Study profiling by inspecting concurrent and parallel Numba CUDA code in Nsight Programs

Optimization is an important a part of writing excessive efficiency code, irrespective of if you’re writing an online server or computational fluid dynamics simulation software program. Profiling lets you make knowledgeable choices concerning your code. In a way, optimization with out profiling is like flying blind: largely high-quality for seasoned professionals with knowledgeable data and fine-tuned instinct, however a recipe for catastrophe for nearly everybody else.

Following my preliminary collection CUDA by Numba Examples (see elements 1, 2, 3, and 4), we are going to examine a comparability between unoptimized, single-stream code and a barely higher model which makes use of stream concurrency and different optimizations. We are going to be taught, from the ground-up, learn how to use NVIDIA Nsight Programs to profile and analyze CUDA code. All of the code on this tutorial will also be discovered within the repo cako/profiling-cuda-nsight-systems.

NVIDIA recommends as finest observe to comply with the APOD framework (Assess, Parallelize, Optimize, Deploy). There are a selection of proprietary, open-source, free, and industrial software program for various kinds of assessments and profiling. Veteran Python customers could also be accustomed to primary profilers akin to cProfile, line_profiler, memory_profiler (sadly, unmaintaned as of 2024) and extra superior instruments like PyInstrument and Memray. These profilers goal particular facets of the “host” akin to CPU and RAM utilization.

Nonetheless, profiling “gadget” (e.g., GPU) code — and its interactions with the host — requires specialised instruments supplied by the gadget vendor. For NVIDIA GPUs, Nsight Programs, Nsight Compute, Nsight Graphics can be found for profiling completely different facets of computation. On this tutorial we are going to concentrate on utilizing Nsight Programs, which is a system-wide profiler. We are going to use it to profile Python code which interacts with the GPU through Numba CUDA.

To get began, you will have Nsight Programs CLI and GUI. The CLI may be put in individually and will likely be used to profile the code in a GPGPU-capable system. The total model consists of each CLI and GUI. Be aware that each variations may very well be put in in a system with no GPU. Seize the model(s) you want from the NVIDIA web site.

To make it simpler to visualise code sections within the GUI, NVIDIA additionally supplies the Python pip and conda-installable library nvtx which we are going to use to annotate sections of our code. Extra on this later.

On this part we are going to set our improvement and profiling surroundings up. Under are two quite simple Python scripts: kernels.py and run_v1.py. The previous will comprise all CUDA kernels, and the latter will function the entry level to run the instance. On this instance we’re following the “scale back” sample launched in article CUDA by Numba Examples Half 3: Streams and Occasions to compute the sum of an array.

#%%writefile kernels.py
import numba
from numba import cuda
THREADS_PER_BLOCK = 256
BLOCKS_PER_GRID = 32 * 40
@cuda.jit
def partial_reduce(array, partial_reduction):
i_start = cuda.grid(1)
threads_per_grid = cuda.blockDim.x * cuda.gridDim.x
s_thread = numba.float32(0.0)
for i_arr in vary(i_start, array.dimension, threads_per_grid):
s_thread += array[i_arr]
s_block = cuda.shared.array((THREADS_PER_BLOCK,), numba.float32)
tid = cuda.threadIdx.x
s_block[tid] = s_thread
cuda.syncthreads()
i = cuda.blockDim.x // 2
whereas i > 0:
if tid < i:
s_block[tid] += s_block[tid + i]
cuda.syncthreads()
i //= 2
if tid == 0:
partial_reduction[cuda.blockIdx.x] = s_block[0]
@cuda.jit
def single_thread_sum(partial_reduction, sum):
sum[0] = numba.float32(0.0)
for factor in partial_reduction:
sum[0] += factor
@cuda.jit
def divide_by(array, val_array):
i_start = cuda.grid(1)
threads_per_grid = cuda.gridsize(1)
for i in vary(i_start, array.dimension, threads_per_grid):
array[i] /= val_array[0]

#%%writefile run_v1.py
import argparse
import warnings
import numpy as np
from numba import cuda
from numba.core.errors import NumbaPerformanceWarning
from kernels import (
BLOCKS_PER_GRID,
THREADS_PER_BLOCK,
divide_by,
partial_reduce,
single_thread_sum,
)
# Ignore NumbaPerformanceWarning
warnings.simplefilter("ignore", class=NumbaPerformanceWarning)
def run(dimension):
# Outline host array
a = np.ones(dimension, dtype=np.float32)
print(f"Previous sum: {a.sum():.3f}")
# Array copy to gadget and array creation on the gadget.
dev_a = cuda.to_device(a)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
# Launching kernels to normalize array
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Array copy to host
dev_a.copy_to_host(a)
cuda.synchronize()
print(f"New sum: {a.sum():.3f}")
def predominant():
parser = argparse.ArgumentParser(description="Easy Instance v1")
parser.add_argument(
"-n",
"--array-size",
kind=int,
default=100_000_000,
metavar="N",
assist="Array dimension",
)
args = parser.parse_args()
run(dimension=args.array_size)
if __name__ == "__main__":
predominant()

This can be a easy script that may simply be run with:

$ python run_v1.py
Previous sum: 100000000.000
New sum: 1.000

We additionally run this code by means of our profiler, which simply entails calling nsys with some choices earlier than the decision to our script:

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v1 
python run_v1.py
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Previous sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-fb78.qdstrm'
[1/1] [========================100%] profile_run_v1.nsys-rep
Generated:
/content material/profile_run_v1.nsys-rep

You possibly can seek the advice of the Nsight CLI docs for all of the obtainable choices to the nsys CLI. For this tutorial we are going to all the time use those above. Let’s dissect this command:

• profile places nsys in profile mode. There are numerous different modes like export and launch.
• --trace cuda,osrt,nvtx ensures we “hear” to all CUDA calls (cuda), OS runtime library calls (osrt) and nvtx annotations (none on this instance). There are numerous extra hint choices akin to cublas, cudnn, mpi,dx11 and a number of other others. Test the docs for all choices.
• --gpu-metrics-device=all data GPU metrics for all GPUs, together with Tensor Core utilization.
• --cuda-memory-usage tracks GPU reminiscence utilization of kernels. It might considerably decelerate execution and requires --trace=cuda. We use it as a result of our scripts our fairly quick in any case.

If the command exited efficiently, we can have a profile_run_v1.nsys-rep within the present folder. We are going to open this file by launching the Nsight Programs GUI, File > Open. The preliminary view is barely complicated. So we are going to begin by decluttering: resize the Occasions View port to the underside, and reduce CPU, GPU and Processes underneath the Timeline View port. Now broaden solely Processes > python > CUDA HW. See Figures 1a and 1b.

First up, let’s discover our kernels. On the CUDA HW line, you’ll find inexperienced and crimson blobs, and really small slivers of sunshine blue (see Determine 1b). Should you hover over these you will notice tooltips saying, “CUDA Reminiscence operations in progress” for crimson and inexperienced, and “CUDA Kernel Working (89.7%)” for the sunshine blues. These are going to be the bread and butter of our profiling. On this line, we will inform when and the way reminiscence is being transferred (crimson and inexperienced) and when and the way our kernels are operating (gentle blue).

Let’s dig in just a little bit extra on our kernels. It is best to see three very small blue slivers, every representing a kernel name. We are going to zoom into the area by clicking and dragging the mouse from simply earlier than the beginning of the primary kernel name to only after the top of the final one, after which urgent Shift + Z. See Determine 2.

Now that we’ve discovered our kernels, let’s see some metrics. We open the GPU > GPU Metrics tabs for that. On this panel, can discover “Warp Occupancy” (beige) for compute kernels. One strategy to optimize CUDA code is to make sure that the warp occupancy is as near 100% as attainable for so long as attainable. Because of this our GPU shouldn’t be idling. We discover that that is occurring for the primary and final kernels however not the center kernel. That’s anticipated as the center kernel launches a single thread. One remaining factor to notice on this part is the GPU > GPU Metrics > SMs Lively > Tensor Lively / FP16 Lively line. This line will present whether or not the tensor cores are getting used. On this case it’s best to confirm that they aren’t.

Now let’s briefly have a look at the Occasions View. Proper click on Processes > python > CUDA HW and click on “Present in Occasions View”. Then type the occasions by descending length. In Determine 3, we see that the slowest occasions are two pageable reminiscence transfers. Now we have seen in CUDA by Numba Examples Half 3: Streams and Occasions that pageable reminiscence transfers may be suboptimal, and we should always desire page-locked or “pinned” reminiscence transfers. If we’ve gradual reminiscence transfers due to make use of of pageable reminiscence, the Occasions View could be a nice location to establish the place these gradual transfers may be discovered. Professional tip: you may isolate reminiscence transfers by proper clicking Processes > python > CUDA HW > XX% Reminiscence as a substitute.

On this part we discovered learn how to profile a Python program which makes use of CUDA, and learn how to visualize primary info of this program within the Nsight Programs GUI. We additionally observed that on this easy program, we’re utilizing pageable as a substitute of pinned reminiscence, that considered one of our kernels shouldn’t be occupying all warps, that the GPU is idle for fairly a while between kernels being run and that we aren’t utilizing tensor cores.

On this part we are going to discover ways to enhance our profiling expertise by annotation sections in Nsight Programs with NVTX. NVTX permits us to mark completely different areas of the code. It might probably mark ranges and instantaneous occasions. For a deeper look, test the docs. Under we create run_v2.py, which, along with annotating run_v1.py, additionally modifications this line:

a = np.ones(dimension, dtype=np.float32)

to those:

a = cuda.pinned_array(dimension, dtype=np.float32)
a[...] = 1.0

Subsequently, along with the annotations, we are actually utilizing a pinned reminiscence. If you wish to be taught extra in regards to the various kinds of recollections that CUDA helps, see the CUDA C++ Programming Information. It’s of relevance that this isn’t the one strategy to pin an array in Numba. A beforehand created Numpy array will also be created with a context, as defined within the Numba documentation.

#%%writefile run_v2.py
import argparse
import warnings
import numpy as np
import nvtx
from numba import cuda
from numba.core.errors import NumbaPerformanceWarning
from kernels import (
BLOCKS_PER_GRID,
THREADS_PER_BLOCK,
divide_by,
partial_reduce,
single_thread_sum,
)
# Ignore NumbaPerformanceWarning
warnings.simplefilter("ignore", class=NumbaPerformanceWarning)
def run(dimension):
with nvtx.annotate("Compilation", colour="crimson"):
dev_a = cuda.device_array((BLOCKS_PER_GRID,), dtype=np.float32)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Outline host array
a = cuda.pinned_array(dimension, dtype=np.float32)
a[...]  = 1.0
print(f"Previous sum: {a.sum():.3f}")
# Array copy to gadget and array creation on the gadget.
with nvtx.annotate("H2D Reminiscence", colour="yellow"):
dev_a = cuda.to_device(a)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
# Launching kernels to normalize array
with nvtx.annotate("Kernels", colour="inexperienced"):
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Array copy to host
with nvtx.annotate("D2H Reminiscence", colour="orange"):
dev_a.copy_to_host(a)
cuda.synchronize()
print(f"New sum: {a.sum():.3f}")
def predominant():
parser = argparse.ArgumentParser(description="Easy Instance v2")
parser.add_argument(
"-n",
"--array-size",
kind=int,
default=100_000_000,
metavar="N",
assist="Array dimension",
)
args = parser.parse_args()
run(dimension=args.array_size)
if __name__ == "__main__":
predominant()

Evaluating the 2 recordsdata, you may see it’s so simple as wrapping some GPU kernel calls with

with nvtx.annotate("Area Title", colour="crimson"):
...

Professional tip: you too can annotate features by putting the @nvtx.annotate decorator above their definition, robotically annotate all the pieces by calling your script with python -m nvtx run_v2.py, or apply the autoannotator selectively in you code by enabling or disabling nvtx.Profile(). See the docs!

Let’s run this new script and open the ends in Nsight Programs.

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v2 
python run_v2.py
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Previous sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-69ab.qdstrm'
[1/1] [========================100%] profile_run_v2.nsys-rep
Generated:
/content material/profile_run_v2.nsys-rep

Once more, we begin by minimizing all the pieces, leaving solely Processes > python > CUDA HW open. See Determine 4. Discover that we now have a brand new line, NVTX. On this line within the timeline window we should always see completely different coloured blocks equivalent to the annotation areas that we created within the code. These are Compilation, H2D Reminiscence, Kernels and D2H Reminiscence. A few of these my be too small to learn, however will likely be legible if you happen to zoom into the area.

The profiler confirms that this reminiscence is pinned, guaranteeing that our code is actually utilizing pinned reminiscence. As well as, H2D Reminiscence and D2H Reminiscence are actually taking lower than half of the time that they had been taking earlier than. Typically we will anticipate higher efficiency utilizing pinned reminiscence or prefetched mapped arrays (not supported by Numba).

Now we are going to examine whether or not we will enhance this code by introducing streams. The concept is that whereas reminiscence transfers are occurring, the GPU can begin processing the information. This permits a degree of concurrency, which hopefully will make sure that we’re occupying our warps as absolutely as attainable.

Within the code beneath we are going to break up the processing of our array into roughly equal elements. Every half will run in a separate stream, together with transferring knowledge and computing the sum of the array. Then, we synchronize all streams and sum their partial sums. At this level we will then launch normalization kernels for every stream independently.

We need to reply a number of questions:

1. Will the code beneath really create concurrency? Might we be introducing a bug?
2. Is it quicker than the code which makes use of a single stream?
3. Is the warp occupancy higher?

#%%writefile run_v3_bug.py
import argparse
import warnings
from math import ceil
import numpy as np
import nvtx
from numba import cuda
from numba.core.errors import NumbaPerformanceWarning
from kernels import (
BLOCKS_PER_GRID,
THREADS_PER_BLOCK,
divide_by,
partial_reduce,
single_thread_sum,
)
# Ignore NumbaPerformanceWarning
warnings.simplefilter("ignore", class=NumbaPerformanceWarning)
def run(dimension, nstreams):
with nvtx.annotate("Compilation", colour="crimson"):
dev_a = cuda.device_array((BLOCKS_PER_GRID,), dtype=np.float32)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Outline host array
a = cuda.pinned_array(dimension, dtype=np.float32)
a[...] = 1.0
# Outline areas for streams
step = ceil(dimension / nstreams)
begins = [i * step for i in range(nstreams)]
ends = [min(s + step, size) for s in starts]
print(f"Previous sum: {a.sum():.3f}")
# Create streams
streams = [cuda.stream()] * nstreams
cpu_sums = [cuda.pinned_array(1, dtype=np.float32) for _ in range(nstreams)]
devs_a = []
with cuda.defer_cleanup():
for i, (stream, begin, finish) in enumerate(zip(streams, begins, ends)):
cpu_sums[i][...] = np.nan
# Array copy to gadget and array creation on the gadget.
with nvtx.annotate(f"H2D Reminiscence Stream {i}", colour="yellow"):
dev_a = cuda.to_device(a[start:end], stream=stream)
dev_a_reduce = cuda.device_array(
(BLOCKS_PER_GRID,), dtype=dev_a.dtype, stream=stream
)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype, stream=stream)
devs_a.append(dev_a)
# Launching kernels to sum array
with nvtx.annotate(f"Sum Kernels Stream {i}", colour="inexperienced"):
for _ in vary(50):  # Make it spend extra time in compute
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK, stream](
dev_a, dev_a_reduce
)
single_thread_sum[1, 1, stream](dev_a_reduce, dev_a_sum)
with nvtx.annotate(f"D2H Reminiscence Stream {i}", colour="orange"):
dev_a_sum.copy_to_host(cpu_sums[i], stream=stream)
# Guarantee all streams are caught up
cuda.synchronize()
# Combination all 1D arrays right into a single 1D array
a_sum_all = sum(cpu_sums)
# Ship it to the GPU
with cuda.pinned(a_sum_all):
with nvtx.annotate("D2H Reminiscence Default Stream", colour="orange"):
dev_a_sum_all = cuda.to_device(a_sum_all)
# Normalize through streams
for i, (stream, begin, finish, dev_a) in enumerate(
zip(streams, begins, ends, devs_a)
):
with nvtx.annotate(f"Divide Kernel Stream {i}", colour="inexperienced"):
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK, stream](
dev_a, dev_a_sum_all
)
# Array copy to host
with nvtx.annotate(f"D2H Reminiscence Stream {i}", colour="orange"):
dev_a.copy_to_host(a[start:end], stream=stream)
cuda.synchronize()
print(f"New sum: {a.sum():.3f}")
def predominant():
parser = argparse.ArgumentParser(description="Easy Instance v3")
parser.add_argument(
"-n",
"--array-size",
kind=int,
default=100_000_000,
metavar="N",
assist="Array dimension",
)
parser.add_argument(
"-s",
"--streams",
kind=int,
default=4,
metavar="N",
assist="Array dimension",
)
args = parser.parse_args()
run(dimension=args.array_size, nstreams=args.streams)
if __name__ == "__main__":
predominant()

Let’s run the code and gather outcomes.

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v3_bug_4streams 
python run_v3_bug.py -s 4
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Previous sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-a666.qdstrm'
[1/1] [========================100%] profile_run_v3_bug_4streams.nsys-rep
Generated:
/content material/profile_run_v3_bug_4streams.nsys-rep

This system ran and yielded the proper reply. However after we open the profiling file (see Determine 6), we discover that there are two streams as a substitute of 4! And one is principally fully idle! What’s occurring right here?

There’s a bug within the creation of the streams. By doing

streams = [cuda.stream()] * nstreams

we are literally making a single stream and repeating it nstreams instances. So why are we seeing two streams as a substitute of 1? The truth that one shouldn’t be doing a lot computation must be an indicator that there’s a stream that we aren’t utilizing. This stream is the default stream, which we aren’t utilizing in any respect in out code since all GPU interactions are given a stream, the stream we created.

We are able to repair this bug with:

streams = [cuda.stream() for _ in range(nstreams)]
# Guarantee they're all completely different
assert all(s1.deal with != s2.deal with for s1, s2 in zip(streams[:-1], streams[1:]))

The code above will even guarantee they’re actually completely different streams, so it might have caught the bug had we had it within the code. It does so by checking the stream pointer worth.

Now we will run the fastened code with 1 stream and eight streams for comparability. See Figures 7 and eight, respectively.

$  nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v3_1stream 
python run_v3.py -s 1
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Previous sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-de65.qdstrm'
[1/1] [========================100%] profile_run_v3_1stream.nsys-rep
Generated:
/content material/profile_run_v3_1stream.nsys-rep

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v3_8streams 
python run_v3.py -s 8
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Previous sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-1fb7.qdstrm'
[1/1] [========================100%] profile_run_v3_8streams.nsys-rep
Generated:
/content material/profile_run_v3_8streams.nsys-rep

Once more, each give right outcomes. By opening the one with 8 streams we see that sure, the bug has been fastened (Determine 7). Certainly, we now see 9 streams (8 created + default). As well as, we see that they’re working on the identical time! So we’ve achieved concurrency!

Sadly, if we dig a bit deeper we discover that the concurrent code shouldn’t be essentially quicker. On my machine the essential part of each variations, from begin of reminiscence switch to the final GPU-CPU copy takes round 160 ms.

A possible wrongdoer is the warp occupancy. We discover that the warp occupancy is considerably higher within the single-stream model. The good points we’re getting on this instance in compute are probably being misplaced by not occupying our GPU as effectively. That is probably associated to the construction of the code which (artificially) calls approach too many kernels. As well as, if all threads are stuffed by a single stream, there is no such thing as a acquire in concurrency, since different streams need to be idle till assets unencumber.

This instance is vital as a result of it reveals that our preconceived notions of efficiency are simply hypotheses. They should be verified.

At this level of APOD, we’ve assessed, parallelized (each by means of threads and concurrency) and so the following step can be to deploy. We additionally observed a slight efficiency regression with concurrency, so for this instance, a single-stream model would probably be the one deployed. In manufacturing, the following step can be to comply with the following piece of code which is finest fitted to parallelization and restarting APOD.

On this article we noticed learn how to arrange, use and interpret outcomes from profiling Python code in NVIDIA Nsight Programs. C and C++ code may be analyzed very equally, and certainly many of the materials on the market makes use of C and C++ examples.

We additionally present how profiling can enable us to catch bugs and efficiency take a look at our packages, guaranteeing that the options we introduce really are bettering efficiency, and if they aren’t, why.