should then be used to establish the connection and QPs as well. At present, openmpi, mvapich1 and mvapich2 do not use PathRecords, but instead hard code attributes like the PKey, SL, etc. In some cases these hardcoded values can be overridden by configurable values such as PKey and SL, but such values must be uniform across all connections and must be provided per job (which can be error prone/tedious). At present opensm supports PKeys and SLs, however MPI cannot easily use these features. Other features, such as lash routing, in opensm do not work properly with MPI because the SL must be uniform across all connections, but for lash it will vary per route. Additionally, applications which do not use PathRecords will have difficulties with advanced features like IB routing, partitioning, etc. All of which are available or being worked on in opensm. From: ewg-boun...@lists.openfabrics.org [mailto:ewg-boun...@lists.openfabrics.org ] On Behalf Of Tziporet Koren Sent: Monday, January 05, 2009 1:00 AM To: ewg@lists.openfabrics.org Cc: gene...@lists.openfabrics.org Subject: [ewg] Agenda for the OFED meeting today (Jan 5, 09) Hello all, I hope we all had nice holidays and vacations, and now it's the time to get back to business. Agenda for OFED meeting today: 1. Conclusions from OFED 1.4 release: Open discussion 2. Do we wish to have OFED 1.4.: Please send pros cons before the meeting 3. OFED 1.5: Schedule and features. This is what we presented in SC08 about 1.5: Preliminary Schedule: * Feature Freeze: 3/20/09 * Alpha Release: 3/20/09 * Beta Release: 4/20/09 * RC1: 5/5/09 * RC2-RCx: About every 2 weeks as needed * Release: June 2009 Features: * Kernel.org: 2.6.28 and 2.6.29 * Multiple Event Queues to support Multi-core CPUs * NFS/RDMA - GA * RDS support for iWARP * OpenMPI 1.3 * Add support/backports for RedHat EL 5.3 and EL 4.8, SLES 11 * Support for Mellanox vNIC (EoIB) and FCoIB with BridgeX device * more TBD... We also presented the OS matrix but I suggest we will close this in the next meeting. My proposal: * Have the release in July and not June - so we will have more time for development * Stick to one kernel version base and not change in the middle since we saw that changing the kernel base caused a delay. We need to decide in the meeting if it is 2.6.29 or we should wait for 2.6.30. * Add IB over Eth - this is similar to iWARP but more like IB (e.g. including UD), and can work over ConnectX. Please send your suggestions to the list before the meeting if possible Tziporet ___ general mailing list gene...@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general -- Jeff Squyres Cisco Systems ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
[ewg] Re: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09)
Hmm. Perhaps I'm not grokking your answer -- did you answer my question? I'm indirectly asking about scalability of the SM to have hundreds/thousands of MPI processes simultaneously querying the SM. On Jan 5, 2009, at 11:47 AM, John Russo wrote: When using routing algorithms such as lash, the SL used per end to end connection will vary based on the route. lash uses multiple VLs to avoid credit loops. As such the SL reported will vary based on fabric topology and which pair of end nodes the path is being requested on behalf of. -Original Message- From: Jeff Squyres [mailto:jsquy...@cisco.com] Sent: Monday, January 05, 2009 11:39 AM To: John Russo Cc: Tziporet Koren; ewg@lists.openfabrics.org; gene...@lists.openfabrics.org Subject: Re: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09) Would all MPI processes need to query the SM for each path that they want to use in a QP? On Jan 5, 2009, at 11:31 AM, John Russo wrote: Another suggestion for 1.5 Implementation of SA queries for Path Records (using IBTA 1.2.1 ServiceId field) in all OFED ULPs, especially for MPI The IBTA standard defines that the proper way to establish a connection is to get a PathRecord from the SM/SA and use it to define all the attributes of the communication path. Ideally the IBTA CM should then be used to establish the connection and QPs as well. At present, openmpi, mvapich1 and mvapich2 do not use PathRecords, but instead hard code attributes like the PKey, SL, etc. In some cases these hardcoded values can be overridden by configurable values such as PKey and SL, but such values must be uniform across all connections and must be provided per job (which can be error prone/tedious). At present opensm supports PKeys and SLs, however MPI cannot easily use these features. Other features, such as lash routing, in opensm do not work properly with MPI because the SL must be uniform across all connections, but for lash it will vary per route. Additionally, applications which do not use PathRecords will have difficulties with advanced features like IB routing, partitioning, etc. All of which are available or being worked on in opensm. From: ewg-boun...@lists.openfabrics.org [mailto:ewg-boun...@lists.openfabrics.org ] On Behalf Of Tziporet Koren Sent: Monday, January 05, 2009 1:00 AM To: ewg@lists.openfabrics.org Cc: gene...@lists.openfabrics.org Subject: [ewg] Agenda for the OFED meeting today (Jan 5, 09) Hello all, I hope we all had nice holidays and vacations, and now it's the time to get back to business. Agenda for OFED meeting today: 1. Conclusions from OFED 1.4 release: Open discussion 2. Do we wish to have OFED 1.4.: Please send pros cons before the meeting 3. OFED 1.5: Schedule and features. This is what we presented in SC08 about 1.5: Preliminary Schedule: * Feature Freeze: 3/20/09 * Alpha Release: 3/20/09 * Beta Release: 4/20/09 * RC1: 5/5/09 * RC2-RCx: About every 2 weeks as needed * Release: June 2009 Features: * Kernel.org: 2.6.28 and 2.6.29 * Multiple Event Queues to support Multi-core CPUs * NFS/RDMA - GA * RDS support for iWARP * OpenMPI 1.3 * Add support/backports for RedHat EL 5.3 and EL 4.8, SLES 11 * Support for Mellanox vNIC (EoIB) and FCoIB with BridgeX device * more TBD... We also presented the OS matrix but I suggest we will close this in the next meeting. My proposal: * Have the release in July and not June - so we will have more time for development * Stick to one kernel version base and not change in the middle since we saw that changing the kernel base caused a delay. We need to decide in the meeting if it is 2.6.29 or we should wait for 2.6.30. * Add IB over Eth - this is similar to iWARP but more like IB (e.g. including UD), and can work over ConnectX. Please send your suggestions to the list before the meeting if possible Tziporet ___ general mailing list gene...@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general -- Jeff Squyres Cisco Systems -- Jeff Squyres Cisco Systems ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
[ewg] RE: Agenda for the OFED meeting today (Jan 5, 09)
There are scaling issues with SA path-record queries. We attempted to be good citizens with Intel MPI using the rdma_cm agent (via uDAPL) but was forced to build hard-coded RC QP support in OFED 1.4 (uDAPL scm) to avoid the many scaling and configuration problems that came with IPoIB requirements, ARP storms, rdma_cm timers, and SA path record query/caching. If someone wants to sign up to design and implement a scalable SA query caching agent we would be happy to look at path record queries again. -arlin Another suggestion for 1.5 Implementation of SA queries for Path Records (using IBTA 1.2.1 ServiceId field) in all OFED ULPs, especially for MPI The IBTA standard defines that the proper way to establish a connection is to get a PathRecord from the SM/SA and use it to define all the attributes of the communication path. Ideally the IBTA CM should then be used to establish the connection and QPs as well. At present, openmpi, mvapich1 and mvapich2 do not use PathRecords, but instead hard code attributes like the PKey, SL, etc. In some cases these hardcoded values can be overridden by configurable values such as PKey and SL, but such values must be uniform across all connections and must be provided per job (which can be error prone/tedious). At present opensm supports PKeys and SLs, however MPI cannot easily use these features. Other features, such as lash routing, in opensm do not work properly with MPI because the SL must be uniform across all connections, but for lash it will vary per route. Additionally, applications which do not use PathRecords will have difficulties with advanced features like IB routing, partitioning, etc. All of which are available or being worked on in opensm. ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
[ewg] Re: [ofa-general] RE: Agenda for the OFED meeting today (Jan 5, 09)
Would all MPI processes need to query the SM for each path that they want to use in a QP? On Jan 5, 2009, at 11:31 AM, John Russo wrote: Another suggestion for 1.5 Implementation of SA queries for Path Records (using IBTA 1.2.1 ServiceId field) in all OFED ULPs, especially for MPI The IBTA standard defines that the proper way to establish a connection is to get a PathRecord from the SM/SA and use it to define all the attributes of the communication path. Ideally the IBTA CM should then be used to establish the connection and QPs as well. At present, openmpi, mvapich1 and mvapich2 do not use PathRecords, but instead hard code attributes like the PKey, SL, etc. In some cases these hardcoded values can be overridden by configurable values such as PKey and SL, but such values must be uniform across all connections and must be provided per job (which can be error prone/tedious). At present opensm supports PKeys and SLs, however MPI cannot easily use these features. Other features, such as lash routing, in opensm do not work properly with MPI because the SL must be uniform across all connections, but for lash it will vary per route. Additionally, applications which do not use PathRecords will have difficulties with advanced features like IB routing, partitioning, etc. All of which are available or being worked on in opensm. From: ewg-boun...@lists.openfabrics.org [mailto:ewg-boun...@lists.openfabrics.org ] On Behalf Of Tziporet Koren Sent: Monday, January 05, 2009 1:00 AM To: ewg@lists.openfabrics.org Cc: gene...@lists.openfabrics.org Subject: [ewg] Agenda for the OFED meeting today (Jan 5, 09) Hello all, I hope we all had nice holidays and vacations, and now it’s the time to get back to business. Agenda for OFED meeting today: 1. Conclusions from OFED 1.4 release: Open discussion 2. Do we wish to have OFED 1.4.: Please send pros cons before the meeting 3. OFED 1.5: Schedule and features. This is what we presented in SC08 about 1.5: Preliminary Schedule: • Feature Freeze: 3/20/09 • Alpha Release: 3/20/09 • Beta Release: 4/20/09 • RC1: 5/5/09 • RC2-RCx: About every 2 weeks as needed • Release: June 2009 Features: • Kernel.org: 2.6.28 and 2.6.29 • Multiple Event Queues to support Multi-core CPUs • NFS/RDMA – GA • RDS support for iWARP • OpenMPI 1.3 • Add support/backports for RedHat EL 5.3 and EL 4.8, SLES 11 • Support for Mellanox vNIC (EoIB) and FCoIB with BridgeX device • more TBD… We also presented the OS matrix but I suggest we will close this in the next meeting. My proposal: • Have the release in July and not June - so we will have more time for development • Stick to one kernel version base and not change in the middle since we saw that changing the kernel base caused a delay. We need to decide in the meeting if it is 2.6.29 or we should wait for 2.6.30. • Add IB over Eth - this is similar to iWARP but more like IB (e.g. including UD), and can work over ConnectX. Please send your suggestions to the list before the meeting if possible Tziporet ___ general mailing list gene...@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general -- Jeff Squyres Cisco Systems ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
Re: [ewg] RE: Agenda for the OFED meeting today (Jan 5, 09)
I chatted with John and Todd from QL on the phone today -- we basically came to the same conclusion: - need to beef-up opensm to be able to scalably handle lots of incoming path record lookups - need to beef-up the CM clients on the host (maybe; this work might already be done?) - need to see the current status of the SA caching stuff / re-open that discussion to see if the work can be completed, etc. It might also be worthwhile to start a whole new discussion about making a better CM (at least from the ULP perspective). One that offers simple mechanisms for those who don't need/care about the details, but also offers complex/detailed mechanisms (perhaps remarkably like today's mechanisms). On Jan 5, 2009, at 2:01 PM, Davis, Arlin R wrote: There are scaling issues with SA path-record queries. We attempted to be good citizens with Intel MPI using the rdma_cm agent (via uDAPL) but was forced to build hard-coded RC QP support in OFED 1.4 (uDAPL scm) to avoid the many scaling and configuration problems that came with IPoIB requirements, ARP storms, rdma_cm timers, and SA path record query/caching. If someone wants to sign up to design and implement a scalable SA query caching agent we would be happy to look at path record queries again. -arlin Another suggestion for 1.5 Implementation of SA queries for Path Records (using IBTA 1.2.1 ServiceId field) in all OFED ULPs, especially for MPI The IBTA standard defines that the proper way to establish a connection is to get a PathRecord from the SM/SA and use it to define all the attributes of the communication path. Ideally the IBTA CM should then be used to establish the connection and QPs as well. At present, openmpi, mvapich1 and mvapich2 do not use PathRecords, but instead hard code attributes like the PKey, SL, etc. In some cases these hardcoded values can be overridden by configurable values such as PKey and SL, but such values must be uniform across all connections and must be provided per job (which can be error prone/tedious). At present opensm supports PKeys and SLs, however MPI cannot easily use these features. Other features, such as lash routing, in opensm do not work properly with MPI because the SL must be uniform across all connections, but for lash it will vary per route. Additionally, applications which do not use PathRecords will have difficulties with advanced features like IB routing, partitioning, etc. All of which are available or being worked on in opensm. ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg -- Jeff Squyres Cisco Systems ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
[ewg] RE: Agenda for the OFED meeting today (Jan 5, 09)
Another suggestion for 1.5 Implementation of SA queries for Path Records (using IBTA 1.2.1 ServiceId field) in all OFED ULPs, especially for MPI The IBTA standard defines that the proper way to establish a connection is to get a PathRecord from the SM/SA and use it to define all the attributes of the communication path. Ideally the IBTA CM should then be used to establish the connection and QPs as well. At present, openmpi, mvapich1 and mvapich2 do not use PathRecords, but instead hard code attributes like the PKey, SL, etc. In some cases these hardcoded values can be overridden by configurable values such as PKey and SL, but such values must be uniform across all connections and must be provided per job (which can be error prone/tedious). At present opensm supports PKeys and SLs, however MPI cannot easily use these features. Other features, such as lash routing, in opensm do not work properly with MPI because the SL must be uniform across all connections, but for lash it will vary per route. Additionally, applications which do not use PathRecords will have difficulties with advanced features like IB routing, partitioning, etc. All of which are available or being worked on in opensm. From: ewg-boun...@lists.openfabrics.org [mailto:ewg-boun...@lists.openfabrics.org] On Behalf Of Tziporet Koren Sent: Monday, January 05, 2009 1:00 AM To: ewg@lists.openfabrics.org Cc: gene...@lists.openfabrics.org Subject: [ewg] Agenda for the OFED meeting today (Jan 5, 09) Hello all, I hope we all had nice holidays and vacations, and now it's the time to get back to business. Agenda for OFED meeting today: 1. Conclusions from OFED 1.4 release: Open discussion 2. Do we wish to have OFED 1.4.: Please send pros cons before the meeting 3. OFED 1.5: Schedule and features. This is what we presented in SC08 about 1.5: Preliminary Schedule: * Feature Freeze: 3/20/09 * Alpha Release: 3/20/09 * Beta Release: 4/20/09 * RC1: 5/5/09 * RC2-RCx: About every 2 weeks as needed * Release: June 2009 Features: * Kernel.org: 2.6.28 and 2.6.29 * Multiple Event Queues to support Multi-core CPUs * NFS/RDMA - GA * RDS support for iWARP * OpenMPI 1.3 * Add support/backports for RedHat EL 5.3 and EL 4.8, SLES 11 * Support for Mellanox vNIC (EoIB) and FCoIB with BridgeX device * more TBD... We also presented the OS matrix but I suggest we will close this in the next meeting. My proposal: * Have the release in July and not June - so we will have more time for development * Stick to one kernel version base and not change in the middle since we saw that changing the kernel base caused a delay. We need to decide in the meeting if it is 2.6.29 or we should wait for 2.6.30. * Add IB over Eth - this is similar to iWARP but more like IB (e.g. including UD), and can work over ConnectX. Please send your suggestions to the list before the meeting if possible Tziporet ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
Re: [ewg] RE: Agenda for the OFED meeting today (Jan 5, 09)
Jeff, On Mon, Jan 5, 2009 at 3:26 PM, Jeff Squyres jsquy...@cisco.com wrote: I chatted with John and Todd from QL on the phone today -- we basically came to the same conclusion: - need to beef-up opensm to be able to scalably handle lots of incoming path record lookups This is the most obvious SA scalability issue but there are some others which may be important (related to SA caching rather than SA distribution as an approach). - need to beef-up the CM clients on the host (maybe; this work might already be done?) - need to see the current status of the SA caching stuff / re-open that discussion to see if the work can be completed, etc. IMO this will aggravate other SA scalability issues as well as there being other limitations with this approach. Don't get me wrong; I'm all for improving the SA scalability; there's no quick solution to this AFAIK. It would be interesting to see an apples to apples comparison of OpenSM and proprietary SMs in terms of running on the same hardware and the transaction rate for various things. I think this warrants an open discussion if people are serious about working on this issue. It might also be worthwhile to start a whole new discussion about making a better CM (at least from the ULP perspective). One that offers simple mechanisms for those who don't need/care about the details, but also offers complex/detailed mechanisms (perhaps remarkably like today's mechanisms). I've heard similar comments before but this too will take significant where-with-all IMO. -- Hal ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
Re: [ewg] RE: Agenda for the OFED meeting today (Jan 5, 09)
On Jan 5, 2009, at 5:16 PM, Hal Rosenstock wrote: Don't get me wrong; I'm all for improving the SA scalability; there's no quick solution to this AFAIK. It would be interesting to see an apples to apples comparison of OpenSM and proprietary SMs in terms of running on the same hardware and the transaction rate for various things. I think this warrants an open discussion if people are serious about working on this issue. Agreed. I agree that this set of issues has come up many times before on the list; it will be interesting to see if anyone will *do* anything about it this time. :-) (obviously, I'm only interested as a consumer of the end result) It might also be worthwhile to start a whole new discussion about making a better CM (at least from the ULP perspective). One that offers simple mechanisms for those who don't need/care about the details, but also offers complex/detailed mechanisms (perhaps remarkably like today's mechanisms). I've heard similar comments before but this too will take significant where-with-all IMO. Ditto my above remarks. :-) -- Jeff Squyres Cisco Systems ___ ewg mailing list ewg@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg