osdir.com

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Using CUDA enabled pyarrow


hi Pearu -- yes, I had thought of this work working on the arrow_gpu
library. Some time ago I opened

https://issues.apache.org/jira/browse/ARROW-1470

thinking that it would be good to combine the MemoryPool* concept and
the AllocateBuffer concept into a single abstract interface. Such an
interface for CUDA could also optimize small allocations by allocating
larger "pages" if desired. So Before adding a CudaMemoryPool we should
consider if we want to define a BufferAllocator interface
On Thu, Oct 4, 2018 at 5:04 AM Pearu Peterson
<pearu.peterson@xxxxxxxxxxxxx> wrote:
>
> Hi,
> Currently, the arrow host memory management includes MemoryPool to
> accelerate memory operations (new/free).
> Would there be interest in supporting the same concept in CUDA memory
> management to reduce the overhead of cudaMalloc/cudaFree?
> Best regards,
> Pearu
>
> On Wed, Oct 3, 2018 at 11:44 PM Pearu Peterson <pearu.peterson@xxxxxxxxxxxxx>
> wrote:
>
> > Hi,
> > I can make the initial design document from the existing comments.
> > Do you have examples of some earlier design documents used for similar
> > purpose? Would shared google docs be OK?
> >
> > Btw, I also figured out an answer to my original question, here is a
> > partial codelet for accessing the batch columns that I was missing:
> >
> > cbuf = <CudaBuffer instance>
> > cbatch = pa.cuda.read_record_batch(cbuf, schema)
> > for col in cbatch:
> >     null_buf, data_buf = col.buffers()
> >     cdata_buf = CudaBuffer.from_buffer(data_buf)
> >     if null_buf is not None: ...
> >     ...
> >
> > This is used in CudaNDArray that allows accessing the items from host,
> > very similar to DeviceNDArray of numba.cuda:
> >   https://github.com/Quansight/pygdf/blob/arrow-cuda/pygdf/cudaarray.py
> > (excuse the coding, its wip and experimental)
> >
> > Best regards,
> > Pearu
> >
> >
> >
> >
> > On Wed, Oct 3, 2018 at 11:29 PM Wes McKinney <wesmckinn@xxxxxxxxx> wrote:
> >
> >> What are the action items on this? Sounds like we need to start a
> >> design document. I'm afraid I don't have the bandwidth to champion GPU
> >> functionality at the moment but I will participate in design
> >> discussions and help break down complex tasks into more accessible
> >> JIRA issues.
> >>
> >> Thanks
> >> Wes
> >> On Fri, Sep 28, 2018 at 9:44 AM Wes McKinney <wesmckinn@xxxxxxxxx> wrote:
> >> >
> >> > Seems like there is a fair bit of work to do to specify APIs and
> >> > semantics. I suggest we create a Google document or something
> >> > collaborative where we can enumerate and discuss the issues we want to
> >> > resolve, and then make a list of the concrete development.
> >> >
> >> > The underlying problem IMHO in ARROW-2446 is that we do not have the
> >> > notion of device. An instance of CudaBuffer is only necessary so that
> >> > the appropriate virtual dtor can be invoked to release the memory. As
> >> > long as a buffer referencing it is aware of the underlying device,
> >> > then our code can dispatch to the correct code paths. At the moment we
> >> > can only really detect whether an arrow::Buffer* is a device buffer by
> >> > dynamic_cast, and then that is not reliable because we may be a slice
> >> > On Fri, Sep 28, 2018 at 7:17 AM Pearu Peterson
> >> > <pearu.peterson@xxxxxxxxxxxxx> wrote:
> >> > >
> >> > > Hi Wes,
> >> > >
> >> > > Yes, it makes sense.
> >> > >
> >> > > If I understand you correctly then defining a device abstraction
> >> would also
> >> > > bring Buffer and CudaBuffer under the same umbrella (that would be
> >> opposite
> >> > > approach to ARROW-2446, btw).
> >> > >
> >> > > This issue is also related to
> >> > >   https://github.com/dmlc/dlpack/blob/master/include/dlpack/dlpack.h
> >> > > that defines a specification for data locality (for ndarrays but the
> >> > > concept is the same for buffers).
> >> > >
> >> > > ARROW-2447 defines API that uses Buffer::cpu_data(), hence also
> >> > > Buffer::cuda_data(), Buffer::disk_data() etc.
> >> > >
> >> > > I would like to propose a more general model (no guarantees that it
> >> would
> >> > > make sense implementation-wise :) ):
> >> > > 0. CPU would be considered as any other device (this would be in line
> >> with
> >> > > dlpack). To name few devices: HOST, CUDA, DISK, FPGA, etc. and why not
> >> > > remote databases defined by URL.
> >> > > 1. A device is defined as a unit that has (i) a memory for holding
> >> data,
> >> > > and (ii) it may have a processor(s) for processing the data
> >> (computations).
> >> > > For instance, HOST device has RAM and CPU(s); a CUDA device has device
> >> > > memory and GPU(s); a DISK device has memory but no processing unit,
> >> etc.
> >> > > 2. Different devices can access other devices memory using the same
> >> API
> >> > > methods (say, Buffer.data()). For processing the data by a device (in
> >> case
> >> > > the device has a processor), the data is copied to device memory
> >> on-demand,
> >> > > unless the data is stored in the same device as the the processor. For
> >> > > instance, for processing the CUDA data with CPU, HOST device would
> >> need to
> >> > > copy CUDA device data to HOST memory (that works currently) and
> >> vice-versa
> >> > > (that works as well, e.g. using CudaHostBuffer). In another setup,
> >> CUDA
> >> > > device might need to use data from DISK: according to this proposal,
> >> the
> >> > > DISK data would be copied directly to CUDA device (bypassing HOST
> >> memory if
> >> > > technically possible).
> >> > > So, in short, the implementation has to check whether the processor
> >> and the
> >> > > memory are on the same device before processing the data, if not, the
> >> data
> >> > > is copied using the on-demand approach. By on-demand approach, I mean
> >> that
> >> > > the data references are passed around as a pair: (device id, device
> >> > > pointer).
> >> > > 3. All the above is controlled from a master device process. Usually,
> >> the
> >> > > master device would be HOST, but it does not have to be always so.
> >> > >
> >> > > PS: I realize that this discussion diverges from the original
> >> subject, feel
> >> > > free to rename the subject if needed.
> >> > >
> >> > > Best regards,
> >> > > Pearu
> >> > >
> >> > >
> >> > >
> >> > >
> >> > >
> >> > >
> >> > >
> >> > > On Fri, Sep 28, 2018 at 12:49 PM Wes McKinney <wesmckinn@xxxxxxxxx>
> >> wrote:
> >> > >
> >> > > > hi Pearu,
> >> > > >
> >> > > > Yes, I think it would be a good idea to develop some tools to make
> >> > > > interacting with device memory using the existing data structures
> >> work
> >> > > > seamlessly.
> >> > > >
> >> > > > This is all closely related to
> >> > > >
> >> > > > https://issues.apache.org/jira/browse/ARROW-2447
> >> > > >
> >> > > > I would say step 1 would be defining the device abstraction. Then we
> >> > > > can add methods or properties to the data structures in pyarrow to
> >> > > > show the location of the memory, whether CUDA or host RAM, etc. We
> >> > > > could also have a memory-mapped device for memory maps to be able to
> >> > > > communicate that data is on disk. We could then define virtual APIs
> >> > > > for host-side data access to ensure that memory is copied to the
> >> host
> >> > > > if needed (e.g. in the case of indexing into the values of an array)
> >> > > >
> >> > > > There are some small details around the handling of device in the
> >> case
> >> > > > of hierarchical memory references. So if we say
> >> `buffer->GetDevice()`
> >> > > > then even if it's a sliced buffer (which will be the case after
> >> using
> >> > > > any IPC reader APIs), it needs to return the right device. This
> >> means
> >> > > > that we probably need to define a SlicedBuffer type that delegates
> >> > > > GetDevice() calls to the parent buffer
> >> > > >
> >> > > > Let me know if what I'm saying makes sense. Kou and Antoine probably
> >> > > > have some thoughts about this also.
> >> > > >
> >> > > > - Wes
> >> > > > On Fri, Sep 28, 2018 at 5:34 AM Pearu Peterson
> >> > > > <pearu.peterson@xxxxxxxxxxxxx> wrote:
> >> > > > >
> >> > > > > Hi,
> >> > > > >
> >> > > > > Consider the following use case:
> >> > > > >
> >> > > > > schema = <pa.Schema instance>
> >> > > > > cbuf = <pa.cuda.CudaBuffer instance>
> >> > > > > cbatch = pa.cuda.read_record_batch(schema, cbuf)
> >> > > > >
> >> > > > > Note that cbatch is pa.RecordBatch instance where data pointers
> >> are
> >> > > > device
> >> > > > > pointers.
> >> > > > >
> >> > > > > for col in cbatch.columns:
> >> > > > >     # here col is, say, FloatArray, that data pointer is a device
> >> pointer
> >> > > > >     # as a result, accessing col data, say, taking a slice, leads
> >> to
> >> > > > > segfaults
> >> > > > >     print(col[0])
> >> > > > >
> >> > > > > The aim of this message would be establishing a user-friendly way
> >> to
> >> > > > > access, say, a slice of the device data so that only the
> >> requested data
> >> > > > is
> >> > > > > copied to host.
> >> > > > >
> >> > > > > Or more generally, should there be a CUDA specific RecordBatch
> >> that
> >> > > > > implements RecordBatch API that can be used from host?
> >> > > > >
> >> > > > > For instance, this would be similar to DeviceNDArray in numba that
> >> > > > > basically implements ndarray API for device data while the API
> >> can be
> >> > > > used
> >> > > > > from host.
> >> > > > >
> >> > > > > What do you think? What would be the proper approach? (I can do
> >> the
> >> > > > > implementation).
> >> > > > >
> >> > > > > Best regards,
> >> > > > > Pearu
> >> > > >
> >>
> >