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 contains 1000’s 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 development in AI improvement has led to a surge within the demand for GPUs, making them tough 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 swimsuit all ML fashions and might, just like the GPU, additionally undergo from restricted availability. On this put up we return to the great old school CPU and revisit its relevance to ML functions. Though CPUs are typically much less suited to ML workloads in comparison with GPUs, they’re much simpler to accumulate. The flexibility to run (a minimum of a few of) our workloads on CPU may have important implications on improvement productiveness.
In earlier posts (e.g., here) we emphasised the significance of analyzing and optimizing the runtime efficiency of AI/ML workloads as a way of accelerating improvement and minimizing prices. Whereas that is essential whatever the compute engine used, the profiling instruments and optimization methods can range significantly between platforms. On this put up, we’ll focus on a few of the efficiency optimization choices that pertain to CPU. Our focus will probably 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’ll run our experiments on an Amazon EC2 c7i occasion with an AWS Deep Learning AMI. Please don’t view our alternative of Cloud platform, CPU model, ML framework, or another software or library we should always point out, as an endorsement over their alternate options.
Our aim will probably be to exhibit that though ML improvement on CPU will not be our first alternative, there are methods to “soften the blow” and — in some instances — even perhaps make it a viable different.
Disclaimers
Our intention on this put up is to exhibit only a few of the ML optimization alternatives out there 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 quite than an inference workload. There are a selection of optimization instruments centered particularly on inference that we’ll not cowl (e.g., see here and here).
Please don’t view this put up as a alternative of the official documentation on any of the instruments or methods that we point out. Understand that given the speedy tempo of AI/ML improvement, a few of the content material, libraries, and/or directions that we point out might grow to be outdated by the point you learn this. Please you should definitely check with the most-up-to-date documentation out there.
Importantly, the influence of the optimizations that we focus on on runtime efficiency is prone to range significantly primarily based on the mannequin and the small print of the setting (e.g., see the excessive diploma of variance between fashions on the official PyTorch TouchInductor CPU Inference Performance Dashboard). The comparative efficiency numbers we’ll share are particular to the toy mannequin and runtime setting that we’ll use. You should definitely reevaluate the entire proposed optimizations by yourself mannequin and runtime setting.
Lastly, our focus will probably be solely on throughput efficiency (as measured in samples per second) — not on coaching convergence. Nonetheless, it needs to be famous that some optimization methods (e.g., batch measurement tuning, combined precision, and extra) may have a detrimental impact on the convergence of sure fashions. In some instances, this may be overcome by means of applicable hyperparameter tuning.
We’ll run our experiments on a easy picture classification mannequin with a ResNet-50 spine (from Deep Residual Learning for Image Recognition). We’ll prepare the mannequin on a faux dataset. The complete coaching script seems within the code block under (loosely primarily based on this example):
import torch
import torchvision
from torch.utils.knowledge 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(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 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. Considering the comparative costs 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 present roughly eleven(!!) occasions higher worth efficiency. Based mostly on these outcomes, the desire for utilizing GPUs to coach ML fashions could be very nicely based. Let’s assess a few of the potentialities for decreasing this hole.
On this part we’ll discover some primary strategies for growing the runtime efficiency of our coaching workload. Though you might acknowledge a few of these from our post on GPU optimization, it is very important 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 aim will probably be to allocate its sources most successfully.
Batch Measurement
Growing the coaching batch measurement can doubtlessly improve efficiency by decreasing the frequency of the mannequin parameter updates. (On GPUs it has the additional advantage of decreasing the overhead of CPU-GPU transactions equivalent 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 damage efficiency on CPU. For causes past the scope of this put up, CPU reminiscence is extra difficult and the perfect method for locating essentially the most optimum batch measurement could also be by means of trial and error. Understand that altering the batch measurement may have an effect on coaching convergence.
The desk under 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 desire decrease batch sizes.
Multi-process Knowledge Loading
A typical approach on GPUs is to assign multiple processes to the information loader in order to cut back the probability of hunger of the GPU. On GPU platforms, a common rule of thumb is to set the variety of employees in accordance with 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 may backfire. As soon as once more, the perfect method for selecting the optimum variety of employees could also be trial and error. The desk under reveals the typical throughput for various decisions of num_workers:
Blended Precision
One other in style approach is to make use of decrease precision floating level datatypes equivalent to torch.float16 or torch.bfloat16 with the dynamic vary of torch.bfloat16 typically thought-about to be extra amiable to ML coaching. Naturally, decreasing the datatype precision can have adversarial results on convergence and needs to be carried out rigorously. PyTorch comes with torch.amp, an computerized combined precision bundle for optimizing using these datatypes. Intel® AVX-512 contains support for the bfloat16 datatype. The modified coaching step seems under:
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 last memory 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. Considered to be more “friendly 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 under:
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 — an extra 56% enchancment and a complete of 415% in comparison with our baseline experiment. We’re on a job!!
Torch Compilation
In a previous post we lined the virtues of PyTorch’s help for graph compilation and its potential influence on runtime efficiency. Opposite to the default keen execution mode through which every operation is run independently (a.ok.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 fashion 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 under:
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 influence of torch compilation is just obvious when the “channels final” optimization is disabled (and improve 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 the use of the underlying CPU resources. These embody optimizing reminiscence administration and thread allocation to the structure of the underlying CPU {hardware}. Reminiscence administration will be improved by means of using advanced memory allocators (equivalent to Jemalloc and TCMalloc) and/or decreasing reminiscence accesses which can be slower (i.e., throughout NUMA nodes). Threading allocation will be improved by means 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 stage understanding of the CPU structure and the options of its supporting SW stack. To simplify issues, PyTorch gives the torch.backends.xeon.run_cpu script for robotically configuring the reminiscence and threading libraries in order to optimize runtime efficiency. The command under will lead to using the devoted reminiscence and threading libraries. We’ll return to the subject of NUMA nodes once we focus on 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 contains many controls for additional tuning efficiency. You should definitely take a look at the documentation with a view to maximize its use.
The Intel® Extension for PyTorch contains further alternatives for training optimization through its ipex.optimize perform. Right here we exhibit 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. (Word {that a} related result’s reached when disabling the “channels final” configuration.)
Intel® Xeon® processors are designed with Non-Uniform Memory Access (NUMA) through which the CPU reminiscence is split into teams, a.ok.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 training across NUMA nodes, the place the CPU cores assigned to every NUMA node act as a single course of in a distributed process group and data distribution throughout nodes is managed by Intel® oneCCL, Intel’s devoted collective communications library.
We will run knowledge distributed coaching throughout NUMA nodes simply utilizing the ipexrun utility. Within the following code block (loosely primarily based on this example) we adapt our script to run knowledge distributed coaching (in accordance with utilization detailed 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 pair of NUMA nodes. We confirm the installation of Intel® oneCCL Bindings for PyTorch and run the next command (as documented 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 extra efficiency tuning choices.
In earlier posts (e.g., here) we mentioned the PyTorch/XLA library and its use of XLA compilation to allow PyTorch primarily based coaching on XLA devices 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 machine. With the institution of the OpenXLA Project, one of many acknowledged objectives was to help excessive efficiency throughout all {hardware} backends, together with CPU (see the CPU RFC here). The code block under 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 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 help matures.
We summarize the outcomes of a subset of our experiments within the desk under. 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 post. 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. 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 may additional shut the hole. As soon as once more, we emphasize that the comparative efficiency outcomes we now have reached are distinctive to this mannequin and runtime setting.
Amazon EC2 Spot Cases Reductions
The elevated availability of cloud-based CPU occasion sorts (in comparison with GPU occasion sorts) might suggest better alternative for acquiring compute energy at discounted charges, e.g., by means of Spot Occasion utilization. Amazon EC2 Spot Instances are situations from surplus cloud service capability which can be supplied for a reduction of as a lot as 90% off the On-Demand pricing. In alternate for the discounted worth, AWS maintains the suitable 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 Instance price 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 worth performances (to 2.43X).
Whereas the runtime efficiency outcomes of the optimized CPU coaching of our toy mannequin (and our chosen setting) 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 parts that aren’t supported by GPU) might outcome within the CPU efficiency matching or beating that of the GPU. And even in instances the place the efficiency hole is just not bridged, there might very nicely be instances 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 may have enormous implications on improvement productiveness and on end-product deployment technique. Whereas the character of the CPU structure is much less amiable to many ML functions when in comparison with the GPU, there are various instruments and methods out there for enhancing its efficiency — a choose few of which we now have mentioned and demonstrated on this put up.
On this put up we centered optimizing coaching on CPU. Please you should definitely take a look at our many other posts on medium overlaying all kinds of matters pertaining to efficiency evaluation and optimization of machine studying workloads.
