CUDA 10 added runtime API calls for putting streams (= queues) in "capture mode", so that instead of executing, they are returned in a "graph". These graphs can then be made to actually execute, or they can be cloned.
But what is the rationale behind this feature? Isn't it unlikely to execute the same "graph" twice? After all, even if you do run the "same code", at least the data is different, i.e. the parameters the kernels take likely change. Or - am I missing something?
PS - I skimmed this slide deck, but still didn't get it.
My experience with graphs is indeed that they are not so mutable. You can change the parameters with 'cudaGraphHostNodeSetParams', but in order for the change of parameters to take effect, I had to rebuild the graph executable with 'cudaGraphInstantiate'. This call takes so long that any gain of using graphs is lost (in my case). Setting the parameters only worked for me when I build the graph manually. When getting the graph through stream capture, I was not able to set the parameters of the nodes as you do not have the node pointers. You would think the call 'cudaGraphGetNodes' on a stream captured graph would return you the nodes. But the node pointer returned was NULL for me even though the 'numNodes' variable had the correct number. The documentation explicitly mentions this as a possibility but fails to explain why.
Task graphs are quite mutable.
There are API calls for changing/setting the parameters of task graph nodes of various kinds, so one can use a task graph as a template, so that instead of enqueueing the individual nodes before every execution, one changes the parameters of every node before every execution (and perhaps not all nodes actually need their parameters changed).
For example, See the documentation for cudaGraphHostNodeGetParams and cudaGraphHostNodeSetParams.
Another useful feature is the concurrent kernel executions. Under manual mode, one can add nodes in the graph with dependencies. It will explore the concurrency automatically using multiple streams. The feature itself is not new but make it automatic becomes useful for certain applications.
When training a deep learning model it happens often to re-run the same set of kernels in the same order but with updated data. Also, I would expect Cuda to do optimizations by knowing statically what will be the next kernels. We can imagine that Cuda can fetch more instructions or adapt its scheduling strategy when knowing the whole graph.
CUDA Graphs is trying to solve the problem that in the presence of too many small kernel invocations, you see quite some time spent on the CPU dispatching work for the GPU (overhead).
It allows you to trade resources (time, memory, etc.) to construct a graph of kernels that you can use a single invocation from the CPU instead of doing multiple invocations. If you don't have enough invocations, or your algorithm is different each time, then it won't worth it to build a graph.
This works really well for anything iterative that uses the same computation underneath (e.g., algorithms that need to converge to something) and it's pretty prominent in a lot of applications that are great for GPUs (e.g., think of the Jacobi method).
You are not going to see great results if you have an algorithm that you invoke once or if your kernels are big; in that case the CPU invocation overhead is not your bottleneck. A succinct explanation of when you need it exists in the Getting Started with CUDA Graphs.
Where task graph based paradigms shine though is when you define your program as tasks with dependencies between them. You give a lot of flexibility to the driver / scheduler / hardware to do scheduling itself without much fine-tuning from the developer's part. There's a reason why we have been spending years exploring the ideas of dataflow programming in HPC.
Related
In my use case, the global GPU memory has many chunks of data. Preferably, the number of these could change, but assuming the number and sizes of these chunks of data to be constant is fine as well. Now, there are a set of functions that take as input some of the chunks of data and modify some of them. Some of these functions should only start processing if others completed already. In other words, these functions could be drawn in graph form with the functions being the nodes and edges being dependencies between them. The ordering of these tasks is quite weak though.
My question is now the following: What is (on a conceptual level) a good way to implement this in CUDA?
An idea that I had, which could serve as a starting point, is the following: A single kernel is launched. That single kernel creates a grid of blocks with the blocks corresponding to the functions mentioned above. Inter-block synchronization ensures that blocks only start processing data once their predecessors completed execution.
I looked up how this could be implemented, but I failed to figure out how inter-block synchronization can be done (if this is possible at all).
I would create for any solution an array in memory 500 node blocks * 10,000 floats (= 20 MB) with each 10,000 floats being stored as one continuous block. (The number of floats be better divisible by 32 => e.g. 10,016 floats for memory alignment reasons).
Solution 1: Runtime Compilation (sequential, but optimized)
Use Python code to generate a sequential order of functions according to the graph and create (printing out the source code into a string) a small program which calls the functions in turn. Each function should read the input from its predecessor blocks in memory and store the output in its own output block. Python should output the glue code (as string) which calls all functions in the correct order.
Use NVRTC (https://docs.nvidia.com/cuda/nvrtc/index.html, https://github.com/NVIDIA/pynvrtc) for runtime compilation and the compiler will optimize a lot.
A further optimization would be to not store the intermediate results in memory, but in local variables. They will be enough for all your specified cases (Maximum of 255 registers per thread). But of course makes the program (a small bit) more complicated. The variables can be freely named. And you can have 500 variables. The compiler will optimize the assignment to registers and reusing registers. So have one variable for each node output. E.g. float node352 = f_352(node45, node182, node416);
Solution 2: Controlled run on device (sequential)
The python program creates a list with the order, in which the functions have to be called. The individual functions know, from what memory blocks to read and in what block to write (either hard-coded, or you have to submit it to them in a memory structure).
On the device kernel a for loop is run, where the order list is went through sequentially and the kernel from the list is called.
How to specify, which functions to call?
The function pointers in the list can be created on the CPU like the following code: https://leimao.github.io/blog/Pass-Function-Pointers-to-Kernels-CUDA/ (not sure, if it works in Python).
Or regardless of host programming language a separate kernel can create a translation table: device function pointers (assign_kernel). Then the list from Python would contain indices into this table.
Solution 3: Dynamic Parallelism (parallel)
With Dynamic Parallelism kernels themselves start other kernels (grids).
https://developer.nvidia.com/blog/cuda-dynamic-parallelism-api-principles/
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#cuda-dynamic-parallelism
There is a maximum depth of 24.
The state of the parent grid could be swapped to memory (which could take a maximum of 860 MB per level, probably not for your program). But this could be a limitation.
All this swapping could make the parallel version slower again.
But the advantage would be that nodes can really be run in parallel.
Solution 4: Use Cuda Streams and Events (parallel)
Each kernel just calls one function. The synchronization and scheduling is done from Python. But the kernels run asynchronously and call a callback as soon as they are finished. Each kernel running in parallel has to be run on a separate stream.
Optimization: You can use the CUDA graph API, with which CUDA learns the order of the kernels and can do additional optimizations, when replaying (with possibly other float input data, but the same graph).
For all methods
You can try different launch configurations from 32 or better 64 threads per block up to 1024 threads per block.
Let's assume that most, or all, of your chunks of data are large; and that you have many distinct functions. If the former does not hold it's not clear you will even benefit from having them on a GPU in the first place. Let's also assume that the functions are black boxes to you, and you don't have the ability to identify fine-graines dependencies between individual values in your different buffers, with simple, local dependency functions.
Given these assumptions - your workload is basically the typical case of GPU work, which CUDA (and OpenCL) have catered for since their inception.
Traditional plain-vanilla approach
You define multiple streams (queues) of tasks; you schedule kernels on these streams for your various functions; and schedule event-fires and event-waits corresponding to your function's inter-dependency (or the buffer processing dependency). The event-waits before kernel launches ensure no buffer is processed until all preconditions have been satisfied. Then you have different CPU threads wait/synchronize with these streams, to get your work going.
Now, as far as the CUDA APIs go - this is bread-and-butter stuff. If you've read the CUDA Programming Guide, or at least the basic sections of it, you know how to do this. You could avail yourself of convenience libraries, like my API wrapper library, or if your workload fits, a higher-level offering such as NVIDIA Thrust might be more appropriate.
The multi-threaded synchronization is a bit less trivial, but this still isn't rocket-science. What is tricky and delicate is choosing how many streams to use and what work to schedule on what stream.
Using CUDA task graphs
With CUDA 10.x, NVIDIA add API functions for explicitly creating task graphs, with kernels and memory copies as nodes and edges for dependencies; and when you've completed the graph-construction API calls, you "schedule the task graph", so to speak, on any stream, and the CUDA runtime essentially takes care of what I've described above, automagically.
For an elaboration on how to do this, please read:
Getting Started with CUDA Graphs
on the NVIDIA developer blog. Or, for a deeper treatment - there's actually a section about them in the programming guide, and a small sample app using them, simpleCudaGraphs .
White-box functions
If you actually do know a lot about your functions, then perhaps you can create larger GPU kernels which perform some dependent processing, by keeping parts of intermediate results in registers or in block shared memory, and continuing to the part of a subsequent function applied to such local results. For example, if your first kernels does c[i] = a[i] + b[i] and your second kernel does e[i] = d[i] * e[i], you could instead write a kernel which performs the second action after the first, with inputs a,b,d (no need for c). Unfortunately I can't be less vague here, since your question was somewhat vague.
We're trying to develop a Natural Language Processing application that has a user facing component. The user can call models through an API, and get the results back.
The models are pretrained using Keras with Theano. We use GPUs to speed up the training. However, prediction is still sped up significantly by using the GPU. Currently, we have a machine with two GPUs. However, at runtime (e.g. when running the user facing bits) there is a problem: multiple Python processes sharing the GPUs via CUDA does not seem to offer a parallelism speed up.
We're using nvidia-docker with libgpuarray (pygpu), Theano and Keras.
The GPUs are still mostly idle, but adding more Python workers does not speed up the process.
What is the preferred way of solving the problem of running GPU models behind an API? Ideally we'd utilize the existing GPUs more efficiently before buying new ones.
I can imagine that we want some sort of buffer before sending it off to the GPU, rather than requesting a lock for each HTTP call?
This is not an answer to your more general question, but rather an answer based on how I understand the scenario you described.
If someone has coded a system which uses a GPU for some computational task, they have (hopefully) taken the time to parallelize its execution so as to benefit from the full resources the GPU can offer, or something close to that.
That means that if you add a second similar task - even in parallel - the total amount of time to complete them should be similar to the amount of time to complete them serially, i.e. one after the other - since there are very little underutilized GPU resources for the second task to benefit from. In fact, it could even be the case that both tasks will be slower (if, say, they both somehow utilize the L2 cache a lot, and when running together they thrash it).
At any rate, when you want to improve performance, a good thing to do is profile your application - in this case, using the nvprof profiler or its nvvp frontend (the first link is the official documentation, the second link is a presentation).
GPU is really fast when it comes to paralleled computation and out performs CPU with being 15-30 ( some have reported even 50 ) times faster however,
GPU memory is very limited compared to CPU memory and communication between GPU memory and CPU is not as fast.
Lets say we have some data what won't fit into GPU ram but we still want to use
it's wonders to compute. What we can do is split that data into pieces and feed it into GPU one by one.
Sending large data to GPU can take time and one might think, what if we would split a data piece into two and feed the first half, run the kernel and then feed the other half while kernel is running.
By that logic we should save some time because data transfer should be going on while computation is, hopefully not interrupting it's job and when finished, it can just, well, continue it's job without needs for waiting a new data path.
I must say that I'm new to gpgpu, new to cuda but I have been experimenting around with simple cuda codes and have noticed that the function cudaMemcpy used to transfer data between CPU and GPU will block if kerner is running. It will wait until kernel is finished and then will do its job.
My question, is it possible to accomplish something like that described above and if so, could one show an example or provide some information source of how it could be done?
Thank you!
is it possible to accomplish something like that described above
Yes, it's possible. What you're describing is a pipelined algorithm, and CUDA has various asynchronous capabilities to enable it.
The asynchronous concurrent execution section of the programming guide covers the necessary elements in CUDA to make it work. To use your example, there exists a non-blocking version of cudaMemcpy, called cudaMemcpyAsync. You'll need to understand CUDA streams and how to use them.
I would also suggest this presentation which covers most of what is needed.
Finally, here is a worked example. That particular example happens to use CUDA stream callbacks, but those are not necessary for basic pipelining. They enable additional host-oriented processing to be asynchronously triggered at various points in the pipeline, but the basic chunking of data, and delivery of data while processing is occurring does not depend on stream callbacks. Note also the linked CUDA sample codes in that answer, which may be useful for study/learning.
Sorry if this is obvious, but I'm studying c++ and Cuda right now and wanted to know if this was possible so I could focus more on the relevant sections.
Basically my problem is highly parallelizable, in fact I'm running it on multiple servers currently. My program gets a work item(very small list) and runs a loop on it and makes one of 3 decisions:
keep the data(saves it),
Discard the data(doesn't do anything with it),
Process data further(its unsure of what to do so it modifies the data and resends it to the queue to process.
This used to be a recursion but I made each part independent and although I'm longer bound by one cpu but the negative effect of it is there's alot of messages that pass back/forth. I understand at a high level how CUDA works and how to submit work to it but is it possible for CUDA to manage the queue on the device itself?
My current thought process was manage the queue on the c++ host and then send the processing to the device, after which the results are returned back to the host and sent back to the device(and so on). I think that could work but I wanted to see if it was possible to have the queue on the CUDA memory itself and kernels take work and send work directly to it.
Is something like this possible with CUDA or is there a better way to do this?
I think what you're asking is if you can keep intermediate results on the device. The answer to that is yes. In other words, you should only need to copy new work items to the device and only copy finished items from the device. The work items that are still undetermined can stay on the device between kernel calls.
You may want to look into CUDA Thrust for this. Thrust has efficient algorithms for transformations, which can be combined with custom logic (search for "kernel fusion" in the Thrust manual.) It sounds like maybe your processing can be considered to be transformations, where you take a vector of work items and create two new vectors, one of items to keep and one of items that are still undetermined.
Is the host aware(or can it monitor) memory on device? My concern is how to be aware and deal with data that starts to exceed GPU onboard memory.
It is possible to allocate and free memory from within a kernel but it's probably not going to be very efficient. Instead, manage memory by running CUDA calls such as cudaMalloc() and cudaFree() or, if you're using Thrust, creating or resizing vectors between kernel calls.
With this "manual" memory management you can keep track of how much memory you have used with cudaMemGetInfo().
Since you will be copying completed work items back to the host, you will know how many work items are left on the device and thus, what the maximum amount of memory that might be required in a kernel call is.
Maybe a good strategy will be to swap source and destination vectors for each transform. To take a simple example, say you have a set of work items that you want to filter in multiple steps. You create vector A and fill it with work items. Then you create vector B of the same size and leave it empty. After the filtering, some portion of the work items in A have been moved to B, and you have the count. Now you run the filter again, this time with B as the source and A as the destination.
TL;DR version: "What's the best way to round-robin kernel calls to multiple GPUs with Python/PyCUDA such that CPU and GPU work can happen in parallel?" with a side of "I can't have been the first person to ask this; anything I should read up on?"
Full version:
I would like to know the best way to design context, etc. handling in an application that uses CUDA on a system with multiple GPUs. I've been trying to find literature that talks about guidelines for when context reuse vs. recreation is appropriate, but so far haven't found anything that outlines best practices, rules of thumb, etc.
The general overview of what we're needing to do is:
Requests come in to a central process.
That process forks to handle a single request.
Data is loaded from the DB (relatively expensive).
The the following is repeated an arbitrary number of times based on the request (dozens):
A few quick kernel calls to compute data that is needed for later kernels.
One slow kernel call (10 sec).
Finally:
Results from the kernel calls are collected and processed on the CPU, then stored.
At the moment, each kernel call creates and then destroys a context, which seems wasteful. Setup is taking about 0.1 sec per context and kernel load, and while that's not huge, it is precluding us from moving other quicker tasks to the GPU.
I am trying to figure out the best way to manage contexts, etc. so that we can use the machine efficiently. I think that in the single-gpu case, it's relatively simple:
Create a context before starting any of the GPU work.
Launch the kernels for the first set of data.
Record an event for after the final kernel call in the series.
Prepare the second set of data on the CPU while the first is computing on the GPU.
Launch the second set, repeat.
Insure that each event gets synchronized before collecting the results and storing them.
That seems like it should do the trick, assuming proper use of overlapped memory copies.
However, I'm unsure what I should do when wanting to round-robin each of the dozens of items to process over multiple GPUs.
The host program is Python 2.7, using PyCUDA to access the GPU. Currently it's not multi-threaded, and while I'd rather keep it that way ("now you have two problems" etc.), if the answer means threads, it means threads. Similarly, it would be nice to just be able to call event.synchronize() in the main thread when it's time to block on data, but for our needs efficient use of the hardware is more important. Since we'll potentially be servicing multiple requests at a time, letting other processes use the GPU when this process isn't using it is important.
I don't think that we have any explicit reason to use Exclusive compute modes (ie. we're not filling up the memory of the card with one work item), so I don't think that solutions that involve long-standing contexts are off the table.
Note that answers in the form of links to other content that covers my questions are completely acceptable (encouraged, even), provided they go into enough detail about the why, not just the API. Thanks for reading!
Caveat: I'm not a PyCUDA user (yet).
With CUDA 4.0+ you don't even need an explicit context per GPU. You can just call cudaSetDevice (or the PyCUDA equivalent) before doing per-device stuff (cudaMalloc, cudaMemcpy, launch kernels, etc.).
If you need to synchronize between GPUs, you will need to potentially create streams and/or events and use cudaEventSynchronize (or the PyCUDA equivalent). You can even have one stream wait on an event inserted in another stream to do sophisticated dependencies.
So I suspect the answer to day is quite a lot simpler than talonmies' excellent pre-CUDA-4.0 answer.
You might also find this answer useful.
(Re)Edit by OP: Per my understanding, PyCUDA supports versions of CUDA prior to 4.0, and so still uses the old API/semantics (the driver API?), so talonmies' answer is still relevant.