After Compute Capability 2.0 (Fermi) was released, I've wondered if there are any use cases left for shared memory. That is, when is it better to use shared memory than just let L1 perform its magic in the background?
Is shared memory simply there to let algorithms designed for CC < 2.0 run efficiently without modifications?
To collaborate via shared memory, threads in a block write to shared memory and synchronize with __syncthreads(). Why not simply write to global memory (through L1), and synchronize with __threadfence_block()? The latter option should be easier to implement since it doesn't have to relate to two different locations of values, and it should be faster because there is no explicit copying from global to shared memory. Since the data gets cached in L1, threads don't have to wait for data to actually make it all the way out to global memory.
With shared memory, one is guaranteed that a value that was put there remains there throughout the duration of the block. This is as opposed to values in L1, which get evicted if they are not used often enough. Are there any cases where it's better too cache such rarely used data in shared memory than to let the L1 manage them based on the usage pattern that the algorithm actually has?
2 big reasons why automatic caching is less efficient than manual scratch pad memory (applies to CPUs as well)
parallel accesses to random addresses are more efficient. Example: histogramming. Let's say you want to increment N bins, and each are > 256 bytes apart. Then due to coalescing rules, that will result in N serial reads/writes since global and cache memory is organized in large ~256byte blocks. Shared memory doesn't have that problem.
Also to access global memory, you have to do virtual to physical address translation. Having a TLB that can do lots of translations in || will be quite expensive. I haven't seen any SIMD architecture that actually does vector loads/stores in || and I believe this is the reason why.
avoids writing back dead values to memory, which wastes bandwidth & power. Example: in an image processing pipeline, you don't want your intermediate images to get flushed to memory.
Also, according to an NVIDIA employee, current L1 caches are write-through (immediately writes to L2 cache), which will slow down your program.
So basically, the caches get in the way if you really want performance.
As far as i know, L1 cache in a GPU behaves much like the cache in a CPU. So your comment that "This is as opposed to values in L1, which get evicted if they are not used often enough" doesn't make much sense to me
Data on L1 cache isn't evicted when it isn't used often enough. Usually it is evicted when a request is made for a memory region that wasn't previously in cache, and whose address resolves to one that is already in use. I don't know the exact caching algorithm employed by NVidia, but assuming a regular n-way associative, then each memory entry can only be cached in a small subset of the entire cache, based on it's address
I suppose this may also answer your question. With shared memory, you get full control as to what gets stored where, while with cache, everything is done automatically. Even though the compiler and the GPU can still be very clever in optimizing memory accesses, you can sometimes still find a better way, since you're the one who knows what input will be given, and what threads will do what (to a certain extent of course)
Caching data through several memory layers always needs to follow a cache-coherency protocol. There are several such protocols and the decision on which one is the most suitable is always a trade off.
You can have a look at some examples:
Related to GPUs
Generally for computing units
I don't want to get in many details, because it is a huge domain and I am not an expert. What I want to point out is that in a shared-memory system (here the term shared does not refer to the so called shared memory of GPUs) where many compute-units (CUs) need data concurrently there is a memory protocol that attempts to keep the data close to the units so that can fetch them as fast as possible. In the example of a GPU when many threads in the same SM (symmetric multiprocessor) access the same data there should be a coherency in the sense that if thread 1 reads a chunk of bytes from the global memory and in the next cycle thread 2 is going to access these data, then an efficient implementation would be such that thread 2 is aware that data are found already in L1 cache and can access it fast. This is what the cache coherency protocol attempts to achieve, to let all compute units be up to date with what data exist in caches L1, L2 and so on.
However, keeping threads up to date, or else, keeping threads in coherent states, comes at some cost which is essentially missing cycles.
In CUDA by defining the memory as shared rather than L1-cache you free it from that coherency protocol. So access to that memory (which is physically the same piece of whatever material it is) is direct and does not implicitly call the functionality of coherency protocol.
I don't know how fast should this be, I didn't perform any such benchmark but the idea is that since you don't pay anymore for this protocol the access should be faster!
Of course, the shared memory on NVIDIA GPUs is split in banks and if someone wants to use it for performance improvement should have a look at this before. The reason is bank conflicts that occur when two threads access the same bank and this causes serialization of the access..., but that's another thing link
Related
What is the best amount of main memory address to map to the same cache block?
I want to give an example to clarify my question.
Assume we want to map 512 KB main memory to 4 KB cache. If we choose to create a direct mapped cache: we are going to map 128 main memory addresses to the same cache block? Is it too much? How to know?
Your question doesn't have enough information to say how many blocks of main memory would map to the same cache line. What's missing is the block size: number of bytes per block.
As to whether any given cache is sufficient, that is a question of the job, also know as the workload — different workloads/jobs will exhibit different behaviors. The same job/program with different sets of input can exhibit different behaviors. Actual experimentation is pretty much the only way to know, but no matter what it will be a judgement call as to whether any given settings are "good enough".
When you have thrashing of cache lines, as is possible with direct mapped cache, that is detrimental to performance, so designers who identify this will try to fix that.
Software designers may chunk the data differently using algorithms to mitigate this given existing hardware they have to work with.
Hardware designers can implement different caching strategies, ranging from larger caches, to 2-, 4-, 8-way set associativity, to fully associative — for the hardware each of these incurs more hardware costs than a direct mapped cache (and may incur performance penalty as well), and further might price the component out of the market (make it too expensive) or exceed the process technology (i.e. be to big for one chip).
I'm current writing a smaller project in OpenCL, and I'm trying to find out what really causes memory coalescing. Every book on GPGPU programming says it's how GPGPUs should be programmed, but not why the hardware would prefer this.
So is it some special hardware component which merges data transfers? Or is it simply to better utilize the cache? Or is it something completely different?
Memory coalescing makes several different things more efficient. It is usually done before the requests hit the cache. Similar to the SIMT execution model it is a architectural trade-off. It enables GPUs to have a more efficient and very high performance memory system but also forces programmers to think carefully about their data layout.
Without coalescing either the cache needs to be able to serve a huge number of requests at the same time or memory access would take a lot longer as the different data transfers would need to be handled one at a time. This is even relevant when just checking if something is a hit or a miss.
Merging requests is rather easy to do, you just pick one transfer and then merge all requests with matching upper address bits. You just generate a single request per cycle and replay the load or store instruction until all threads have been handled.
Caches also stores consecutive bytes, 32/64/128Byte, this fits most applications well, is a good fit to modern DRAM and reduces the overhead for cache bookeeping information: The cache is organized in cachelines and each cacheline has a tag that indicates which addresses are stored in the line.
Modern DRAM uses wide interfaces and also long bursts: The memory of a GPU is typically organized in 32-bit or 64-bit wide channels with GDDR5 memory that has a burst length of 8. This means that every transaction at the DRAM interface has to fetch at least 32-bit*8=32 byte or 64-bit*8=64 byte at a time, even if just a single byte is required from these bytes. Designing data layouts that lead to coalesced requests helps to use the DRAM interface efficiently.
GPUs also have a huge number of parallel threads active at the same time and rather small cache at the same time. CPUs are often able to use their caches to reorder their memory requests to DRAM friendly patterns. The larger number of threads and smaller caches on GPUs make this "cache based coalescing" less efficient on GPUs, as the data will often not stay long enough in the cache to get merged at the cache with other requests to the same cacheline.
Despite the "random access" name on "RAM" (Random-access Memory), Double-Data-Rate #3 Random-Access Memory (DDR3-RAM) is faster at accessing consecutive positions rather than randomly.
Case in point: "CAS Latency" is the amount of time that DDR3 RAM will stall when you're accessing a new "column", as your RAM chip is literally charging up to serve the new data from another location on the chip.
EDIT: Jan Lucas argues that RAS Latency is more important in practice. See his comment for details.
There's roughly a 10ns delay whenever you switch columns. So, if you have a bunch of memory accesses, if you keep access a bunch of data 'close' to each other, then you don't invoke a CAS delay.
So if you have 20-words to access at a particular location, its more efficient to access those 20-words before moving to a new memory location (invoking a CAS delay). Otherwise, you'll have to invoke ANOTHER CAS delay to "switch back" between memory locations.
Its just around 10 nanoseconds, but that amount of time adds up over time.
I have encountered some Intel compiler intrinsic functions which I believe allow developers to bypass the cache?
http://software.intel.com/sites/products/documentation/doclib/stdxe/2013/composerxe/compiler/fortran-mac/GUID-AF42A867-B796-4D29-8FED-C20193FD87E0.htm
I have also come across the GCC compiler prefetch keyword, although I cannot admit to fully appreciating what this does.
With the above in mind I wondered if any members could either elaborate on the above (which I badly described) or provide other techniques which allow the developer to have close control over which data (or instructions) is/isn't loaded in the CPU cache?
This page contains a lot of information about all intrinsics:
Intel Intrinsics Guide
The series of instructions that will write data to memory, avoiding cache evictions are generally named _mm_stream_.... As the name implies, these are ideal for applications that write a large stream of data that is basically contiguous in memory and unlikely to be accessed again in the near future. So, for example, if you are mixing audio buffers and producing a single waveform output this would work well.
One of the keys to using these instructions effectively is taking advantage of write combining. If your write locations are scattered throughout memory, these instructions will stall as badly, or possibly worse than any other kind of memory storage instruction you attempt. Since these writes do not wind up in cache, if you're not filling an entire write buffer then essentially your operation becomes a write-through operation, requiring a stall until the write is completed. If you are writing contiguous memory locations then write combining will apply, and make your data writes much more efficient.
The flip side of that coin is prefetching. Prefetching tells the system to start pulling a memory address into the desired level of cache so that by the time the memory read is complete, you are ready to use the data. This is much harder to use, and requires an appropriate data "stride" which takes into account the cache sizes, cache line size, and the number of instructions which can execute before the memory read completes. Using the hinting parameter, you can "suggest" that the data goes into the L1, L2, or L3 cache, or that it is "non-temporal", meaning that you're just going to use it once and it should be evicted first before any other cache evictions. The hardware has its own prefetching heuristics that work well for most problems without explicit prefetching instructions, but the classic counter-example is a matrix transpose:
Prefetching examples
Prefetching is generally very difficult to use effectively except in some very specific cases like this. Without a more specific problem statement from you, this is about all I can provide.
I've always been wondering about the caching behaviour of global data in OpenCL.
Lets say I have a pointer to global memory in a kernel.
Now I read the location the pointer points to.
Later in the kernel I might need the same data again, so I read it again through the pointer.
Now the question is, will this data be cached, or will it be reread from global memory every single time because other threads could have modified it?
If it's not cached, then I'd have to make a local copy every time so I don't lose tons of performance by constantly accessing global memory.
I know this might be vendor specific, but what do the specs say about this?
There is some caching but the key to great GPU compute performance it is move "accessed many time" data to private or shared local memory and not re-read it. In a way, you can think of this as "you control the caching". In OpenCL this would be done in your kernel (in parallel!) and then you'd have a memory barrier (to ensure all work items have finished the copy) then your algorithm has access to the data in fast memory. See the matrix multiply example (since each column and row contributes to multiple output values, copying them to shared local memory accelerates the algorithm.
Those who want benefits of local cashing for work-items within a work-group, for example on FPGAs, can read this paper by Andrew Ling at IWOCL2017 https://www.iwocl.org/wp-content/uploads/iwocl2017-andrew-ling-fpga-sdk.pdf. It is a good example of having correct usage of local caches and clever communication for dataflow computing. Those who want convenience of cache in parallel peer-to-peer setting and still have hardware do this for them should consider POWER8 or POWER9 chips. These conveniences come at the cost: caching global or virtual memory cluster interconnect may have to have several TBs of bandwidth. Real question is: What is the value of caching for dataflow compute e.g. ML, especially on clusters, vs. reducing communication and increasing data reuse by other means.
Paging creates illusion that each process has infinite RAM by moving pages to and from disk. So if we have infinite memory(in some hypothetical situation), do we still need Paging? If yes, then why? I faced this question in an interview.
Assuming that "infinite memory" means infinite randomly accessable memory, or RAM, we will still need paging. Although paging is often associated with the ability to swap pages in and out of RAM to a hard disk to conserve memory, this is merely one aspect of paging. Here are some other reasons to have paging:
Security. Paging is a method to enforce operating system security and memory protection by ensuring that a processes cannot access the memory of another process and that it cannot modify the resident kernel.
Multitasking. Paging aids in multitasking by virtualizing the memory space, that is, address 0xFOO in Process A can be something completely different than 0xFOO in Process B
Memory Allocation. Paging aids in memory allocation by reducing fragmentation and ensuring RAM is only allocated when accessed. What this means is that although a process needs, suppose, 100MB of continuous RAM space, this need not be continuous physically. Additionally, when a program requests 100MB of space, the operating system will tell the program it's safe to use that 100MB of space, yet it will not be actually allocated until the program uses that space to its fullest.
Admittedly, the latter would not be entirely necessary if one had infinite RAM, nonetheless; it is always good practice to be eifficient even when we are not resource constrained. It also demonstrates a use for paging that is sometimes not considered.
This is a philosophical question, so here's a philosophical answer :)
The trick in this question is that you make assumptions about the infinite memory. It's ok to say "no, no need to use paging, BUT". And follow by:
The infinite memory has to be accessible within the acceptable time limit for memory access. If it's not (because infinity takes a lot of space, and the memory sits farther away from the processing unit), there's no difference between it and the disk, both are not satisfying the readily available memory requirement, which is what caching via pages tries to solve.
Take for example Amazon's S3, which for all practical purposes is infinite. If you can rely on S3 to satisfy all your memory requirements in the sense that when you need something fetched within time x you can fetch it from S3, there's no need to page anything or even hold it in "local" memory. Just get it from S3 whenever you need it, as many times as you need it. (Obviously this would have other repercussions like cost and network, but let's ignore that for now).
Of course you can always say that optimally you want memory access to be as fast as possible, and "fast enough" is probably slower than "fastest", so local memory access would give you better performance etc.
And last, if I had to envision a memory which is infinite and has the same access time no matter how "far" the memory unit is from the fetching unit, I would have to envision a sphere where the processing unit is in the middle, so that you can't argue that one memory unit is slower than the other because of the distance. Otherwise you could say that paging would be done internally within the memory so that access is faster to the memory units that are most used (or whatever algorithm you choose to use).