Understanding Application Performance with Roofline Modeling

with calculating an software’s efficiency is that the real-world efficiency and theoretical efficiency can differ. With an ecosystem of merchandise that’s rising with excessive efficiency wants resembling Excessive Efficiency Computing (HPC), gaming, or within the present panorama – Giant Language Fashions (LLMs), it’s important to calculate precisely the efficiency of an software.

Merely measuring theoretical GFLOPs (Floating-Level Operations Per Second) shouldn’t be sufficient, as purposes hardly ever attain these maximums in the true world. That is the place the Roofline Mannequin is available in, providing a transparent visible methodology to estimate an software’s efficiency and highlighting the essential function of hardware-specific optimizations.

Why easy metrics aren’t sufficient

Once we take into consideration measuring efficiency, there are a couple of metrics that come to thoughts:

• Execution time: This tells you how lengthy a activity took however gives no perception into why.
• Cycles per Directions (CPI): This only measures the processor’s compute efficiency.
• Serial vs Parallel execution: Measures compute efficiency overlooking any {hardware} optimizations.
• Floating Level Operations Per Second (FLOP/s): This only represents a theoretical most which is commonly not achievable in a real-world state of affairs.

Whereas these are good metrics, they often don’t present sufficient info. For example, utilizing the Floating Level Operations Per Seconds is a theoretical restrict which isn’t typically achieved. So utilizing that because the solely metric shouldn’t be sufficient because it ignores a standard efficiency limiter – knowledge motion.

Roofline Modeling

The Roofline Mannequin is a robust device that visually maps an software’s efficiency in opposition to the capabilities of a particular {hardware} structure, resembling a CPU or GPU. The mannequin will get its title from the form of the graph it produces, which contains a “roof” composed of a slanted line and a flat, horizontal line. This form represents the last word efficiency limits imposed by the {hardware}.

From this modeling approach, there are two parameters which outline the achievable limits with {hardware}:

• Information motion: The time it takes to maneuver knowledge, calculated as the full knowledge dimension divided by the system’s peak reminiscence bandwidth.
• Computation: The time required for calculations, decided by dividing the full variety of floating-point operations by the system’s peak compute efficiency (generally measured in GFLOP/s).

The full execution time of an software is set by the larger of those two values: max {data_movement, computation}.

Regardless of the {hardware} having higher compute efficiency, knowledge motion can typically turn into the bottleneck. Roofline Modeling introduces the idea of Arithmetic Depth (AI). AI is the ratio of floating-point operations carried out for each byte of information moved from reminiscence.

• An algorithm with excessive Arithmetic Depth is taken into account compute-hungry. Its efficiency is proscribed by how rapidly calculations will be carried out.
• An algorithm with low Arithmetic Depth is taken into account data-hungry. Its efficiency is proscribed by how rapidly knowledge will be moved.

Understanding the graph

A Roofline graph plots the Attainable FLOP/s (y-axis) in opposition to the Arithmetic Depth (x-axis). The “roof” itself exhibits the {hardware}’s limitations. The slanted a part of the roof represents the height knowledge bandwidth (in GB/s), whereas the flat half represents the height computational efficiency (in GFLOPS). Observe that all the pieces within the picture is in a logarithmic scale.

• Factors beneath the roof: Point out suboptimal efficiency indicating scope of enchancment.
• Factors hitting the slanted line: Information hungry software. Its efficiency is proscribed by knowledge bandwidth.
• Factors hitting the flat line: Compute hungry software. It’s utilizing the total computational energy of the processor.

Why is Roofline Modeling necessary?

Roofline Modeling gives a visible, intuitive strategy to perceive software efficiency, displaying key traits like Operational Depth, GPU capabilities, and attainable FLOP/s. This type of modeling helps the programmer make focused optimizations to their software for {hardware} with which higher outcomes will be obtained.

• Bottleneck evaluation: Having a visible assist makes it straightforward for the developer to determine the place the bottleneck is – reminiscence or efficiency. If the applying is reminiscence intensive, a developer can concentrate on bettering knowledge locality with strategies like caching or loop tiling. If it’s compute intensive, the main target can shift to enabling extra parallel computations or leveraging compiler optimizations.
• {Hardware} and software program design: Software program engineers mustn’t worry the underlying {hardware}. As an alternative, the {hardware} design needs to be embraced and optimized. Software program engineers can use insights from Roofline Modeling to embrace and optimize for the precise structure they’re utilizing.

Roofline Modeling in Motion

To carry out Roofline Modeling, we have to profile the applying to grasp the efficiency. From profiling, we will get metrics resembling Floating Level Operations (FLOPs) and reminiscence bandwidth utilization, each of that are required for Roofline Modeling. This text explores two of those instruments – Nvidia’s ncu which is the Nsight Compute CLI for GPU evaluation and PyTorch’s profiler, particularly for purposes utilizing PyTorch.

For detailed CUDA kernel optimization and exact FLOP/byte calculations, ncu gives direct GPU {hardware} counter info. In distinction, torch.profiler.profile gives a higher-level perspective inside PyTorch, serving to within the understanding of operator-level efficiency, tensor reminiscence utilization, and the general software conduct encompassing each CPU and GPU actions.

Profiling with ncu

ncu is the command line interface which is used for profiling CUDA kernels [2]. It might probably show outcomes immediately within the terminal or save them to a log file for later evaluation. To construct a Roofline mannequin, we have to seize the precise metrics that may enable us to calculate Arithmetic Depth.

We’ll use the PyTorch ImageNet repository [3] as our instance. It’s a sensible choice as a result of it’s straightforward to grasp, well-documented by PyTorch, and works with their profiler, so we will actually dig into the efficiency.

Step 1: Run the ncu command to gather metrics

Step one is to run the applying by means of ncu to gather the mandatory hardware-level knowledge. The command seems to be like this:

ncu --log-file  
    --metrics  
    --target-processes all 
    python3 

• log-file: The log file during which we need to retailer the outcomes.
• metrics: That is crucial parameter and depicts the metrics that we need to seize. For calculating Arithmetic Depth, we think about:
  • dram__sectors_write.sum : sum of DRAM sectors written
  • dram__sectors_read.sum : sum of DRAM sectors learn
  • smsp__sass_thread_inst_executed_op_fadd_pred_on.sum : sum of floating-point additions
  • smsp__sass_thread_inst_executed_op_fmul_pred_on.sum : sum of floating-point multiplications
  • smsp__sass_thread_inst_executed_op_ffma_pred_on.sum : sum of floating-point fused multiply add operations
• target-process: all flag ensures that we profile your complete software.

Our ncu command adjustments to:

ncu --log-file logs_example --metrics dram__sectors_write.sum, 
dram__sectors_read.sum, 
smsp__sass_thread_inst_executed_op_fadd_pred_on.sum,  
smsp__sass_thread_inst_executed_op_fmul_pred_on.sum, 
smsp__sass_thread_inst_executed_op_ffma_pred_on.sum 
--target-processes all python3 
predominant.py /imagenet --arch resnet50 --epochs 1 --batch-size 10 
--print-freq 10 --seed 42

Step 2: Calculating FLOPs from the metrics

As soon as the profiler has run, we will mixture the collected metrics to calculate the full floating-point operations. The method is:

[FLOPs = 2 * FMA_count + FADD_count + FMUL_count]

• FLOPs: Depend of Floating Level Operations.
• FMA_count: Fused Multiply-Add (FMA) operations usually rely as 2 FLOPs (one multiplication and one addition). That is represented by the smsp__sass_thread_inst_executed_op_ffma_pred_on.sum metric.
• FADD_count: That is represented by the smsp__sass_thread_inst_executed_op_fadd_pred_on.sum metric.
• FMUL_count: That is represented by the smsp__sass_thread_inst_executed_op_fmul_pred_on.sum metric.

Step 3: Calculate the bytes transferred

Subsequent, we calculate the full knowledge transferred to and from DRAM. The ncu metrics present the variety of DRAM sectors learn and written. Assuming a standard sector dimension of 32 bytes for contemporary GPUs:

[Total_DRAM_bytes = (dram__sectors_read.sum + dram__sectors_write.sum) * 32]

Step 4: Calculate the Arithmetic Depth

With FLOPs and complete bytes, we will now calculate the Arithmetic Depth:

[AI = FLOPs / Total_DRAM_Bytes]

Step 5: Calculate execution time

To search out the applying’s efficiency in FLOP/s, we additionally want the execution time. For this, we will use NVIDIA Nsight Methods (nsys), a system-wide profiler that may precisely measure the runtime of software segments. We run our software once more, this time with nsys, to generate a time-based report. From this report, we will extract the full GPU operating time.

nsys profile -f true -o  python3 

Our nsys command adjustments to:

nsys profile -f true -o time.qdrep python3 predominant.py /imagenet 
--arch resnet50 --epochs 1 --batch-size 10 --print-freq 10 
--seed 42

After operating this command, we will get the GPU_RUNNING_TIME.

Step 6: Calculate the applying efficiency

Lastly, we calculate the achieved efficiency in FLOP/s by dividing the full FLOPs by the execution time:

[FLOP/s = FLOPs / GPU_RUNNING_TIME]

This worth offers us the “attainable FLOP/s” that we will plot on our Roofline graph.

Profiling with torch

For purposes written in PyTorch, the built-in torch.profiler.profile gives a user-friendly strategy to collect efficiency knowledge. There are 2 choices which might be supplied to the builders:

• Use the Profiler Context Supervisor
• Focusing on Profiling for particular neural community layers

Profiler Context Supervisor

The a part of the code that we need to profile will be wrapped throughout the with torch.profiler.profile() context supervisor. Within the with assertion, you possibly can outline the actions to hint (CPU, CUDA, or each), set a schedule to profile particular coaching steps, and select whether or not to report tensor shapes, reminiscence utilization, or FLOPs. As soon as contained in the context, you could name prof.step() on the finish of every iteration to sign the profiler to advance, particularly when a schedule is used.

with profile(
    actions=,
    schedule=torch.profiler.schedule(),
    record_shapes=,
    profile_memory=,
    with_flops=
) as prof:

    ....
    prof.step()

• actions: Specify whether or not to profile the CPU, CUDA or each.
• schedule: Helpful for profiling a number of steps within the coaching loop. If the schedule parameter is used, the profiler must name prof.step() to maneuver to the subsequent step.
• record_shapes: Whether or not to report the shapes of the tensors.
• profile_memory: To seize reminiscence utilization
• with_flops: That is experimental however is used to FLOPs with operators.

Our profiler command adjustments to:

with profile(
    actions=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
    schedule=torch.profiler.schedule(wait=1, warmup=1, lively=3, repeat=2),
    record_shapes=True,
    profile_memory=True,
    with_flops=True
) as prof:

Focusing on Profiling for particular neural community layers

The profiler will also be utilized in a extra focused method to research particular layers of a neural community. That is helpful to test whether or not some particular layer is contributing extra to the efficiency than the opposite layers giving the developer the choice of modifying particular layers. Whereas utilizing that is very straightforward to make use of, most often, the primary possibility works higher. The PyTorch profiler outcomes will also be exported and visualized on a TensorBoard.

profiler.begin()
self.conv2(x)
profiler.cease()

LLMs and Roofline Modeling

Coming to the subject everybody has been ready for – does Roofline Modeling assist with LLM efficiency calculation? The quick reply is sure.

LLMs are complicated neural community architectures with billions of parameters and the large datasets that they course of. Whereas coaching is a really resource-intensive activity, inference and advantageous tuning the mannequin additionally must be environment friendly.

• Bottlenecks: LLMs throughout inference can endure from bottlenecks as a result of sheer quantity of parameters that it’s working with. These parameters are the weights of the fashions and so they trigger reminiscence bandwidth points. Utilizing Roofline Modeling, the precise layers will be profiled for the bottlenecks.
• {Hardware} choice: As most organizations fine-tune current fashions quite than coaching them from scratch, selecting the best infrastructure is essential for managing prices. This underscores the significance of selecting optimum infrastructure for coaching. For instance, selecting the {hardware} in line with your LLM structure or optimizing your mannequin to run on a particular structure can lower coaching and inference prices.

Conclusion

The Roofline Mannequin gives a robust visible evaluation of software efficiency optimization. By visualizing the applying efficiency throughout reminiscence and compute, a transparent steering is supplied in selecting the easiest way to strategy optimizations. Whereas this text solely thought-about Naive Roofline Fashions, there are extra superior strategies resembling Hierarchical Roofline Fashions or including ceilings for particular compute optimizations.

References

[1] https://docs.nersc.gov/tools/performance/roofline/

[2] https://docs.nvidia.com/nsight-compute/NsightComputeCli/index.html

[3] https://github.com/pytorch/examples/tree/main/imagenet

[4] https://developer.nvidia.com/nsight-systems