A low-latency GPU memory copy library based on NVIDIA GPUDirect RDMA technology.
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.
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:
- sanity, which contains unit tests for the library and the driver.
- copybw, a minimal application which calculates the R/W bandwidth for a specific buffer size.
- copylat, a benchmark application which calculates the R/W copy latency for a range of buffer sizes.
GPUDirect RDMA requires NVIDIA Tesla or Quadro class GPUs based on Kepler, Pascal, Volta, or Turing, see GPUDirect RDMA. For more technical informations, 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. Additionally, the sanity test requires check >= 0.9.8 and subunit.
# On RHEL
$ sudo yum install check check-devel subunit subunit-devel
# On Debian
$ sudo apt install check libsubunit0 libsubunit-devDeveloped and tested on RH7.x and Ubuntu18_04. The supported architectures are Linux x86_64 and ppc64le.
Root privileges are necessary to load/install the kernel-mode device driver.
We provide three ways for building and installing GDRCopy
$ sudo yum groupinstall 'Development Tools'
$ sudo yum install rpm-build make check check-devel subunit subunit-devel
$ cd packages
$ CUDA=<cuda-install-top-dir> ./build-rpm-packages.sh
$ sudo rpm -Uvh gdrcopy-kmod-<version>.<platform>.rpm
$ sudo rpm -Uvh gdrcopy-<version>.<platform>.rpm
$ sudo rpm -Uvh gdrcopy-devel-<version>.<platform>.rpm$ sudo apt install build-essential devscripts debhelper check libsubunit-dev
$ cd packages
$ CUDA=<cuda-install-top-dir> ./build-deb-packages.sh
$ sudo dpkg -i gdrdrv-dkms_<version>_<platform>.deb
$ sudo dpkg -i gdrcopy_<version>_<platform>.deb$ make PREFIX=<install-to-this-location> CUDA=<cuda-install-top-dir> all install
$ sudo ./insmod.shIf libcheck is installed in a non-standard path and therefore is not
picked by pkg-config, you can set the PKG_CONFIG_PATH environment
variable to the directory which contains the check.pc file and pass it
down to make:
$ PKG_CONFIG_PATH=/check_install_path/lib/pkgconfig/ make <...>Execute provided tests:
$ sanity
Running suite(s): Sanity
100%: Checks: 11, Failures: 0, Errors: 0
$ copybw
testing size: 4096
rounded size: 65536
device ptr: 5046c0000
bar_ptr: 0x7f8cff410000
info.va: 5046c0000
info.mapped_size: 65536
info.page_size: 65536
page offset: 0
user-space pointer:0x7f8cff410000
BAR writing test...
BAR1 write BW: 9549.25MB/s
BAR reading test...
BAR1 read BW: 1.50172MB/s
unmapping buffer
unpinning buffer
closing gdrdrv
$ copylat
GPU id:0; name: Tesla P100-PCIE-16GB; Bus id: 0000:09:00
selecting device 0
device ptr: 0x7f6aca800000
allocated size: 16777216
map_d_ptr: 0x7f6ae5000000
info.va: 7f6aca800000
info.mapped_size: 16777216
info.page_size: 65536
info.mapped: 1
info.wc_mapping: 1
page offset: 0
user-space pointer: 0x7f6ae5000000
gdr_copy_to_mapping num iters for each size: 10000
WARNING: Measuring the issue 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 0.0969
gdr_copy_to_mapping 2 0.0988
gdr_copy_to_mapping 4 0.0982
gdr_copy_to_mapping 8 0.0983
gdr_copy_to_mapping 16 0.1000
gdr_copy_to_mapping 32 0.0997
gdr_copy_to_mapping 64 0.1018
gdr_copy_to_mapping 128 0.1011
gdr_copy_to_mapping 256 0.1134
gdr_copy_to_mapping 512 0.1342
gdr_copy_to_mapping 1024 0.1751
gdr_copy_to_mapping 2048 0.2606
gdr_copy_to_mapping 4096 0.4336
gdr_copy_to_mapping 8192 0.8141
gdr_copy_to_mapping 16384 1.6070
gdr_copy_to_mapping 32768 3.1999
gdr_copy_to_mapping 65536 6.3869
gdr_copy_to_mapping 131072 12.7635
gdr_copy_to_mapping 262144 25.5032
gdr_copy_to_mapping 524288 51.0073
gdr_copy_to_mapping 1048576 102.0074
gdr_copy_to_mapping 2097152 203.9973
gdr_copy_to_mapping 4194304 408.1637
gdr_copy_to_mapping 8388608 817.4134
gdr_copy_to_mapping 16777216 1634.5638
gdr_copy_from_mapping num iters for each size: 100
Test Size(B) Avg.Time(us)
gdr_copy_from_mapping 1 1.0986
gdr_copy_from_mapping 2 1.9074
gdr_copy_from_mapping 4 1.7588
gdr_copy_from_mapping 8 1.7593
gdr_copy_from_mapping 16 0.8822
gdr_copy_from_mapping 32 1.7350
gdr_copy_from_mapping 64 3.0681
gdr_copy_from_mapping 128 3.4641
gdr_copy_from_mapping 256 2.9769
gdr_copy_from_mapping 512 3.5207
gdr_copy_from_mapping 1024 3.6279
gdr_copy_from_mapping 2048 5.5507
gdr_copy_from_mapping 4096 10.5047
gdr_copy_from_mapping 8192 17.8014
gdr_copy_from_mapping 16384 30.0232
gdr_copy_from_mapping 32768 58.1767
gdr_copy_from_mapping 65536 118.7792
gdr_copy_from_mapping 131072 241.5278
gdr_copy_from_mapping 262144 506.1804
gdr_copy_from_mapping 524288 1014.1972
gdr_copy_from_mapping 1048576 2026.6072
gdr_copy_from_mapping 2097152 4048.9970
gdr_copy_from_mapping 4194304 8103.9561
gdr_copy_from_mapping 8388608 19230.3878
gdr_copy_from_mapping 16777216 38474.8613
unmapping buffer
unpinning buffer
closing gdrdrvDepending 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, the K40m and K80 GPU are respectively hosted by socket0 and socket1. By explicitly playing with the OS process and memory affinity, it is possible to run the test onto the optimal processor:
$ GDRCOPY_ENABLE_LOGGING=1 GDRCOPY_LOG_LEVEL=0 LD_LIBRARY_PATH=$PWD:$LD_LIBRARY_PATH numactl -N 0 -l copybw -d 0 -s $((64 * 1024)) -o $((0 * 1024)) -c $((64 * 1024))
GPU id:0 name:Tesla K40m PCI domain: 0 bus: 2 device: 0
GPU id:1 name:Tesla K80 PCI domain: 0 bus: 132 device: 0
GPU id:2 name:Tesla K80 PCI domain: 0 bus: 133 device: 0
selecting device 0
testing size: 65536
rounded size: 65536
device ptr: 2305ba0000
bar_ptr: 0x7fe60956c000
info.va: 2305ba0000
info.mapped_size: 65536
info.page_size: 65536
page offset: 0
user-space pointer:0x7fe60956c000
BAR writing test, size=65536 offset=0 num_iters=10000
DBG: sse4_1=1 avx=1 sse=1 sse2=1
DBG: using AVX implementation of gdr_copy_to_bar
BAR1 write BW: 9793.23MB/s
BAR reading test, size=65536 offset=0 num_iters=100
DBG: using SSE4_1 implementation of gdr_copy_from_bar
BAR1 read BW: 787.957MB/s
unmapping buffer
unpinning buffer
closing gdrdrvor on the other one:
drossetti@drossetti-hsw0 16:52 (1181) gdrcopy>GDRCOPY_ENABLE_LOGGING=1 GDRCOPY_LOG_LEVEL=0 LD_LIBRARY_PATH=$PWD:$LD_LIBRARY_PATH numactl -N 1 -l copybw -d 0 -s $((64 * 1024)) -o $((0 * 1024)) -c $((64 * 1024))
GPU id:0 name:Tesla K40m PCI domain: 0 bus: 2 device: 0
GPU id:1 name:Tesla K80 PCI domain: 0 bus: 132 device: 0
GPU id:2 name:Tesla K80 PCI domain: 0 bus: 133 device: 0
selecting device 0
testing size: 65536
rounded size: 65536
device ptr: 2305ba0000
bar_ptr: 0x7f2299166000
info.va: 2305ba0000
info.mapped_size: 65536
info.page_size: 65536
page offset: 0
user-space pointer:0x7f2299166000
BAR writing test, size=65536 offset=0 num_iters=10000
DBG: sse4_1=1 avx=1 sse=1 sse2=1
DBG: using AVX implementation of gdr_copy_to_bar
BAR1 write BW: 6812.08MB/s
BAR reading test, size=65536 offset=0 num_iters=100
DBG: using SSE4_1 implementation of gdr_copy_from_bar
BAR1 read BW: 669.825MB/s
unmapping buffer
unpinning buffer
closing gdrdrvGDRCopy 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.
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.
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.
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.