On 10/12/2017 10:48 AM, Francois Ozog wrote:
On 12 October 2017 at 16:57, Jonathan Cameron
<Jonathan.Cameron@xxxxxxxxxx> wrote:
On Thu, 12 Oct 2017 08:31:36 -0500
Douglas Miller <dougmill@xxxxxxxxxxxxxxxxxx> wrote:
Not sure if you're already plugged-in to this, but the OpenMP group is
(has been) working on Accelerator support.
http://www.openmp.org/updates/openmp-accelerator-support-gpus/
Maybe you are talking about a different aspect of accelerator support,
but it seems prudent to involve OpenMP as much as makes sense.
That's certainly interesting and sits in the area of 'standard'
userspace code but it is (I think) really addressing only one aspect
of the wider support problem.
I do like the emphasis on balancing between CPU and accelerator,
that is certainly an open question even a the lowest levels in
areas such as cryptography acceleration where you either run
out of hardware resources on your accelerator or you actually
have a usage pattern that would be quicker on the CPU due
to inherent overheads in (current) non cpu crypto engines.
Thanks for the pointer. I can see we are going to need some location
for resources like this to be gathered together.
Jonathan
On 10/12/2017 12:22 AM, Andrew Donnellan wrote:
On 10/10/17 22:28, Jonathan Cameron wrote:
Hi All,
Please forward this email to anyone you think may be interested.
Have forwarded this to a number of relevant IBMers.
On behalf of Huawei, I am looking into options to foster a wider
community
around the various ongoing projects related to Accelerator support
within
Linux. The particular area of interest to Huawei is that of harnessing
accelerators from userspace, but in a collaborative way with the kernel
still able to make efficient use of them, where appropriate.
We are keen to foster a wider community than one just focused on
our own current technology. This is a field with no clear answers,
so the
widest possible range of input is needed!
The address list of this email is drawn from people we have had
discussions
with or who have been suggested in response to Kenneth Lee's wrapdrive
presentation at Linaro Connect and earlier presentations on the more
general
issue. A few relevant lists added to hopefully catch anyone we missed.
My apologies to anyone who got swept up in this and isn't interested!
Here we are defining accelerators fairly broadly - suggestions for a
better
term are also welcome.
The infrastructure may be appropriate for:
* Traditional offload engines - cryptography, compression and similar
* Upcoming AI accelerators
* ODP type requirements for access to elements of networking
* Systems utilizing SVM including CCIX and other cache coherent buses
* Many things we haven't thought of yet...
As I see it, there are several aspects to this:
1) Kernel drivers for accelerators themselves.
* Traditional drivers such as crypto etc
- These already have their own communities. The main
focus of such work will always be through them.
- What a more general community could add here would be an
overview of the shared infrastructure of such devices.
This is particularly true around VFIO based (or similar)
userspace interfaces with a non trivial userspace component.
* How to support new types of accelerator?
2) The need for lightweight access paths from userspace that 'play
well' and
share resources etc with standard in-kernel drivers. This is the
area
that Kenneth Lee and Huawei have been focusing on with their
wrapdrive
effort. We know there are other similar efforts going on in other
companies.
* This may involve interacting with existing kernel communities
such as
those around VFIO and mdev.
* Resource management when we may have many consumers - not all
hardware
has appropriate features to deal with this.
3) Usecases for accelerators. e.g.
* kTLS
* Storage encryption
* ODP - networking dataplane
* AI toolkits
Discussions we want to get started include:
* A wider range of hardware than we are currently considering. What
makes
sense to target / what hardware do people have they would like to
support?
* Upstream paths - potential blockers and how to overcome them. The
standard
kernel drivers should be fairly straightforward, but once we start
looking at
systems with a heavier userspace component, things will get more
controversial!
* Fostering stronger userspace communities to allow these these
accelerators
to be easily harnessed.
So as ever with a linux community focusing on a particular topic, the
obvious solution is a mailing list. There are a number of options on how
do this.
1) Ask one of the industry bodies to host? Who?
2) Put together a compelling argument for
linux-accelerators@xxxxxxxxxxxxxxx
as probably the most generic location for such a list.
Happy to offer linux-accelerators@xxxxxxxxxxxxxxxx, which I can get
set up immediately (and if we want patchwork, patchwork.ozlabs.org is
available as always, no matter where the list is hosted).
More open questions are
1) Scope?
* Would anyone ever use such an overarching list?
* Are we better off with the usual adhoc list of 'interested
parties' + lkml?
* Do we actually need to define the full scope - are we better with
a vague
definition?
I think a list with a broad and vaguely defined scope is a good idea -
it would certainly be helpful to us to be able to follow what other
contributors are doing that could be relevant to our CAPI and OpenCAPI
work.
2) Is there an existing community we can use to discuss these issues?
(beyond the obvious firehose of LKML).
3) Who else to approach for input on these general questions?
In parallel to this there are elements such as git / patchwork etc but
they can all be done as they are needed.
Thanks
--
Jonathan Cameron
Huawei
I'd like to keep sharing thoughts on this.
I understand accelerators can be fixed/parameterized, reconfigurable
(FPGA), programmable (GPUs, NPUs...).
With that in mind, there is a preparation phase that can as simple as
set some parameters, or as complex as loading a "kernel" to a GPU or
send a bitstream to an FPGA.
In some cases, there may even be a slicing phase where the accelerator
is actually sliced to accommodate different "customers" on the host it
serves.
Then there is the data supply to the accelerator.
Is it fair to say that one of the main concerns of your proposal is to
focus on having the userland data supply to the accelerator be as
native/direct as possible ?
And if so, then OpenMP would be a user of the userland IO framework
when it comes to data supply?
It also reminds me some work done by the media community and GStreamer
arround DMA buf which specializes in a domain where large video
"chunks" passes from one functional block to the other with specific
caching policies (write combining is a friend here). While for 100Gbps
networking were we need to handle 142Mpps the nature of the datapath
is very different.
Would you like to address both classes of problems? (I mean class 1:
large chunks of data to be shared between few consummers; class 2:
very large number of small chunks of data shared with a few to a large
number of consumers?)
I've been out of touch with OpenMP for a number of years now, but that
standard is a programming paradigm, and not (necessarily) limited to
userland (or kernel). My reason for bringing it up is to make sure the
right people get involved to help keep OpenMP relevant for things like
CAPI and intended uses in the kernel. I believe the intent of OpenMP is
to create a paradigm that will work is (most) all cases.
