CPU Affinity for PyG Workloads
The performance of PyG workloads using CPU can be significantly improved by setting a proper affinity mask. Processor affinity, or core binding, is a modification of the native OS queue scheduling algorithm that enables an application to assign a specific set of cores to processes or threads launched during its execution on the CPU. In consequence, it increases the overall effective hardware utilisation by minimizing core stalls and memory bounds. It also secures CPU resources to critical processes or threads, even if the system is under heavy load.
CPU affinity targets the two main performance-critical regions:
Execution bind: Indicates a core where process/thread will run.
Memory bind: Indicates a preferred memory area where memory pages will be bound (local areas in NUMA machine).
The following article discusses readily available tools and environment settings that one can use to maximize the performance of Intel CPUs with PyG.
Overall, CPU affinity can be a useful tool for improving the performance and predictability of certain types of applications, but one configuration does not necessarily fit all cases: it is important to carefully consider whether CPU affinity is appropriate for your use case, and to test and measure the impact of any changes you make.
Using CPU affinity
Each PyG workload can be parallelized using the PyTorch iterator class
MultiProcessingDataLoaderIter, which is automatically enabled in case
num_workers > 0 is passed to a
Under the hood, it creates
num_workers many sub-processes that will run in parallel to the main process.
Setting a CPU affinity mask for the data loading processes places
DataLoader worker threads on specific CPU cores.
In effect, it allows for more efficient data batch preparation by allocating pre-fetched batches in local memory.
Every time a process or thread moves from one core to another, registers and caches need to be flushed and reloaded.
This can become very costly if it happens often, and threads may also no longer be close to their data, or be able to share data in a cache.
Since PyG (2.3 and beyond),
LinkLoader classes officially support a native solution for CPU affinity using the
torch_geometric.loader.AffinityMixin context manager.
CPU affinity can be enabled via the
enable_cpu_affinity() method for
num_workers > 0 use-cases,
and will guarantee that a separate core is assigned to each worker at initialization.
A user-defined list of core IDs may be assigned using the
Otherwise, cores will be assigned automatically, starting at core ID 0.
As of now, only a single core can be assigned to a worker, hence multi-threading is disabled in workers’ processes by default.
The recommended number of workers to start with lies between
[2, 4], and the optimum may vary based on workload characteristics:
loader = NeigborLoader( data, num_workers=3, ..., ) with loader.enable_cpu_affinity(loader_cores=[0, 1, 2]): for batch in loader: pass
It is generally adivisable to use
filter_per_worker=True for any multi-process CPU workloads (
True by default).
The workers then prepare each mini-batch: first by sampling the node indices using pre-defined a sampler, and secondly filtering node and edge features according to sampled nodes and edges.
The filtering function selects node feature vectors from the complete input
Data tensor loaded into DRAM.
filter_per_worker is set to
True, each worker’s subprocess performs the filtering within it’s CPU resource.
Hence, main process resources are relieved and can be secured only for GNN computation.
Binding processes to physical cores
Following general performance tuning principles, it is advisable to use only physical cores for deep learning workloads.
For example, while two logical threads run
GEMM at the same time, they will be sharing the same core resources causing front end bound, such that the overhead from this front end bound is greater than the gain from running both logical threads at the same time.
This is because OpenMP threads will contend for the same
GEMM execution units, see here.
The binding can be done in many ways, however the most common tools are:
numactl(only on Linux):
--physcpubind=<cpus>, -C <cpus> or --cpunodebind=<nodes>, -N <nodes>
For best performance, it is required combine main process affinity using the tools listed above, with the multi-process
DataLoader affinity settings.
In each parallelized PyG workload execution, the main process performs message passing updates over GNN layers, while the
DataLoader workers sub-processes take care of fetching and pre-processing data to be passed to a GNN model.
It is advisable to isolate the CPU resources made available to these two processes to achieve the best results.
To do this, CPUs assigned to each affinity mask should be mutually exclusive.
For example, if four
DataLoader workers are assigned to CPUs
[0, 1, 2, 3], the main process should use the rest of available cores, i.e. by calling:
numactl -C 4-(N-1) --localalloc python …
N is the total number of physical cores, with the last CPU having core ID
--localalloc improves local memory allocation and keeps the cache closer to active cores.
Dual socket CPU separation
With dual-socket CPUs, it might be beneficial to further isolate the processes between the sockets.
This leads to decreased frequency of remote memory calls for the main process.
The goal is to utilize high-speed cache on local memory and reduces memory bound caused by migrating cached data between NUMA nodes.
This can be achieved by using
DataLoader affinity, and launching main process on the cores of the second socket, i.e. with:
numactl -C M-(N-1) -m 1 python …
M is the
cpuid of the first core of the second CPU socket.
Adding a complementary memory-allocation flag
-m 1 prioritizes cache allocation on the same NUMA node, where the main process is running (alternatively for less strict memory allocation use
This makes the data readily available on the same socket where the computation takes place.
Using this setting is very workload-specific and may require some fine-tuning, as one needs to manage a trade-off between using more OMP threads vs. limiting the number of remote memory calls.
Improving memory bounds
Following the CPU performance optimization guidelines for PyTorch, it is also advised for PyG to use
These generally can reach better memory usage than the default PyTorch memory allocator
A non-default memory allocator can be specified using
LD_PRELOAD prior to script execution.
Quick start guidelines
The general guidelines for achieving the best performance with CPU affinity can be summarized in the following steps:
Test if your dataset benefits from using parallel data loaders. For some datasets, it might be more beneficial to use a plain serial data loader, especially when the dimensions of the input
Dataare relatively small.
Enable multi-process data loaders by setting
num_workers > 0. A good estimate for
num_workerslies in the range
[2, 4]. However, for more complex datasets you might want to experiment with larger number of workers. Use the
enable_cpu_affinity()feature to affinitize
Bind execution to physical cores. Alternatively, hyperthreading can be disabled completely at a system-level.
Separate the cores used for main process from the data loader workers’ cores by using
Find the optimum number of OMP threads for your workload. A good starting point is
N - num_workers. Generally, well-parallelized models will benefit from many OMP threads. However, if your model computation flow has interlaced parallel and serial regions, the performance will decrease due to resource allocation needed for spawning and maintaining threads between parallel regions.
When using a dual-socket CPU, you might want to experiment with assigning data loading to one socket and main process to another socket with memory allocation (
numactl -m) on the same socket where the main process is executed. This leads to best cache-allocation and often overweighs the benefit of using more OMP threads.
An additional boost in performance can be obtained by using non-default memory allocator, such as
Finding an optimal setup for the CPU affinity mask is a problem of managing the proportion of CPU time spent in each iteration for loading and preparing the data vs. time spent during GNN execution. Different results may be obtained by changing model hyperparameters, such as the batch size, number of sampled neighbors, and the number of layers. As a general rule, workloads which require sampling a complex graph may benefit more from reserving some CPU resources just for the data preparation step.
The figure below presents the outcome of applying CPU affinity mask to
Measurements were taken for a variable number of workers, while other hyperparameters for each benchmark were constant:
--warmup 0 --use-sparse-tensor --num-layers 3 --num-hidden-channels 128 --batch-sizes 2048.
Three different affinity configurations are presented:
Baseline - only
OMP_NUM_THREADS=(N-num_workers) python training_benchmark.py --num-workers …
Aff - data loader process on first socket, main process on first and second socket, 98-110 threads:
LD_PRELOAD=(path)/libjemalloc.so (path)/libiomp5.so MALLOC_CONF=oversize_threshold:1,background_thread:true,metadata_thp:auto OMP_NUM_THREADS=(N-num_workers) KMP_AFFINITY=granularity=fine,compact,1,0 KMP_BLOCKTIME=0 numactl -C <num_workers-(N-1)> --localalloc python training_benchmark.py --cpu-affinity --num-workers …
Aff+SocketSep - data loader process on first socket, main process on second socket, 60 threads:
LD_PRELOAD=(path)/libjemalloc.so (path)/libiomp5.so MALLOC_CONF=oversize_threshold:1,background_thread:true,metadata_thp:auto OMP_NUM_THREADS=(N-M) KMP_AFFINITY=granularity=fine,compact,1,0 KMP_BLOCKTIME=0 numactl -C <M-(N-1)> -m 1 python training_benchmark.py --cpu-affinity --num-workers ...
Training times for each model/dataset combination were obtained by taking a mean of results at a variable number of dataloader workers:
[0, 2, 4, 8, 16] for the baseline and
[2, 4, 8, 16] workers for each affinity configuration.
Then, the affinity means were normalized with respect to the mean baseline measurement.
This value is denoted on the \(y\)-axis.
The labels above each result indicate the end-to-end performance gain from using the discussed configuration.
Over all model/dataset samples, the average training time is decreased by 1.53x for plain affinity and 1.85x for the affinity with socket separation.