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 usually consists of 1000’s of multi-processors, high-speed reminiscence, devoted tensor cores, and extra, is especially well-suited to fulfill the intensive calls for of AI/ML workloads. Sadly, the speedy progress in AI growth has led to a surge within the demand for GPUs, making them tough to acquire. Because of this, ML builders are more and more exploring different {hardware} choices for coaching and working their fashions. In earlier posts, we mentioned the potential of coaching on devoted AI ASICs akin to Google Cloud TPU, Haban Gaudi, and AWS Trainium. Whereas these choices supply vital 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 submit we return to the great old school 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 flexibility to run (at the least a few of) our workloads on CPU might have vital implications on growth 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 way of accelerating growth and minimizing prices. Whereas that is essential whatever the compute engine used, the profiling instruments and optimization strategies can fluctuate drastically between platforms. On this submit, we’ll talk about a few of the efficiency optimization choices that pertain to CPU. Our focus shall be on Intel® Xeon® CPU processors (with Intel® AVX-512) and on the PyTorch (model 2.4) framework (though comparable strategies could be utilized to different CPUs and frameworks, as nicely). Extra particularly, we’ll 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 software or library we should always point out, as an endorsement over their options.
Our purpose shall be to display that though ML growth on CPU will not be our first selection, there are methods to “soften the blow” and — in some instances — maybe even make it a viable different.
Disclaimers
Our intention on this submit is to display just some 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’ll concentrate on a coaching workload fairly than an inference workload. There are a selection of optimization instruments centered particularly on inference that we are going to not cowl (e.g., see right here and right here).
Please don’t view this submit as a substitute of the official documentation on any of the instruments or strategies that we point out. Needless to say given the speedy tempo of AI/ML growth, a few of the content material, libraries, and/or directions that we point out could change into outdated by the point you learn this. Please make sure to check with the most-up-to-date documentation accessible.
Importantly, the affect of the optimizations that we talk about on runtime efficiency is prone to fluctuate drastically based mostly on the mannequin and the main points of the surroundings (e.g., see the excessive diploma of variance between fashions on the official PyTorch TouchInductor CPU Inference Efficiency Dashboard). The comparative efficiency numbers we’ll share are particular to the toy mannequin and runtime surroundings that we are going to use. Be sure you reevaluate the entire proposed optimizations by yourself mannequin and runtime surroundings.
Lastly, our focus shall be solely on throughput efficiency (as measured in samples per second) — not on coaching convergence. Nonetheless, it must be famous that some optimization strategies (e.g., batch measurement tuning, blended precision, and extra) might have a destructive impact on the convergence of sure fashions. In some instances, this may be overcome via applicable hyperparameter tuning.
We’ll run our experiments on a easy picture classification mannequin with a ResNet-50 spine (from Deep Residual Studying for Picture Recognition). We’ll prepare the mannequin on a faux dataset. The total coaching script seems within the code block beneath (loosely based mostly on this instance):
import torch
import torchvision
from torch.utils.knowledge import Dataset, DataLoader
import time# A dataset with random pictures 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(knowledge=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
depend = 0
for idx, (knowledge, goal) in enumerate(train_loader):
optimizer.zero_grad()
output = mannequin(knowledge)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
batch_time = time.perf_counter() - t0
if idx > 10: # skip first steps
summ += batch_time
depend += 1
t0 = time.perf_counter()
if idx > 100:
break
print(f'common step time: {summ/depend}')
print(f'throughput: {depend*batch_size/summ}')
Working this script on a c7i.2xlarge (with 8 vCPUs) and the CPU model of PyTorch 2.4, ends in a throughput of 9.12 samples per second. For the sake of comparability, we notice 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 varieties ($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 present roughly eleven(!!) occasions higher value efficiency. Based mostly on these outcomes, the choice for utilizing GPUs to coach ML fashions could be very nicely based. Let’s assess a few of the prospects for lowering this hole.
On this part we’ll discover some primary strategies for rising the runtime efficiency of our coaching workload. Though you might acknowledge a few of these from our submit on GPU optimization, you will need to spotlight a major 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 knowledge preprocessing on) the CPU and (the mannequin coaching on) the GPU. On CPU platforms the entire processing happens on the CPU and our purpose shall be to allocate its sources most successfully.
Batch Dimension
Growing the coaching batch measurement can probably improve efficiency by lowering the frequency of the mannequin parameter updates. (On GPUs it has the additional advantage of lowering the overhead of CPU-GPU transactions akin to kernel loading). Nonetheless, whereas on GPU we aimed for a batch measurement that will maximize the utilization of the GPU reminiscence, the identical technique would possibly harm efficiency on CPU. For causes past the scope of this submit, CPU reminiscence is extra sophisticated and the most effective method for locating probably the most optimum batch measurement could also be via trial and error. Needless to say altering the batch measurement might have an effect on coaching convergence.
The desk beneath summarizes the throughput of our coaching workload for a couple of (arbitrary) decisions of batch measurement:
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 standard method on GPUs is to assign a number of processes to the information loader in order to scale back the probability of hunger of the GPU. On GPU platforms, a common rule of thumb is to set the variety of staff based on the variety of CPU cores. Nonetheless, 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, the most effective method for selecting the optimum variety of staff could also be trial and error. The desk beneath exhibits the common throughput for various decisions of num_workers:
Combined Precision
One other in style method is to make use of decrease precision floating level datatypes akin to torch.float16 or torch.bfloat16 with the dynamic vary of torch.bfloat16 usually thought-about to be extra amiable to ML coaching. Naturally, lowering the datatype precision can have hostile results on convergence and must be completed fastidiously. PyTorch comes with torch.amp, an automated blended precision package deal for optimizing the usage of these datatypes. Intel® AVX-512 consists of assist for the bfloat16 datatype. The modified coaching step seems beneath:
for idx, (knowledge, goal) in enumerate(train_loader):
optimizer.zero_grad()
with torch.amp.autocast('cpu',dtype=torch.bfloat16):
output = mannequin(knowledge)
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 ends in the entire knowledge 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, (knowledge, goal) in enumerate(train_loader):
knowledge = knowledge.to(memory_format=torch.channels_last)
optimizer.zero_grad()
with torch.amp.autocast('cpu',dtype=torch.bfloat16):
output = mannequin(knowledge)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
The ensuing throughput is 37.93 samples per second — a further 56% enchancment and a complete of 415% in comparison with our baseline experiment. We’re on a task!!
Torch Compilation
In a earlier submit we coated the virtues of PyTorch’s assist for graph compilation and its potential affect on runtime efficiency. Opposite to the default keen execution mode wherein 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 totally different backend libraries and with a number of configuration choices. Right here we’ll 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 affect of torch compilation is barely obvious when the “channels final” optimization is disabled (and improve of ~27% for every of the backends). When “channels final” is utilized, the efficiency truly drops. Because of this, we drop this optimization from our subsequent experiments.
There are a selection of alternatives for optimizing the usage of the underlying CPU sources. These embody optimizing reminiscence administration and thread allocation to the construction of the underlying CPU {hardware}. Reminiscence administration could be improved via the usage of superior reminiscence allocators (akin to Jemalloc and TCMalloc) and/or lowering reminiscence accesses which can be slower (i.e., throughout NUMA nodes). Threading allocation could be improved via 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 stage understanding of the CPU structure and the options of its supporting SW stack. To simplify issues, PyTorch provides the torch.backends.xeon.run_cpu script for mechanically configuring the reminiscence and threading libraries in order to optimize runtime efficiency. The command beneath will end in the usage of the devoted reminiscence and threading libraries. We’ll 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. Observe that the run_cpu script consists of many controls for additional tuning efficiency. Be sure you take a look at the documentation to be able to maximize its use.
The Intel® Extension for PyTorch consists of further alternatives for coaching optimization through its ipex.optimize perform. Right here we display its default use. Please see the documentation to study 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. (Observe {that a} comparable result’s reached when disabling the “channels final” configuration.)
Intel® Xeon® processors are designed with Non-Uniform Reminiscence Entry (NUMA) wherein 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 way quicker. This provides 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 knowledge distribution throughout nodes is managed by Intel® oneCCL, Intel’s devoted collective communications library.
We are able to run knowledge distributed coaching throughout NUMA nodes simply utilizing the ipexrun utility. Within the following code block (loosely based mostly on this instance) we adapt our script to run knowledge distributed coaching (based on utilization detailed right here):
import os, time
import torch
from torch.utils.knowledge import Dataset, DataLoader
from torch.utils.knowledge.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(knowledge=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, knowledge 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 based mostly coaching on XLA units akin to TPU, GPU, and CPU. Much like torch compilation, XLA makes use of graph compilation to generate machine code that’s optimized for the goal machine. With the institution of the OpenXLA Challenge, one of many acknowledged targets was to assist excessive efficiency throughout all {hardware} backends, together with CPU (see the CPU RFC right here). The code block beneath demonstrates the changes to our authentic (unoptimized) script required to coach utilizing PyTorch/XLA:
import torch
import torchvision
import timeimport torch_xla
import torch_xla.core.xla_model as xmmachine = xm.xla_device()
mannequin = torchvision.fashions.resnet50().to(machine)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(mannequin.parameters())
mannequin.prepare()
for idx, (knowledge, goal) in enumerate(train_loader):
knowledge = knowledge.to(machine)
goal = goal.to(machine)
optimizer.zero_grad()
output = mannequin(knowledge)
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 assist 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 submit. The samples per greenback was calculated based mostly 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. Based mostly on our outcomes, coaching on GPU can be ~6.7 occasions cheaper. It’s seemingly that with further efficiency tuning and/or making use of further 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 surroundings.
Amazon EC2 Spot Situations Reductions
The elevated availability of cloud-based CPU occasion varieties (in comparison with GPU occasion varieties) could suggest better alternative for acquiring compute energy at discounted charges, e.g., via Spot Occasion utilization. Amazon EC2 Spot Situations are situations from surplus cloud service capability which can be provided for a reduction of as a lot as 90% off the On-Demand pricing. In change for the discounted value, AWS maintains the proper to preempt the occasion with little to no warning. Given the excessive demand for GPUs, you might 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 consequence 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 surroundings) had been decrease than the GPU outcomes, it’s seemingly that the identical optimization steps utilized to different mannequin architectures (e.g., ones that embody elements that aren’t supported by GPU) could consequence within the CPU efficiency matching or beating that of the GPU. And even in instances the place the efficiency hole just isn’t bridged, there could very nicely be instances the place the scarcity of GPU compute capability would justify working 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 working ML workloads might have large implications on growth 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 strategies accessible for reinforcing its efficiency — a choose few of which we’ve got mentioned and demonstrated on this submit.
On this submit we centered optimizing coaching on CPU. Please make sure to take a look at our many different posts on medium masking all kinds of subjects pertaining to efficiency evaluation and optimization of machine studying workloads.
