[Ofa_boardplus] Jason's draft preso to the Linux Plumbers' Conference

Woodruff, Robert J robert.j.woodruff at intel.com
Fri Sep 1 07:36:06 PDT 2017


There are actually a number of new RDMA devices coming on the market recently, such as the new Cavium and Broadcom Roce adapters and the new Intel iWarp NIC that is integrated into the chipset of the recently released Xeon (skylake) platform. I know that Intel is still interested in interop testing between the new i40 NIC and the Chelsio iWarp NICs. 
I believe that Cavium has also been participating in various debug events. Not sure about Broadcom. Check with Paul Bowden and/or Rupert Dance. They would have the details on who is most active in the events. 

As for InfiniBand, you are correct, Mellanox is pretty much the sole provider of that technology going forward. 
Intel does have and older InfiniBand adapter, but it not sold anymore and just in sustaining mode since Intel released its Omni-Path fabric products. 

-----Original Message-----
From: Ofa_boardplus [mailto:ofa_boardplus-bounces at lists.openfabrics.org] On Behalf Of Christoph Lameter
Sent: Thursday, August 31, 2017 5:38 PM
To: Jim Ryan <jimdryan at gmail.com>
Cc: ofa_boardplus at lists.openfabrics.org, <ofa_boardplus at lists.openfabrics.org>
Subject: Re: [Ofa_boardplus] Jason's draft preso to the Linux Plumbers' Conference

Which companies are participating in the IWG? It seems that Mellanox is at this point the overwhelming provider of most of the Infiniband technology with just minor niches occupied by others. Our own experience with trying to use technology from Qlogic on a single system (2011) failed miserably. Intel is moving to Omnipath. Others simply focus on Ethernet to have a vendor neutral shared standard.

It seems to us therefore (Jump Trading) that there is a trend for interconnect technology to become vendor specific whereas the APIs to access the interconnect (RDMA subsystem in the kernel) provide a common way of using these fabrics by essentially providing a hardware abstraction layer for fabrics in the Linux Kernel.

Given this I would like to see a list of participants in the IWG and if its is as dominated as I think it is by Mellanox then we should slowly shut down the program and focus on the software layer while continuing to provide a forum for the exchange of innovative ideas about fabrics.



On Thu, Aug 31, 2017 at 7:28 PM, Jim Ryan <jimdryan at gmail.com> wrote:
> Ok, I think I see your point. Your consciously blending membership 
> dues with interop program participation to make a point. I don't want 
> to argue that point, but I do want to be painfully clear about 
> something. The approach we take is conscious and, for example, 
> specifically contrary to, for example the IBTA. We view interop as a 
> program of value to participants who make the biz decision to fund it. 
> The IBTA, in contrast, views testing as a member benefit available to all based on membership dues.
>
> Those of us who have been involved with this issue, actually for 
> years, share your frustration, if I may characterize it as such, with 
> the apparent indifference on the part of the vast majority of members. 
> There have been important contributions in the past in the form of 
> donated equipment, but I'll quickly ack this has been from a small 
> number of donors and not recently. The actual components are, AFAIK, 
> left behind after testing. I have requested another call for donations 
> but, for whatever reason, that hasn't happened.
>
> I *do* have to ask you to not use terms along the lines of "membership 
> funding"; there is no such thing. There is participant funding and a 
> degree of "membership" indifference. I'm trying to ack an element of 
> your argument but continue to make the distinction clear.
>
> Re the quality of testing, that's a challenge for the IWG. One of, if 
> not
> *the* most important thing they're responsible for is quality of 
> testing. If something is broken there, I'm not aware of it, and we 
> need to come to understand this.
>
> Finally, I realized I failed to respond to a point you made earlier. 
> It's kinda delicate, but important. The OFA is specifically not 
> "chartered" to develop specs and the IBTA and others are. There are IP 
> provisions that need to exist if this is part of our mission or not. I 
> can give you boring details if you want to hear more.
>
> The reason this is delicate is because the OFIWG has had to go right 
> to the edge of what we can do, to use MAN pages to document expected 
> API functionality. We have agreed this is short of a spec, but you get 
> the point; it's a fine but important distinction.
>
> I hope this helps, Jim
>
> On Thu, Aug 31, 2017 at 4:50 PM, Jason Gunthorpe 
> <jgunthorpe at obsidianresearch.com> wrote:
>>
>> On Thu, Aug 31, 2017 at 04:05:49PM -0700, Jim Ryan wrote:
>>
>> > First, Jason, it appears you're working under a misunderstanding 
>> > which I think Bill tried to correct, but perhaps he could have used 
>> > different words. The interop program is self-funding in the sense 
>> > the cost of the testing is borne
>>
>> No, I do understand this very well.
>>
>> The fact IWG participation is optional, and that it is billed on a 
>> cost recovery basis, is not relevant to my point. I am observing that 
>> of all the funds the membership chooses to send to the OFA, 50% are 
>> directed to UNH-IOL (by the direct choice of the membership).
>>
>> While at the same time the membership cannot be bothered to properly 
>> equip UNH-IOL to actually test the software, and does not seem 
>> interested in the logo program.
>>
>> So, exactly why, is the membership choosing to continue fund this?
>>
>> > Re the logo program being out of date and maybe valueless, I can 
>> > simply say two things. I know of vendors for whom this *testing* is 
>> > extremely important. I view that info as being confidential so as 
>> > loathsome as it is to me, I have to ask you to trust me. Notice I'm 
>> > not making the same claim for the logo.
>>
>> From what I can see the testing is far less useful than I assumed it 
>> was. Maybe your sources are also operating under poor assumptions?
>>
>> Jason
>
>
>
> _______________________________________________
> Ofa_boardplus mailing list
> Ofa_boardplus at lists.openfabrics.org
> http://lists.openfabrics.org/mailman/listinfo/ofa_boardplus
>
_______________________________________________
Ofa_boardplus mailing list
Ofa_boardplus at lists.openfabrics.org
http://lists.openfabrics.org/mailman/listinfo/ofa_boardplus


More information about the Ofa_boardplus mailing list