Revisiting CPU for ML in an Period of GPU Shortage
The current successes in AI are sometimes attributed to the emergence and evolutions of the GPU. The GPU’s structure, which generally consists of hundreds of multi-processors, high-speed reminiscence, devoted tensor cores, and extra, is especially well-suited to satisfy the intensive calls for of AI/ML workloads. Sadly, the speedy progress in AI improvement has led to a surge within the demand for GPUs, making them troublesome to acquire. In consequence, ML builders are more and more exploring different {hardware} choices for coaching and operating their fashions. In earlier posts, we mentioned the potential for coaching on devoted AI ASICs equivalent to Google Cloud TPU, Haban Gaudi, and AWS Trainium. Whereas these choices provide important cost-saving alternatives, they don’t go well with all ML fashions and might, just like the GPU, additionally endure from restricted availability. On this put up we return to the great old style CPU and revisit its relevance to ML purposes. Though CPUs are usually much less suited to ML workloads in comparison with GPUs, they’re much simpler to accumulate. The power to run (at the very least a few of) our workloads on CPU might have important implications on improvement productiveness.
In earlier posts (e.g., right here) we emphasised the significance of analyzing and optimizing the runtime efficiency of AI/ML workloads as a method of accelerating improvement and minimizing prices. Whereas that is essential whatever the compute engine used, the profiling instruments and optimization methods can fluctuate drastically between platforms. On this put up, we are going to talk about a number of the efficiency optimization choices that pertain to CPU. Our focus will likely be on Intel® Xeon® CPU processors (with Intel® AVX-512) and on the PyTorch (model 2.4) framework (though related methods will be utilized to different CPUs and frameworks, as nicely). Extra particularly, we are going to run our experiments on an Amazon EC2 c7i occasion with an AWS Deep Studying AMI. Please don’t view our selection of Cloud platform, CPU model, ML framework, or some other device or library we must always point out, as an endorsement over their options.
Our purpose will likely be to display that though ML improvement on CPU will not be our first selection, there are methods to “soften the blow” and — in some circumstances — even perhaps make it a viable different.
Disclaimers
Our intention on this put up is to display only a few of the ML optimization alternatives accessible on CPU. Opposite to a lot of the on-line tutorials on the subject of ML optimization on CPU, we are going to give attention to a coaching workload reasonably than an inference workload. There are a selection of optimization instruments centered particularly on inference that we’ll not cowl (e.g., see right here and right here).
Please don’t view this put up as a substitute of the official documentation on any of the instruments or methods that we point out. Remember the fact that given the speedy tempo of AI/ML improvement, a number of the content material, libraries, and/or directions that we point out might develop into outdated by the point you learn this. Please be sure you check with the most-up-to-date documentation accessible.
Importantly, the impression of the optimizations that we talk about on runtime efficiency is prone to fluctuate drastically primarily based on the mannequin and the small print of the atmosphere (e.g., see the excessive diploma of variance between fashions on the official PyTorch TouchInductor CPU Inference Efficiency Dashboard). The comparative efficiency numbers we are going to share are particular to the toy mannequin and runtime atmosphere that we’ll use. You’ll want to reevaluate all the proposed optimizations by yourself mannequin and runtime atmosphere.
Lastly, our focus will likely be solely on throughput efficiency (as measured in samples per second) — not on coaching convergence. Nevertheless, it needs to be famous that some optimization methods (e.g., batch dimension tuning, combined precision, and extra) might have a detrimental impact on the convergence of sure fashions. In some circumstances, this may be overcome by way of applicable hyperparameter tuning.
We are going to run our experiments on a easy picture classification mannequin with a ResNet-50 spine (from Deep Residual Studying for Picture Recognition). We are going to prepare the mannequin on a pretend dataset. The complete coaching script seems within the code block beneath (loosely primarily based on this instance):
import torch
import torchvision
from torch.utils.information import Dataset, DataLoader
import time# A dataset with random photos and labels
class FakeDataset(Dataset):
def __len__(self):
return 1000000
def __getitem__(self, index):
rand_image = torch.randn([3, 224, 224], dtype=torch.float32)
label = torch.tensor(information=index % 10, dtype=torch.uint8)
return rand_image, label
train_set = FakeDataset()
batch_size=128
num_workers=0
train_loader = DataLoader(
dataset=train_set,
batch_size=batch_size,
num_workers=num_workers
)
mannequin = torchvision.fashions.resnet50()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(mannequin.parameters())
mannequin.prepare()
t0 = time.perf_counter()
summ = 0
rely = 0
for idx, (information, goal) in enumerate(train_loader):
optimizer.zero_grad()
output = mannequin(information)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
batch_time = time.perf_counter() - t0
if idx > 10: # skip first steps
summ += batch_time
rely += 1
t0 = time.perf_counter()
if idx > 100:
break
print(f'common step time: {summ/rely}')
print(f'throughput: {rely*batch_size/summ}')
Operating this script on a c7i.2xlarge (with 8 vCPUs) and the CPU model of PyTorch 2.4, leads to a throughput of 9.12 samples per second. For the sake of comparability, we be aware that the throughput of the identical (unoptimized script) on an Amazon EC2 g5.2xlarge occasion (with 1 GPU and eight vCPUs) is 340 samples per second. Bearing in mind the comparative prices of those two occasion sorts ($0.357 per hour for a c7i.2xlarge and $1.212 for a g5.2xlarge, as of the time of this writing), we discover that coaching on the GPU occasion to offer roughly eleven(!!) instances higher value efficiency. Primarily based on these outcomes, the desire for utilizing GPUs to coach ML fashions could be very nicely based. Let’s assess a number of the prospects for lowering this hole.
On this part we are going to discover some primary strategies for rising the runtime efficiency of our coaching workload. Though it’s possible you’ll acknowledge a few of these from our put up on GPU optimization, you will need to spotlight a big distinction between coaching optimization on CPU and GPU platforms. On GPU platforms a lot of our effort was devoted to maximizing the parallelization between (the coaching information preprocessing on) the CPU and (the mannequin coaching on) the GPU. On CPU platforms all the processing happens on the CPU and our purpose will likely be to allocate its sources most successfully.
Batch Measurement
Rising the coaching batch dimension can doubtlessly enhance efficiency by lowering the frequency of the mannequin parameter updates. (On GPUs it has the additional benefit of lowering the overhead of CPU-GPU transactions equivalent to kernel loading). Nevertheless, whereas on GPU we aimed for a batch dimension that might maximize the utilization of the GPU reminiscence, the identical technique may damage efficiency on CPU. For causes past the scope of this put up, CPU reminiscence is extra sophisticated and one of the best method for locating probably the most optimum batch dimension could also be by way of trial and error. Remember the fact that altering the batch dimension might have an effect on coaching convergence.
The desk beneath summarizes the throughput of our coaching workload for just a few (arbitrary) decisions of batch dimension:
Opposite to our findings on GPU, on the c7i.2xlarge occasion sort our mannequin seems to choose decrease batch sizes.
Multi-process Knowledge Loading
A typical method on GPUs is to assign a number of processes to the information loader in order to scale back the chance of hunger of the GPU. On GPU platforms, a normal rule of thumb is to set the variety of employees in accordance with the variety of CPU cores. Nevertheless, on CPU platforms, the place the mannequin coaching makes use of the identical sources as the information loader, this method might backfire. As soon as once more, one of the best method for selecting the optimum variety of employees could also be trial and error. The desk beneath exhibits the typical throughput for various decisions of num_workers:
Combined Precision
One other common method is to make use of decrease precision floating level datatypes equivalent to torch.float16 or torch.bfloat16 with the dynamic vary of torch.bfloat16 usually thought of to be extra amiable to ML coaching. Naturally, lowering the datatype precision can have adversarial results on convergence and needs to be carried out rigorously. PyTorch comes with torch.amp, an automated combined precision bundle for optimizing using these datatypes. Intel® AVX-512 consists of help for the bfloat16 datatype. The modified coaching step seems beneath:
for idx, (information, goal) in enumerate(train_loader):
optimizer.zero_grad()
with torch.amp.autocast('cpu',dtype=torch.bfloat16):
output = mannequin(information)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
The throughput following this optimization is 24.34 samples per second, a rise of 86%!!
Channels Final Reminiscence Format
Channels final reminiscence format is a beta-level optimization (on the time of this writing), pertaining primarily to imaginative and prescient fashions, that helps storing 4 dimensional (NCHW) tensors in reminiscence such that the channels are the final dimension. This leads to all the information of every pixel being saved collectively. This optimization pertains primarily to imaginative and prescient fashions. Thought of to be extra “pleasant to Intel platforms”, this reminiscence format is reported enhance the efficiency of a ResNet-50 on an Intel® Xeon® CPU. The adjusted coaching step seems beneath:
for idx, (information, goal) in enumerate(train_loader):
information = information.to(memory_format=torch.channels_last)
optimizer.zero_grad()
with torch.amp.autocast('cpu',dtype=torch.bfloat16):
output = mannequin(information)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
The ensuing throughput is 37.93 samples per second — an extra 56% enchancment and a complete of 415% in comparison with our baseline experiment. We’re on a task!!
Torch Compilation
In a earlier put up we coated the virtues of PyTorch’s help for graph compilation and its potential impression on runtime efficiency. Opposite to the default keen execution mode by which every operation is run independently (a.okay.a., “eagerly”), the compile API converts the mannequin into an intermediate computation graph which is then JIT-compiled into low-level machine code in a way that’s optimum for the underlying coaching engine. The API helps compilation through completely different backend libraries and with a number of configuration choices. Right here we are going to restrict our analysis to the default (TorchInductor) backend and the ipex backend from the Intel® Extension for PyTorch, a library with devoted optimizations for Intel {hardware}. Please see the documentation for applicable set up and utilization directions. The up to date mannequin definition seems beneath:
import intel_extension_for_pytorch as ipexmannequin = torchvision.fashions.resnet50()
backend='inductor' # optionally change to 'ipex'
mannequin = torch.compile(mannequin, backend=backend)
Within the case of our toy mannequin, the impression of torch compilation is simply obvious when the “channels final” optimization is disabled (and enhance of ~27% for every of the backends). When “channels final” is utilized, the efficiency really drops. In consequence, we drop this optimization from our subsequent experiments.
There are a selection of alternatives for optimizing using the underlying CPU sources. These embody optimizing reminiscence administration and thread allocation to the construction of the underlying CPU {hardware}. Reminiscence administration will be improved by way of using superior reminiscence allocators (equivalent to Jemalloc and TCMalloc) and/or lowering reminiscence accesses which are slower (i.e., throughout NUMA nodes). Threading allocation will be improved by way of applicable configuration of the OpenMP threading library and/or use of Intel’s Open MP library.
Typically talking, these sorts of optimizations require a deep degree understanding of the CPU structure and the options of its supporting SW stack. To simplify issues, PyTorch affords the torch.backends.xeon.run_cpu script for routinely configuring the reminiscence and threading libraries in order to optimize runtime efficiency. The command beneath will lead to using the devoted reminiscence and threading libraries. We are going to return to the subject of NUMA nodes once we talk about the choice of distributed coaching.
We confirm applicable set up of TCMalloc (conda set up conda-forge::gperftools) and Intel’s Open MP library (pip set up intel-openmp), and run the next command.
python -m torch.backends.xeon.run_cpu prepare.py
The usage of the run_cpu script additional boosts our runtime efficiency to 39.05 samples per second. Word that the run_cpu script consists of many controls for additional tuning efficiency. You’ll want to try the documentation to be able to maximize its use.
The Intel® Extension for PyTorch consists of extra alternatives for coaching optimization through its ipex.optimize operate. Right here we display its default use. Please see the documentation to be taught of its full capabilities.
mannequin = torchvision.fashions.resnet50()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(mannequin.parameters())
mannequin.prepare()
mannequin, optimizer = ipex.optimize(
mannequin,
optimizer=optimizer,
dtype=torch.bfloat16
)
Mixed with the reminiscence and thread optimizations mentioned above, the resultant throughput is 40.73 samples per second. (Word {that a} related result’s reached when disabling the “channels final” configuration.)
Intel® Xeon® processors are designed with Non-Uniform Reminiscence Entry (NUMA) by which the CPU reminiscence is split into teams, a.okay.a., NUMA nodes, and every of the CPU cores is assigned to 1 node. Though any CPU core can entry the reminiscence of any NUMA node, the entry to its personal node (i.e., its native reminiscence) is far quicker. This offers rise to the notion of distributing coaching throughout NUMA nodes, the place the CPU cores assigned to every NUMA node act as a single course of in a distributed course of group and information distribution throughout nodes is managed by Intel® oneCCL, Intel’s devoted collective communications library.
We will run information distributed coaching throughout NUMA nodes simply utilizing the ipexrun utility. Within the following code block (loosely primarily based on this instance) we adapt our script to run information distributed coaching (in accordance with utilization detailed right here):
import os, time
import torch
from torch.utils.information import Dataset, DataLoader
from torch.utils.information.distributed import DistributedSampler
import torch.distributed as dist
import torchvision
import oneccl_bindings_for_pytorch as torch_ccl
import intel_extension_for_pytorch as ipexos.environ["MASTER_ADDR"] = "127.0.0.1"
os.environ["MASTER_PORT"] = "29500"
os.environ["RANK"] = os.environ.get("PMI_RANK", "0")
os.environ["WORLD_SIZE"] = os.environ.get("PMI_SIZE", "1")
dist.init_process_group(backend="ccl", init_method="env://")
rank = os.environ["RANK"]
world_size = os.environ["WORLD_SIZE"]
batch_size = 128
num_workers = 0
# outline dataset and dataloader
class FakeDataset(Dataset):
def __len__(self):
return 1000000
def __getitem__(self, index):
rand_image = torch.randn([3, 224, 224], dtype=torch.float32)
label = torch.tensor(information=index % 10, dtype=torch.uint8)
return rand_image, label
train_dataset = FakeDataset()
dist_sampler = DistributedSampler(train_dataset)
train_loader = DataLoader(
dataset=train_dataset,
batch_size=batch_size,
num_workers=num_workers,
sampler=dist_sampler
)
# outline mannequin artifacts
mannequin = torchvision.fashions.resnet50()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(mannequin.parameters())
mannequin.prepare()
mannequin, optimizer = ipex.optimize(
mannequin,
optimizer=optimizer,
dtype=torch.bfloat16
)
# configure DDP
mannequin = torch.nn.parallel.DistributedDataParallel(mannequin)
# run coaching loop
# destroy the method group
dist.destroy_process_group()
Sadly, as of the time of this writing, the Amazon EC2 c7i occasion household doesn’t embody a multi-NUMA occasion sort. To check our distributed coaching script, we revert again to a Amazon EC2 c6i.32xlarge occasion with 64 vCPUs and a couple of NUMA nodes. We confirm the set up of Intel® oneCCL Bindings for PyTorch and run the next command (as documented right here):
supply $(python -c "import oneccl_bindings_for_pytorch as torch_ccl;print(torch_ccl.cwd)")/env/setvars.sh# This instance command would make the most of all of the numa sockets of the processor, taking every socket as a rank.
ipexrun cpu --nnodes 1 --omp_runtime intel prepare.py
The next desk compares the efficiency outcomes on the c6i.32xlarge occasion with and with out distributed coaching:
In our experiment, information distribution did not enhance the runtime efficiency. Please see ipexrun documentation for added efficiency tuning choices.
In earlier posts (e.g., right here) we mentioned the PyTorch/XLA library and its use of XLA compilation to allow PyTorch primarily based coaching on XLA units equivalent to TPU, GPU, and CPU. Just like torch compilation, XLA makes use of graph compilation to generate machine code that’s optimized for the goal gadget. With the institution of the OpenXLA Mission, one of many acknowledged targets was to help excessive efficiency throughout all {hardware} backends, together with CPU (see the CPU RFC right here). The code block beneath demonstrates the changes to our unique (unoptimized) script required to coach utilizing PyTorch/XLA:
import torch
import torchvision
import timeimport torch_xla
import torch_xla.core.xla_model as xmgadget = xm.xla_device()
mannequin = torchvision.fashions.resnet50().to(gadget)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(mannequin.parameters())
mannequin.prepare()
for idx, (information, goal) in enumerate(train_loader):
information = information.to(gadget)
goal = goal.to(gadget)
optimizer.zero_grad()
output = mannequin(information)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
xm.mark_step()
Sadly, (as of the time of this writing) the XLA outcomes on our toy mannequin appear far inferior to the (unoptimized) outcomes we noticed above (— by as a lot as 7X). We anticipate this to enhance as PyTorch/XLA’s CPU help matures.
We summarize the outcomes of a subset of our experiments within the desk beneath. For the sake of comparability, we add the throughput of coaching our mannequin on Amazon EC2 g5.2xlarge GPU occasion following the optimization steps mentioned in this put up. The samples per greenback was calculated primarily based on the Amazon EC2 On-demand pricing web page ($0.357 per hour for a c7i.2xlarge and $1.212 for a g5.2xlarge, as of the time of this writing).
Though we succeeded in boosting the coaching efficiency of our toy mannequin on the CPU occasion by a substantial margin (446%), it stays inferior to the (optimized) efficiency on the GPU occasion. Primarily based on our outcomes, coaching on GPU can be ~6.7 instances cheaper. It’s probably that with extra efficiency tuning and/or making use of extra optimizations methods, we might additional shut the hole. As soon as once more, we emphasize that the comparative efficiency outcomes we’ve got reached are distinctive to this mannequin and runtime atmosphere.
Amazon EC2 Spot Situations Reductions
The elevated availability of cloud-based CPU occasion sorts (in comparison with GPU occasion sorts) might indicate higher alternative for acquiring compute energy at discounted charges, e.g., by way of Spot Occasion utilization. Amazon EC2 Spot Situations are situations from surplus cloud service capability which are provided for a reduction of as a lot as 90% off the On-Demand pricing. In trade for the discounted value, AWS maintains the appropriate to preempt the occasion with little to no warning. Given the excessive demand for GPUs, it’s possible you’ll discover CPU spot situations simpler to get ahold of than their GPU counterparts. On the time of this writing, c7i.2xlarge Spot Occasion value is $0.1291 which might enhance our samples per greenback outcome to 1135.76 and additional reduces the hole between the optimized GPU and CPU value performances (to 2.43X).
Whereas the runtime efficiency outcomes of the optimized CPU coaching of our toy mannequin (and our chosen atmosphere) had been decrease than the GPU outcomes, it’s probably that the identical optimization steps utilized to different mannequin architectures (e.g., ones that embody elements that aren’t supported by GPU) might outcome within the CPU efficiency matching or beating that of the GPU. And even in circumstances the place the efficiency hole will not be bridged, there might very nicely be circumstances the place the scarcity of GPU compute capability would justify operating a few of our ML workloads on CPU.
Given the ubiquity of the CPU, the power to make use of them successfully for coaching and/or operating ML workloads might have big implications on improvement productiveness and on end-product deployment technique. Whereas the character of the CPU structure is much less amiable to many ML purposes when in comparison with the GPU, there are lots of instruments and methods accessible for reinforcing its efficiency — a choose few of which we’ve got mentioned and demonstrated on this put up.
On this put up we centered optimizing coaching on CPU. Please be sure you try our many different posts on medium overlaying all kinds of matters pertaining to efficiency evaluation and optimization of machine studying workloads.
