[Ofvwg] OpenFabrics Verbs Working Group meeting minutes 3/3/15

[Gilad Shainer]

ORNL
Mellanox
Chelsio
Oracle
LLNL
Unisys
Broadcom
Intel
OSU
LANL


Accelerated Verbs discussion

-          Liran presented

-          Are accelerated Verbs applicable to multiple architectures?

o   As an interface specification, there is no dependency on a specific architecture

-          Concern re. type vs. OOP approach

o   The robustness benefits of an OOP approach seem to out-weight the relative overhead, which seems to be negligible in real-world applications since apps already hold additional context information related to the notion of an endpoint.

o   In any case, we should quantify these tradeoffs with actual empirical data

-          Accelerated family completeness and inter-dependencies

o   If a provider supports an interface, all functions contained in this interface must be implemented

o   No explicit dependencies are defined between families, so providers are free to mix and match the families that are supported

-          Version info seems useful

o   To relate between 2 families that one is an extension to the other

o   Each new version must return all functions of previous versions

?  Similar to OOP single-inheritance

o   Another advantage is that code simplicity is maintained regardless of the version of an interface

-          Device capabilities and Accelerated Verbs:

o   Accelerated verbs do not introduce new features that aren't already supported according to an existing provider device capabilities

?  This means that applications can *always* fall back to the standard Verbs supported by a given provider

o   While the granularity of device capabilities are implicit by the definition of the specific capability, the granularity of Accelerated Verbs are explicit

?  The query_intf() call could be used both to query availability and obtain the interface

?  If an interface is returned, it is always complete

o   If query_intf() fails, it could be useful to elaborate on the reason of the failure. For example:

?  Interface not supported by the provider

?  Incorrect object type

?  Incorrect version (+ latest supported version)

?  Incorrect object state

-          It is important to standardize the name-space of interface families in advance

o   So programmers can know about common prefixes

o   To prevent name-space collisions between vendors

Verbs extensions discussion

-          Tzahi Oved presented

-          Tzahi gave a brief introduction

-          The discussion will continue next week

Gilad proposed to take UCX, a SHMEM communication library being developed by ORNL and other organizations, bring it into the working group and use it as the first case for designing Accelerated Verbs

-          UCX is library that will be used in future systems and for future applications

-          Fits the charter of the working group

-          A great candidate for enhancing the verbs interface

-          Potentially can be included in the working group distribution

-          There was a discussion in the WG, no objection raised, Pasha and Liran supported the proposal, so approved



Regards,
Gilad Shainer
Vice President, Marketing
Mellanox Technologies
350 Oakmead Parkway, Suite 100, Sunnyvale CA, 94085
Office: 408-916-0048, Mobile: 408-421-0048, Fax: 408-585-0348

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ofvwg/attachments/20150304/73de8dc3/attachment.html>