CUDA: How many default streams are there on a single device? - cuda

I have read the documentation carefully but still, confused due to the large amount of information for different CUDA versions.
Is it that there is only one default stream on the entire device or
there is one default stream per-process on the HOST CPU? If the answer depends on the version of CUDA, could you also list the situation for different CUDA versions?

By default, CUDA has a per-process default stream. There is a compiler flag --default-stream per-thread which changes the behaviour to per-host-thread default stream, see the documentation.
Note that streams and host threads are programming-level abstractions for hardware details. Even with a single process, there is a limited number of streams you can use concurrently, depending on the hardware. For example, on the Fermi architecture, all streams were multiplexed into a single hardware queue, but since Kepler there are 32 separate hardware queues (see CUDA Streams: Best Practices and Common Pitfalls).
Since the programming guide does not talk about multiple processes in this part, I believe these abstractions do not define the behaviour of multi-process scenarios. As for multi-process, the right term is "CUDA context" which is created for each process and even each host thread (when using the runtime API). The question of how many contexts can be active on a device at the same time: the guide says in 3.4 Compute modes that in the default mode, "Multiple host threads can use the device". Since the following exclusive-process mode talks about CUDA contexts instead, I assume that this means that the description of the default mode covers even multiple host threads from multiple processes.
For more info about multi-process concurrency see e.g. How do I use Nvidia Multi-process Service (MPS) to run multiple non-MPI CUDA applications?, Unleash legacy MPI codes with Kepler's Hyper-Q and CUDA Streams: Best Practices and Common Pitfalls.
Finally, note that multi-process concurrency works this way since the Kepler architecture, which is the oldest supported architecture nowadays. Since the Pascal architecture there is support for compute preemption (see 3.4 Compute modes for details).

Related

multiple processes of video processing to GPU cores

Can we assign a number of processes (i.e. 100-500 processes) to GPU, each process running on a GPU core?
In my application of video processing, I have to use ffmpeg library to proceed video and audio. If there are like more than 100 or even 500 such independent processes, I guess it's faster to assign each process to a GPU. However, I don't know if we can do it, and to do it, what libraries, tools are necessary? CUDA?
Can we assign a number of processes (i.e. 100-500 processes) to GPU, each process running on a GPU core?
No, you can't. In general it's not possible to schedule anything on a GPU core per se. This level of "scheduling" is handled mainly by the mechanics of the CUDA architecture and runtime system.
The basic idea is to expose parallelism at a fairly low level in your code (e.g. at the loop level) and with proper use of a GPU acceleration syntax (such as CUDA, OpenACC, OpenCL, etc.) the GPU can often make such elements of your program run faster.
But the GPU is not designed to be a drop-in replacement for CPU cores. There is the scheduling factor that I mentioned already, as well as the fact that codes generally need to be compiled for the GPU specifically.

Multiple GPUs and Multiple Executables

Suppose I have 4 GPUs and would like to run 50 CUDA programs in parallel. My question is: is the NVIDIA driver smart enough to run the 50 CUDA programs on the different GPUs or do I have to set the CUDA device for each program?
thank you
The first point to make is that you cannot run 50 applications in parallel on 4 GPUs on just about any CUDA platform. If you have a Hyper-Q capable GPU, there is the possibility of up to 32 threads or MPI processes queuing work to the GPU. Otherwise there is a single command queue.
For anything other than the latest Kepler Tesla cards, CUDA driver only supports a single active context at a time. If you run more that one application on a GPU, the processes will both have contexts which just contend with one another in a "first come, first serve" basis. If one application blocks the other with a long running kernel or similar, there is no pre-emption or anything else which makes the process yield to another process. When the GPU is shared with a display manager, there is a watchdog timer that will impose an upper limit of a few seconds before the application will get its context killed. The result is that only one context ever runs on the hardware at a time. Context switching isn't free, and there is a performance penalty to having multiple processes contending for a single device.
Furthermore, every context present on a GPU requires device memory. On the platform you are asking about, linux, there is no memory paging, so every context's resources must coexist in GPU memory. I don't believe it would be possible to have 12 non-trivial contexts running on any current GPU simultaneously - you would run out of available memory well before that number. Trying to run more applications would result in an context establishment failure.
As for the behaviour of the driver distributing multiple applications on multiple GPUs, AFAIK the linux driver doesn't do any intelligent distribution of processes amongst GPUs, except when one or more of the GPUs are in a non-default compute mode. If no device is specifically requested, the driver will always try and find the first valid, free GPU it can run a process or thread on. If a GPU is busy and marked compute exclusive (either thread or process) or marked prohibited, then the driver will skip over it when trying to find a GPU to run on. If all GPUs are exclusive and occupied or prohibited, then the application will fail with a no valid device available error.
So in summary,for everything other than Hyper-Q devices, there is no performance gain in doing what you are asking about (quite the opposite) and I would expected it to break if you tried. A much saner approach would be to use compute exclusivity in combination with a resource managing task scheduler like Torque or one of the (former) Sun Grid Engine versions, which could schedule your processes to run in an orderly fashion according to the availability of GPUs. This is how most general purpose HPC clusters deal with scheduling in multi-gpu environments.

Do all GPUs use the same architecture?

I have some experience with nVIDIA CUDA and am now thinking about learning openCL too. I would like to be able to run my programs on any GPU. My question is: does every GPU use the same architecture as nVIDIA (multi-processors, SIMT stracture, global memory, local memory, registers, cashes, ...)?
Thank you very much!
Starting with your stated goal:
"I would like to be able to run my programs on any GPU."
Then yes, you should learn OpenCL.
In answer to your overall question, other GPU vendors do use different architectures than Nvidia GPUs. In fact, GPU designs from a single vendor can vary by quite a bit, depending on the model.
This is one reason that a given OpenCL code may perform quite differently (depending on your performance metric) from one GPU to the next. In fact, to achieve optimized performance on any GPU, an algorithm should be "profiled" by varying, for example, local memory size, to find the best algorithm settings for a given hardware design.
But even with these hardware differences, the goal of OpenCL is to provide a level of core functionality that is supported by all devices (CPUs, GPUs, FPGAs, etc) and include "extensions" which allow vendors to expose unique hardware features. Although OpenCL cannot hide significant differences in hardware, it does guarantee portability. This makes it much easier for a developer to start with an OpenCL program tuned for one device and then develop a program optimized for another architecture.
To complicate matters with identifying hardware differences, the terminology used by CUDA is different than that used by OpenCL, for example, the following are roughly equivalent in meaning:
CUDA: OpenCL:
Thread Work-item
Thread block Work-group
Global memory Global memory
Constant memory Constant memory
Shared memory Local memory
Local memory Private memory
More comparisons and discussion can be found here.
You will find that the kinds of abstraction provided by OpenCL and CUDA are very similar. You can also usually count on your hardware having similar features: global mem, local mem, streaming multiprocessors, etc...
Switching from CUDA to OpenCL, you may be confused by the fact that many of the same concepts have different names (for example: CUDA "warp" == OpenCL "wavefront").

differences between virtual and real architecture of cuda

Trying to understand the differences between virtual and real architecture of cuda, and how the different configurations will affect the performance of the program, e.g.
-gencode arch=compute_20,code=sm_20
-gencode arch=compute_20,code=sm_21
-gencode arch=compute_21,code=sm_21
...
The following explanation was given in NVCC manual,
GPU compilation is performed via an intermediate representation, PTX
([...]), which can
be considered as assembly for a virtual GPU architecture. Contrary to an actual graphics
processor, such a virtual GPU is defined entirely by the set of capabilities, or features,
that it provides to the application. In particular, a virtual GPU architecture provides a
(largely) generic instruction set, and binary instruction encoding is a non-issue because
PTX programs are always represented in text format.
Hence, a nvcc compilation command always uses two architectures: a compute
architecture to specify the virtual intermediate architecture, plus a real GPU architecture
to specify the intended processor to execute on. For such an nvcc command to be valid,
the real architecture must be an implementation (someway or another) of the virtual
architecture. This is further explained below.
The chosen virtual architecture is more of a statement on the GPU capabilities that
the application requires: using a smallest virtual architecture still allows a widest range
of actual architectures for the second nvcc stage. Conversely, specifying a virtual
architecture that provides features unused by the application unnecessarily restricts the
set of possible GPUs that can be specified in the second nvcc stage.
But still don't quite get how the performance will be affected by different configurations (or, maybe only affect the selection of the physical GPU devices?). In particular, this statement is most confusing to me:
In particular, a virtual GPU architecture provides a
(largely) generic instruction set, and binary instruction encoding is a non-issue because
PTX programs are always represented in text format.
The NVIDIA CUDA Compiler Driver NVCC User Guide Section on GPU Compilation provides a very thorough description of virtual and physical architecture and how the concepts are used in the build process.
The virtual architecture specifies the feature set that is targeted by the code. The table listed below shows some of the evolution of the virtual architecture. When compiling you should specify the lowest virtual architecture that has a sufficient feature set to enable the program to be executed on the widest range of physical architectures.
Virtual Architecture Feature List (from the User Guide)
compute_10 Basic features
compute_11 + atomic memory operations on global memory
compute_12 + atomic memory operations on shared memory
+ vote instructions
compute_13 + double precision floating point support
compute_20 + Fermi support
compute_30 + Kepler support
The physical architecture specifies the implementation of the GPU. This provides the compiler with the instruction set, instruction latency, instruction throughput, resource sizes, etc. so that the compiler can optimally translate the virtual architecture to binary code.
It is possible to specify multiple virtual and physical architecture pairs to the compiler and have the compiler back the final PTX and binary into a single binary. At runtime the CUDA driver will choose the best representation for the physical device that is installed. If binary code is not provided in the fatbinary file the driver can use the JIT runtime for the best PTX implementation.
"Virtual architecture" code will get compiled by a just-in-time compiler before being loaded on the device. AFAIK, it is the same compiler as the one NVCC invokes when building "physical architecture" code offline - so I don't know if there will be any differences in the resulting application performance.
Basically, every generation of the CUDA hardware is binary incompatible with previous generation - imagine next generation of Intel processors sporting ARM instruction set. This way, virtual architectures provide an intermediate representation of the CUDA application that can be compiled for compatible hardware. Every hardware generation introduces new features (e.g. atomics, CUDA Dynamic Parallelism) that require new instructions - that's why you need new virtual architectures.
Basically, if you want to use CDP you should compile for SM 3.5. You can compile it to device binary that will have assembly code for specific CUDA device generation or you can compile it to PTX code that can be compiled into device assembly for any device generation that provides these features.
The virtual architecture specifies what capabilities a GPU has and the real architecture specifies how it does it.
I can't think of any specific examples off hand. A (probably not correct) example may be a virtual GPU specifying the number of cores a card has, so code is generated targeting that number of cores, whereas the real card may have a few more for redundancy (or a few less due to manufacturing errors) and some methods of mapping to the cores that are actually in use, which can be placed on top of the more generic code generated in the first step.
You can think of the PTX code sort of like assembly code, which targets a certain architecture, which can then be compiled to machine code for a specific processor. Targeting the assembly code for the right kind of processor will, in general, generate better machine code.
well usually what nvidia writes as document causes people (including myself) to become more confused! (just me maybe!)
you are concerned with the performance, basically what this says is that don't be (probably) but you should.basically the GPU architecture is like nature. they run something on it and something happens. then they try to explain it. and then they feed it to you.
at the end should probably run some tests and see what configuration gives the best result.
the virtual architecture is what is designed to let you think freely. you should obey that, use as much as threads as you want, you can assign virtually everything as number of threads and blocks, doesn't matter, it will be translated to PTX and the device will run it.
the only problem is, if you assign more than 1024 threads per a single block you will get 0 s as the result, because the device(the real architecture) doesn't support it.
or for example your device support the CUDA 1.2, you can define double pointing variables in your code, but again you will get 0 s as the result because simply the device can't run it.
performance wise you have to know that every 32 thread (e.g. warps) have to access a single position in memory or else your access will be serialized and so on.
So I hope you've got the point by now, It is a relatively new science and GPU is a really sophisticated piece of hardware architecture, everybody is trying to make the best of it but it's a game of testing and a little knowledge of actual architecture behind CUDA. I suggest that search for GPU architecture and see how the virtual threads and thread blocks are actually implemented.

Concurrent GPU kernel execution from multiple processes

I have an application in which I would like to share a single GPU between multiple processes. That is, each of these processes would create its own CUDA or OpenCL context, targeting the same GPU. According to the Fermi white paper[1], application-level context switching is less then 25 microseconds, but the launches are effectively serialized as they launch on the GPU -- so Fermi wouldn't work well for this. According to the Kepler white paper[2], there is something called Hyper-Q that allows for up to 32 simultaneous connections from multiple CUDA streams, MPI processes, or threads within a process.
My questions: Has anyone tried this on a Kepler GPU and verified that its kernels are run concurrently when scheduled from distinct processes? Is this just a CUDA feature, or can it also be used with OpenCL on Nvidia GPUs? Do AMD's GPUs support something similar?
[1] http://www.nvidia.com/content/PDF/fermi_white_papers/NVIDIA_Fermi_Compute_Architecture_Whitepaper.pdf
[2] http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf
In response to the first question, NVIDIA has published some hyper-Q results in a blog here. The blog is pointing out that the developers who were porting CP2K were able to get to accelerated results more quickly because hyper-Q allowed them to use the application's MPI structure more or less as-is and run multiple ranks on a single GPU, and get higher effective GPU utilization that way. As mentioned in the comments, this (hyper-Q) feature is only available on K20 processors currently, as it is dependent on the GK110 GPU.
I've run simultaneous kernels from Fermi architecture it works wonderfully and in fact, is often the only way to get high occupancy from your hardware. I used OpenCL and you need to run a separate command queue from a separate cpu thread in order to do this. Hyper-Q is the ability to dispatch new data parallel kernels from within another kernel. This is only on Kepler.