On my application I need to transform each line of an image, apply a filter and transform it back.
I want to be able to make multiple FFT at the same time using the GPU. More precisely, I'm using NVIDIA's CUDA. Now, some considerations:
CUDA's FFT library, CUFFT is only able to make calls from the host ( https://devtalk.nvidia.com/default/topic/523177/cufft-device-callable-library/).
On this topic (running FFTW on GPU vs using CUFFT), Robert Corvella says
"cufft routines can be called by multiple host threads".
I believed that doing all this FFTs in parallel would increase performance, but Robert comments
"the FFT operations are of reasonably large size, then just calling the cufft library routines as indicated should give you good speedup and approximately fully utilize the machine"
So,
Is this it? Is there no gain in performing more than one FFT at a time?
Is there any library that supports calls from the device?
Shoud I just use cufftPlanMany() instead (as refered in "is-there-a-method-of-fft-that-will-run-inside-cuda-kernel" by hang or as referred in the previous topic, by Robert)?
Or the best option is to call mutiple host threads?
(this 2 links limit is killing me...)
My objective is to get some discussion on what's the best solution to this problem, since many have faced similar situations.
This might be obsolete once NVIDIA implements device calls on CUFFT.
(something they said they are working on but there is no expected date for the release - something said on the discussion at the NVIDIA forum (first link))
So, Is this it? Is there no gain in performing more than one FFT at a time?
If the individual FFT's are large enough to fully utilize the device, there is no gain in performing more than one FFT at a time. You can still use standard methods like overlap of copy and compute to get the most performance out of the machine.
If the FFT's are small then the batched plan is a good way to get the most performance. If you go this route, I recommend using CUDA 5.5, as there have been some API improvements.
Is there any library that supports calls from the device?
cuFFT library cannot be used by making calls from device code.
There are other CUDA libraries, of course, such as ArrayFire, which may have options I'm not familiar with.
Shoud I just use cufftPlanMany() instead (as refered in "is-there-a-method-of-fft-that-will-run-inside-cuda-kernel" by hang or as referred in the previous topic, by Robert)?
Or the best option is to call mutiple host threads?
Batched plan is preferred over multiple host threads - the API can do a better job of resource management that way, and you will have more API-level visibility (such as through the resource estimation functions in CUDA 5.5) as to what is possible.
Related
I am somewhat familiar with the CUDA visual profiler and the occupancy spreadsheet, although I am probably not leveraging them as well as I could. Profiling & optimizing CUDA code is not like profiling & optimizing code that runs on a CPU. So I am hoping to learn from your experiences about how to get the most out of my code.
There was a post recently looking for the fastest possible code to identify self numbers, and I provided a CUDA implementation. I'm not satisfied that this code is as fast as it can be, but I'm at a loss as to figure out both what the right questions are and what tool I can get the answers from.
How do you identify ways to make your CUDA kernels perform faster?
If you're developing on Linux then the CUDA Visual Profiler gives you a whole load of information, knowing what to do with it can be a little tricky. On Windows you can also use the CUDA Visual Profiler, or (on Vista/7/2008) you can use Nexus which integrates nicely with Visual Studio and gives you combined host and GPU profile information.
Once you've got the data, you need to know how to interpret it. The Advanced CUDA C presentation from GTC has some useful tips. The main things to look out for are:
Optimal memory accesses: you need to know what you expect your code to do and then look for exceptions. So if you are always loading floats, and each thread loads a different float from an array, then you would expect to see only 64-byte loads (on current h/w). Any other loads are inefficient. The profiling information will probably improve in future h/w.
Minimise serialization: the "warp serialize" counter indicates that you have shared memory bank conflicts or constant serialization, the presentation goes into more detail and what to do about this as does the SDK (e.g. the reduction sample)
Overlap I/O and compute: this is where Nexus really shines (you can get the same info manually using cudaEvents), if you have a large amount of data transfer you want to overlap the compute and the I/O
Execution configuration: the occupancy calculator can help with this, but simple methods like commenting the compute to measure expected vs. measured bandwidth is really useful (and vice versa for compute throughput)
This is just a start, check out the GTC presentation and the other webinars on the NVIDIA website.
If you are using Windows... Check Nexus:
http://developer.nvidia.com/object/nexus.html
The CUDA profiler is rather crude and doesn't provide a lot of useful information. The only way to seriously micro-optimize your code (assuming you have already chosen the best possible algorithm) is to have a deep understanding of the GPU architecture, particularly with regard to using shared memory, external memory access patterns, register usage, thread occupancy, warps, etc.
Maybe you could post your kernel code here and get some feedback ?
The nVidia CUDA developer forum forum is also a good place to go for help with this kind of problem.
I hung back because I'm no CUDA expert, and the other answers are pretty good IF the code is already pretty near optimal. In my experience, that's a big IF, and there's no harm in verifying it.
To verify it, you need to find out if the code is for sure not doing anything it doesn't really have to do. Here are ways I can see to verify that:
Run the same code on the vanilla processor, and either take stackshots of it, or use a profiler such as Oprofile or RotateRight/Zoom that can give you equivalent information.
Running it on a CUDA processor, and doing the same thing, if possible.
What you're looking for are lines of code that have high occupancy on the call stack, as shown by the fraction of stack samples containing them. Those are your "bottlenecks". It does not take a very large number of samples to locate them.
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).
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.
Does anyone have any experience implementing a hash map on a CUDA Device? Specifically, I'm wondering how one might go about allocating memory on the Device and copying the result back to the Host, or whether there are any useful libraries that can facilitate this task.
It seems like I would need to know the maximum size of the hash map a priori in order to allocate Device memory. All my previous CUDA endeavors have used arrays and memcpys and therefore been fairly straightforward.
Any insight into this problem are appreciated. Thanks.
There is a GPU Hash Table implementation presented in "CUDA by example", from Jason Sanders and Edward Kandrot.
Fortunately, you can get information on this book and download the examples source code freely on this page:
http://developer.nvidia.com/object/cuda-by-example.html
In this implementation, the table is pre-allocated on CPU and safe multithreaded access is ensured by a lock function based upon the atomic function atomicCAS (Compare And Swap).
Moreover, newer hardware generation (from 2.0) combined with CUDA >= 4.0 are supposed to be able to use directly new/delete operators on the GPU ( http://developer.nvidia.com/object/cuda_4_0_RC_downloads.html?utm_source=http://forums.nvidia.com&utm_medium=http://forums.nvidia.com&utm_term=Developers&utm_content=Developers&utm_campaign=CUDA4 ), which could serve your implementation. I haven't tested these features yet.
cuCollections is a relatively new open-source library started by NVIDIA engineers aiming at implementing efficient containers on the GPU.
cuCollections (cuco) is an open-source, header-only library of GPU-accelerated, concurrent data structures.
Similar to how Thrust and CUB provide STL-like, GPU accelerated algorithms and primitives, cuCollections provides STL-like concurrent data structures. cuCollections is not a one-to-one, drop-in replacement for STL data structures like std::unordered_map. Instead, it provides functionally similar data structures tailored for efficient use with GPUs.
cuCollections is still under heavy development. Users should expect breaking changes and refactoring to be common.
At the moment it provides a fixed size hashtable cuco::static_map and one that can grow cuco::dynamic_map.
I recall someone developed a straightforward hash map implementation on top of thrust. There is some code for it here, although whether it works with current thrust releases is something I don't know. It might at least give you some ideas.
AFAIK, the hash table given in "Cuda by Example" does not perform too well.
Currently, I believe, the fastest hash table on CUDA is given in Dan Alcantara's PhD dissertation. Look at chapter 6.
BTW, warpcore is a framework for creating high-throughput, purpose-built hashing data structures on CUDA-accelerators. Hashing at the speed of light on modern CUDA-accelerators. You can find it here:
https://github.com/sleeepyjack/warpcore
There are ways of using cuda:
auto-paralleing tools such as PGI workstation;
wrapper such as Thrust(in STL style)
NVidia GPUSDK(runtime/driver API)
Which one is better for performance or learning curve or other factors?
Any suggestion?
Performance rankings will likely be 3, 2, 1.
Learning curve is (1+2), 3.
If you become a CUDA expert, then it will be next to impossible to beat the performance of your hand-rolled code using all the tricks in the book using the GPU SDK due to the control that it gives you.
That said, a wrapper like Thrust is written by NVIDIA engineers and shown on several problems to have 90-95+% efficiency compared with hand-rolled CUDA. The reductions, scans, and many cool iterators they have are useful for a wide class of problems too.
Auto-parallelizing tools tend to not do quite as good a job with the different memory types as karlphillip mentioned.
My preferred workflow is using Thrust to write as much as I can and then using the GPU SDK for the rest. This is largely a factor of not trading away too much performance to reduce development time and increase maintainability.
Go with the traditional CUDA SDK, for both performance and smaller learning curve.
CUDA exposes several types of memory (global, shared, texture) which have a dramatic impact on the performance of your application, there are great articles about it on the web.
This page is very interesting and mentions the great series of articles about CUDA on Dr. Dobb's.
I believe that the NVIDIA GPU SDK is the best, with a few caveats. For example, try to avoid using the cutil.h functions, as these were written solely for use with the SDK, and I've personally, as well as many others, have run into some problems and bugs in them, that are hard to fix (There also is no documentation for this "library" and I've heard that NVIDIA does not support it at all)
Instead, as you mentioned, use the one of the two provided APIs. In particular I recommend the Runtime API, as it is a higher level API, and so you don't have to worry quite as much about all of the low level implementation details as you do in the Device API.
Both APIs are fully documented in the CUDA Programming Guide and CUDA Reference Guide, both of which are updated and provided with each CUDA release.
It depends on what you want to do on the GPU. If your algorithm would highly benefit from the things thrust can offer, like reduction, prefix, sum, then thrust is definitely worth a try and I bet you can't write the code faster yourself in pure CUDA C.
However if you're porting already parallel algorithms from the CPU to the GPU, it might be easier to write them in plain CUDA C. I had already successful projects with a good speedup going this route, and the CPU/GPU code that does the actual calculations is almost identical.
You can combine the two paradigms to some extend, but as far as I know you're launching new kernels for each thrust call, if you want to have all in one big fat kernel (taking too frequent kernel starts out of the equation), you have to use plain CUDA C with the SDK.
I find the pure CUDA C actually easier to learn, as it gives you quite a good understanding on what is going on on the GPU. Thrust adds a lot of magic between your lines of code.
I never used auto-paralleing tools such as PGI workstation, but I wouldn't advise to add even more "magic" into the equation.