GDRCopy
A low-latency GPU memory copy library based on NVIDIA GPUDirect RDMA technology.
Introduction
While GPUDirect RDMA is meant for direct access to GPU memory from third-party devices, it is possible to use these same APIs to create perfectly valid CPU mappings of the GPU memory.
The advantage of a CPU driven copy is the very small overhead involved. That might be useful when low latencies are required.
What is inside
GDRCopy offers the infrastructure to create user-space mappings of GPU memory, which can then be manipulated as if it was plain host memory (caveats apply here).
A simple by-product of it is a copy library with the following characteristics:
very low overhead, as it is driven by the CPU. As a reference, currently a cudaMemcpy can incur in a 6-…
GDRCopy
A low-latency GPU memory copy library based on NVIDIA GPUDirect RDMA technology.
Introduction
While GPUDirect RDMA is meant for direct access to GPU memory from third-party devices, it is possible to use these same APIs to create perfectly valid CPU mappings of the GPU memory.
The advantage of a CPU driven copy is the very small overhead involved. That might be useful when low latencies are required.
What is inside
GDRCopy offers the infrastructure to create user-space mappings of GPU memory, which can then be manipulated as if it was plain host memory (caveats apply here).
A simple by-product of it is a copy library with the following characteristics:
very low overhead, as it is driven by the CPU. As a reference, currently a cudaMemcpy can incur in a 6-7us overhead.
An initial memory pinning phase is required, which is potentially expensive, 10us-1ms depending on the buffer size.
Fast H-D, because of write-combining. H-D bandwidth is 6-8GB/s on Ivy Bridge Xeon but it is subject to NUMA effects.
Slow D-H, because the GPU BAR, which backs the mappings, can’t be prefetched and so burst reads transactions are not generated through PCIE
The library comes with a few tests like:
- gdrcopy_sanity, which contains unit tests for the library and the driver.
- gdrcopy_copybw, a minimal application which calculates the R/W bandwidth for a specific buffer size.
- gdrcopy_copylat, a benchmark application which calculates the R/W copy latency for a range of buffer sizes.
- gdrcopy_apiperf, an application for benchmarking the latency of each GDRCopy API call.
- gdrcopy_pplat, a benchmark application which calculates the round-trip ping-pong latency between GPU and CPU.
Requirements
GPUDirect RDMA requires NVIDIA Data Center GPU or NVIDIA RTX GPU (formerly Tesla and Quadro) based on Kepler or newer generations, see GPUDirect RDMA. For more general information, please refer to the official GPUDirect RDMA design document.
The device driver requires GPU display driver >= 418.40 on ppc64le and >= 331.14 on other platforms. The library and tests require CUDA >= 6.0.
DKMS is a prerequisite for installing GDRCopy kernel module package. On RHEL or SLE, however, users have an option to build kmod and install it instead of the DKMS package. See Build and installation section for more details.
# On RHEL
# dkms can be installed from epel-release. See https://fedoraproject.org/wiki/EPEL.
$ sudo yum install dkms
# On Debian - No additional dependency
# On SLE / Leap
# On SLE dkms can be installed from PackageHub.
$ sudo zypper install dkms rpmbuild
CUDA and GPU display driver must be installed before building and/or installing GDRCopy. The installation instructions can be found in https://developer.nvidia.com/cuda-downloads.
GPU display driver header files are also required. They are installed as a part of the driver (or CUDA) installation with runfile. If you install the driver via package management, we suggest
- On RHEL,
sudo dnf module install nvidia-driver:latest-dkms. - On Debian,
sudo apt install nvidia-dkms-<your-nvidia-driver-version>. - On SLE,
sudo zypper install nvidia-gfx<your-nvidia-driver-version>-kmp.
The supported architectures are Linux x86_64, ppc64le, and arm64. The supported platforms are RHEL8, RHEL9, Ubuntu20_04, Ubuntu22_04, SLE-15 (any SP) and Leap 15.x.
Root privileges are necessary to load/install the kernel-mode device driver.
Build and installation
We provide three ways for building and installing GDRCopy.
rpm package
# For RHEL:
$ sudo yum groupinstall 'Development Tools'
$ sudo yum install dkms rpm-build make
# For SLE:
$ sudo zypper in dkms rpmbuild
$ cd packages
$ CUDA=<cuda-install-top-dir> ./build-rpm-packages.sh
$ sudo rpm -Uvh gdrcopy-kmod-<version>dkms.noarch.<platform>.rpm
$ sudo rpm -Uvh gdrcopy-<version>.<arch>.<platform>.rpm
$ sudo rpm -Uvh gdrcopy-devel-<version>.noarch.<platform>.rpm
DKMS package is the default kernel module package that build-rpm-packages.sh generates. To create kmod package, -m option must be passed to the script. Unlike the DKMS package, the kmod package contains a prebuilt GDRCopy kernel module which is specific to the NVIDIA driver version and the Linux kernel version used to build it.
deb package
$ sudo apt install build-essential devscripts debhelper fakeroot pkg-config dkms
$ cd packages
$ CUDA=<cuda-install-top-dir> ./build-deb-packages.sh
$ sudo dpkg -i gdrdrv-dkms_<version>_<arch>.<platform>.deb
$ sudo dpkg -i libgdrapi_<version>_<arch>.<platform>.deb
$ sudo dpkg -i gdrcopy-tests_<version>_<arch>.<platform>.deb
$ sudo dpkg -i gdrcopy_<version>_<arch>.<platform>.deb
from source
$ make prefix=<install-to-this-location> CUDA=<cuda-install-top-dir> all install
$ sudo ./insmod.sh
Notes
Compiling the gdrdrv driver requires the NVIDIA driver source code, which is typically installed at /usr/src/nvidia-<version>. Our make file automatically detects and picks that source code. In case there are multiple versions installed, it is possible to pass the correct path by defining the NVIDIA_SRC_DIR variable, e.g. export NVIDIA_SRC_DIR=/usr/src/nvidia-520.61.05/nvidia before building the gdrdrv module.
There are two major flavors of NVIDIA driver: 1) proprietary, and 2) opensource. We detect the flavor when compiling gdrdrv based on the source code of the NVIDIA driver. Different flavors come with different features and restrictions:
- gdrdrv compiled with the opensource flavor will provide functionality and high performance on all platforms. However, you will not be able to load this gdrdrv driver when the proprietary NVIDIA driver is loaded.
- gdrdrv compiled with the proprietary flavor can always be loaded regardless of the flavor of NVIDIA driver you have loaded. However, it may have suboptimal performance on coherent platforms such as Grace-Hopper. Functionally, it will not work correctly on Intel CPUs with Linux kernel built with confidential compute (CC) support, i.e.
CONFIG_ARCH_HAS_CC_PLATFORM=y, WHEN CC is enabled at runtime.
Tests
Execute provided tests:
$ gdrcopy_sanity
Total: 28, Passed: 28, Failed: 0, Waived: 0
List of passed tests:
basic_child_thread_pins_buffer_cumemalloc
basic_child_thread_pins_buffer_vmmalloc
basic_cumemalloc
basic_small_buffers_mapping
basic_unaligned_mapping
basic_vmmalloc
basic_with_tokens
data_validation_cumemalloc
data_validation_vmmalloc
invalidation_access_after_free_cumemalloc
invalidation_access_after_free_vmmalloc
invalidation_access_after_gdr_close_cumemalloc
invalidation_access_after_gdr_close_vmmalloc
invalidation_fork_access_after_free_cumemalloc
invalidation_fork_access_after_free_vmmalloc
invalidation_fork_after_gdr_map_cumemalloc
invalidation_fork_after_gdr_map_vmmalloc
invalidation_fork_child_gdr_map_parent_cumemalloc
invalidation_fork_child_gdr_map_parent_vmmalloc
invalidation_fork_child_gdr_pin_parent_with_tokens
invalidation_fork_map_and_free_cumemalloc
invalidation_fork_map_and_free_vmmalloc
invalidation_two_mappings_cumemalloc
invalidation_two_mappings_vmmalloc
invalidation_unix_sock_shared_fd_gdr_map_cumemalloc
invalidation_unix_sock_shared_fd_gdr_map_vmmalloc
invalidation_unix_sock_shared_fd_gdr_pin_buffer_cumemalloc
invalidation_unix_sock_shared_fd_gdr_pin_buffer_vmmalloc
$ gdrcopy_copybw
GPU id:0; name: Tesla V100-SXM2-32GB; Bus id: 0000:06:00
GPU id:1; name: Tesla V100-SXM2-32GB; Bus id: 0000:07:00
GPU id:2; name: Tesla V100-SXM2-32GB; Bus id: 0000:0a:00
GPU id:3; name: Tesla V100-SXM2-32GB; Bus id: 0000:0b:00
GPU id:4; name: Tesla V100-SXM2-32GB; Bus id: 0000:85:00
GPU id:5; name: Tesla V100-SXM2-32GB; Bus id: 0000:86:00
GPU id:6; name: Tesla V100-SXM2-32GB; Bus id: 0000:89:00
GPU id:7; name: Tesla V100-SXM2-32GB; Bus id: 0000:8a:00
selecting device 0
testing size: 131072
rounded size: 131072
gpu alloc fn: cuMemAlloc
device ptr: 7f1153a00000
map_d_ptr: 0x7f1172257000
info.va: 7f1153a00000
info.mapped_size: 131072
info.page_size: 65536
info.mapped: 1
info.wc_mapping: 1
page offset: 0
user-space pointer:0x7f1172257000
writing test, size=131072 offset=0 num_iters=10000
write BW: 9638.54MB/s
reading test, size=131072 offset=0 num_iters=100
read BW: 530.135MB/s
unmapping buffer
unpinning buffer
closing gdrdrv
$ gdrcopy_copylat
GPU id:0; name: Tesla V100-SXM2-32GB; Bus id: 0000:06:00
GPU id:1; name: Tesla V100-SXM2-32GB; Bus id: 0000:07:00
GPU id:2; name: Tesla V100-SXM2-32GB; Bus id: 0000:0a:00
GPU id:3; name: Tesla V100-SXM2-32GB; Bus id: 0000:0b:00
GPU id:4; name: Tesla V100-SXM2-32GB; Bus id: 0000:85:00
GPU id:5; name: Tesla V100-SXM2-32GB; Bus id: 0000:86:00
GPU id:6; name: Tesla V100-SXM2-32GB; Bus id: 0000:89:00
GPU id:7; name: Tesla V100-SXM2-32GB; Bus id: 0000:8a:00
selecting device 0
device ptr: 0x7fa2c6000000
allocated size: 16777216
gpu alloc fn: cuMemAlloc
map_d_ptr: 0x7fa2f9af9000
info.va: 7fa2c6000000
info.mapped_size: 16777216
info.page_size: 65536
info.mapped: 1
info.wc_mapping: 1
page offset: 0
user-space pointer: 0x7fa2f9af9000
gdr_copy_to_mapping num iters for each size: 10000
WARNING: Measuring the API invocation overhead as observed by the CPU. Data
might not be ordered all the way to the GPU internal visibility.
Test Size(B) Avg.Time(us)
gdr_copy_to_mapping 1 0.0889
gdr_copy_to_mapping 2 0.0884
gdr_copy_to_mapping 4 0.0884
gdr_copy_to_mapping 8 0.0884
gdr_copy_to_mapping 16 0.0905
gdr_copy_to_mapping 32 0.0902
gdr_copy_to_mapping 64 0.0902
gdr_copy_to_mapping 128 0.0952
gdr_copy_to_mapping 256 0.0983
gdr_copy_to_mapping 512 0.1176
gdr_copy_to_mapping 1024 0.1825
gdr_copy_to_mapping 2048 0.2549
gdr_copy_to_mapping 4096 0.4366
gdr_copy_to_mapping 8192 0.8141
gdr_copy_to_mapping 16384 1.6155
gdr_copy_to_mapping 32768 3.2284
gdr_copy_to_mapping 65536 6.4906
gdr_copy_to_mapping 131072 12.9761
gdr_copy_to_mapping 262144 25.9459
gdr_copy_to_mapping 524288 51.9100
gdr_copy_to_mapping 1048576 103.8028
gdr_copy_to_mapping 2097152 207.5990
gdr_copy_to_mapping 4194304 415.2856
gdr_copy_to_mapping 8388608 830.6355
gdr_copy_to_mapping 16777216 1661.3285
gdr_copy_from_mapping num iters for each size: 100
Test Size(B) Avg.Time(us)
gdr_copy_from_mapping 1 0.9069
gdr_copy_from_mapping 2 1.7170
gdr_copy_from_mapping 4 1.7169
gdr_copy_from_mapping 8 1.7164
gdr_copy_from_mapping 16 0.8601
gdr_copy_from_mapping 32 1.7024
gdr_copy_from_mapping 64 3.1016
gdr_copy_from_mapping 128 3.4944
gdr_copy_from_mapping 256 3.6400
gdr_copy_from_mapping 512 2.4394
gdr_copy_from_mapping 1024 2.8022
gdr_copy_from_mapping 2048 4.6615
gdr_copy_from_mapping 4096 7.9783
gdr_copy_from_mapping 8192 14.9209
gdr_copy_from_mapping 16384 28.9571
gdr_copy_from_mapping 32768 56.9373
gdr_copy_from_mapping 65536 114.1008
gdr_copy_from_mapping 131072 234.9382
gdr_copy_from_mapping 262144 496.4011
gdr_copy_from_mapping 524288 985.5196
gdr_copy_from_mapping 1048576 1970.7057
gdr_copy_from_mapping 2097152 3942.5611
gdr_copy_from_mapping 4194304 7888.9468
gdr_copy_from_mapping 8388608 18361.5673
gdr_copy_from_mapping 16777216 36758.8342
unmapping buffer
unpinning buffer
closing gdrdrv
$ gdrcopy_apiperf -s 8
GPU id:0; name: Tesla V100-SXM2-32GB; Bus id: 0000:06:00
GPU id:1; name: Tesla V100-SXM2-32GB; Bus id: 0000:07:00
GPU id:2; name: Tesla V100-SXM2-32GB; Bus id: 0000:0a:00
GPU id:3; name: Tesla V100-SXM2-32GB; Bus id: 0000:0b:00
GPU id:4; name: Tesla V100-SXM2-32GB; Bus id: 0000:85:00
GPU id:5; name: Tesla V100-SXM2-32GB; Bus id: 0000:86:00
GPU id:6; name: Tesla V100-SXM2-32GB; Bus id: 0000:89:00
GPU id:7; name: Tesla V100-SXM2-32GB; Bus id: 0000:8a:00
selecting device 0
device ptr: 0x7f1563a00000
allocated size: 65536
Size(B) pin.Time(us) map.Time(us) get_info.Time(us) unmap.Time(us)
unpin.Time(us)
65536 1346.034060 3.603800 0.340270 4.700930 676.612800
Histogram of gdr_pin_buffer latency for 65536 bytes
[1303.852000 - 2607.704000] 93
[2607.704000 - 3911.556000] 0
[3911.556000 - 5215.408000] 0
[5215.408000 - 6519.260000] 0
[6519.260000 - 7823.112000] 0
[7823.112000 - 9126.964000] 0
[9126.964000 - 10430.816000] 0
[10430.816000 - 11734.668000] 0
[11734.668000 - 13038.520000] 0
[13038.520000 - 14342.372000] 2
closing gdrdrv
$ numactl -N 1 -l gdrcopy_pplat
GPU id:0; name: NVIDIA A40; Bus id: 0000:09:00
selecting device 0
device ptr: 0x7f99d2600000
gpu alloc fn: cuMemAlloc
map_d_ptr: 0x7f9a054fb000
info.va: 7f99d2600000
info.mapped_size: 4
info.page_size: 65536
info.mapped: 1
info.wc_mapping: 1
page offset: 0
user-space pointer: 0x7f9a054fb000
CPU does gdr_copy_to_mapping and GPU writes back via cuMemHostAlloc'd buffer.
Running 1000 iterations with data size 4 bytes.
Round-trip latency per iteration is 1.08762 us
unmapping buffer
unpinning buffer
closing gdrdrv
NUMA effects
Depending on the platform architecture, like where the GPU are placed in the PCIe topology, performance may suffer if the processor which is driving the copy is not the one which is hosting the GPU, for example in a multi-socket server.
In the example below, GPU ID 0 is hosted by CPU socket 0. By explicitly playing with the OS process and memory affinity, it is possible to run the test onto the optimal processor:
$ numactl -N 0 -l gdrcopy_copybw -d 0 -s $((64 * 1024)) -o $((0 * 1024)) -c $((64 * 1024))
GPU id:0; name: Tesla V100-SXM2-32GB; Bus id: 0000:06:00
GPU id:1; name: Tesla V100-SXM2-32GB; Bus id: 0000:07:00
GPU id:2; name: Tesla V100-SXM2-32GB; Bus id: 0000:0a:00
GPU id:3; name: Tesla V100-SXM2-32GB; Bus id: 0000:0b:00
GPU id:4; name: Tesla V100-SXM2-32GB; Bus id: 0000:85:00
GPU id:5; name: Tesla V100-SXM2-32GB; Bus id: 0000:86:00
GPU id:6; name: Tesla V100-SXM2-32GB; Bus id: 0000:89:00
GPU id:7; name: Tesla V100-SXM2-32GB; Bus id: 0000:8a:00
selecting device 0
testing size: 65536
rounded size: 65536
gpu alloc fn: cuMemAlloc
device ptr: 7f5817a00000
map_d_ptr: 0x7f583b186000
info.va: 7f5817a00000
info.mapped_size: 65536
info.page_size: 65536
info.mapped: 1
info.wc_mapping: 1
page offset: 0
user-space pointer:0x7f583b186000
writing test, size=65536 offset=0 num_iters=1000
write BW: 9768.3MB/s
reading test, size=65536 offset=0 num_iters=1000
read BW: 548.423MB/s
unmapping buffer
unpinning buffer
closing gdrdrv
or on the other socket:
$ numactl -N 1 -l gdrcopy_copybw -d 0 -s $((64 * 1024)) -o $((0 * 1024)) -c $((64 * 1024))
GPU id:0; name: Tesla V100-SXM2-32GB; Bus id: 0000:06:00
GPU id:1; name: Tesla V100-SXM2-32GB; Bus id: 0000:07:00
GPU id:2; name: Tesla V100-SXM2-32GB; Bus id: 0000:0a:00
GPU id:3; name: Tesla V100-SXM2-32GB; Bus id: 0000:0b:00
GPU id:4; name: Tesla V100-SXM2-32GB; Bus id: 0000:85:00
GPU id:5; name: Tesla V100-SXM2-32GB; Bus id: 0000:86:00
GPU id:6; name: Tesla V100-SXM2-32GB; Bus id: 0000:89:00
GPU id:7; name: Tesla V100-SXM2-32GB; Bus id: 0000:8a:00
selecting device 0
testing size: 65536
rounded size: 65536
gpu alloc fn: cuMemAlloc
device ptr: 7fbb63a00000
map_d_ptr: 0x7fbb82ab0000
info.va: 7fbb63a00000
info.mapped_size: 65536
info.page_size: 65536
info.mapped: 1
info.wc_mapping: 1
page offset: 0
user-space pointer:0x7fbb82ab0000
writing test, size=65536 offset=0 num_iters=1000
write BW: 9224.36MB/s
reading test, size=65536 offset=0 num_iters=1000
read BW: 521.262MB/s
unmapping buffer
unpinning buffer
closing gdrdrv
Restrictions and known issues
GDRCopy works with regular CUDA device memory only, as returned by cudaMalloc. In particular, it does not work with CUDA managed memory.
gdr_pin_buffer() accepts any addresses returned by cudaMalloc and its family. In contrast, gdr_map() requires that the pinned address is aligned to the GPU page. Neither CUDA Runtime nor Driver APIs guarantees that GPU memory allocation functions return aligned addresses. Users are responsible for proper alignment of addresses passed to the library.
Two cudaMalloc’d memory regions may be contiguous. Users may call gdr_pin_buffer and gdr_map with address and size that extend across these two regions. This use case is not well-supported in GDRCopy. On rare occassions, users may experience 1.) an error in gdr_map, or 2.) low copy performance because gdr_map cannot provide write-combined mapping.
In some GPU driver versions, pinning the same GPU address multiple times consumes additional BAR1 space. This is because the space is not properly reused. If you encounter this issue, we suggest that you try the latest version of NVIDIA GPU driver.
On POWER9 where CPU and GPU are connected via NVLink, CUDA9.2 and GPU Driver v396.37 are the minimum requirements in order to achieve the full performance. GDRCopy works with ealier CUDA and GPU driver versions but the achievable bandwidth is substantially lower.
If gdrdrv is compiled with the proprietary flavor of NVIDIA driver, GDRCopy does not fully support Linux with the confidential computing (CC) configuration with Intel CPU. In particular, it does not functional if CONFIG_ARCH_HAS_CC_PLATFORM=y and CC is enabled at runtime. However, it works if CC is disabled or CONFIG_ARCH_HAS_CC_PLATFORM=n. This issue is not applied to AMD CPU. To avoid this issue, please compile and load gdrdrv with the opensource flavor of NVIDIA driver.
To allow the loading of unsupported 3rd party modules in SLE, set allow_unsupported_modules 1 in /etc/modprobe.d/unsupported-modules. After making this change, modules missing the “supported” flag, will be allowed to load.
Bug filing
For reporting issues you may be having using any of NVIDIA software or reporting suspected bugs we would recommend you use the bug filing system which is available to NVIDIA registered developers on the developer site.
If you are not a member you can sign up.
Once a member you can submit issues using this form. Be sure to select GPUDirect in the “Relevant Area” field.
You can later track their progress using the My Bugs link on the left of this view.
Acknowledgment
If you find this software useful in your work, please cite: R. Shi et al., “Designing efficient small message transfer mechanism for inter-node MPI communication on InfiniBand GPU clusters,” 2014 21st International Conference on High Performance Computing (HiPC), Dona Paula, 2014, pp. 1-10, doi: 10.1109/HiPC.2014.7116873.