CUDA offers us the cuDeviceSetGraphMemAttribute() function for setting one of two attributes:
CU_GRAPH_MEM_ATTR_USED_MEM_HIGH: High watermark of memory, in bytes, associated with graphs since the last time it was reset. High watermark can only be reset to zero.
CU_GRAPH_MEM_ATTR_RESERVED_MEM_HIGH: High watermark of memory, in bytes, currently allocated for use by the CUDA graphs asynchronous allocator.
My intuition would be that any high watermark should only be "settable" to 0, to signify stopping taking the past into consideration and starting measurements from this point on. So - why is there a difference here between the high-watermarks for used memory and for reserved memory?
Related
I'm experimenting with cache blocking. To do that, I implemented 2 convolution based smoothing algorithms. The gaussian kernel I'm using looks like this:
The first algorithm is just the simple double for loop, looping from left to right, top to bottom as shown below.
Image source: (https://people.engr.ncsu.edu/efg/521/f02/common/lectures/notes/lec9.html)
In the second algorithm I tried to play with cache blocking by spliting the loops into chunks, which became something like the following. I used a BLOCK size of 512x512.
Image source: (https://people.engr.ncsu.edu/efg/521/f02/common/lectures/notes/lec9.html)
I'm running the code on a raspberry pi 3B+, which has a Cortex-A53 with 32KB of L1 and 256KB of L2, I believe. I ran the two algorithms with different image sizes (2048x1536, 6000x4000, 12000x8000, 16000x12000. 8bit gray scale images). But across different image sizes, I saw the run time being very similar.
The question is shouldn't the first algorithm experience access latency which the second should not, especially when using large size image (like 12000x8000). Base on the description of cache blocking in this link, when processing data at the end of image rows using the 1st algorithm, the data at the beginning of the rows should have been evicted from the L1 cache. Using 12000x8000 size image as an example, since we are using 5x5 kernel, 5 rows of data is need, which is 12000x5=60KB, already larger than the 32KB L1 size. When we start processing data for a new row, 4 rows of previous data are still needed but they are likely gone in L1 so needs to be re-fetched. But for the second algorithm it shouldn't have this problem because the block size is small. Can anyone please tell me what am I missing?
I also profiled the algorithm using oprofile with the following data:
Algorithm 1
event
count
L1D_CACHE_REFILL
13,933,254
PREFETCH_LINEFILL
13,281,559
Algorithm 2
event
count
L1D_CACHE_REFILL
9,456,369
PREFETCH_LINEFILL
8,725,250
So it looks like the 1st algorithm does have more cache miss compared to the second, reflecting by the L1D_CACHE_REFILL counts. But it also has higher data prefetching rate, which maybe due to the simple behavior of the loop. So is the whole story of cache blocking not taking into account data prefetching?
Conceptually, you're right blocking will reduce cache misses by keeping the input window in cache.
I suspect the main reason you're not seeing a speedup is because the cache is prefetching from all 5 input rows. Your performance counters show more prefetch loads in the unblocked implementation. I suspect many textbook examples are out of date since cache prefetching has kept getting better. Intel's L2 cache can detect and prefetch from up to 16 linear streams about 10 years ago, I think.
Assume the filter takes 5 * 5 cycles. So that would be 20.8 ns = 25 / 1.2GHz on RPI3. The IO cost will be reading a 5 high column of new input pixels. The amortized IO cost will be 5 bytes / 20.8ns = 229 MiB/s, which is much less than the ~2 GiB/s DRAM bandwidth. So in theory, the relatively slow computation combined with prefetching (I'm not certain how effective) means that memory access isn't a bottleneck.
Try increasing the filter height. The cache can only detect and prefetch from a certain # streams. Or try vectorizing the computation so that memory access becomes the bottleneck.
From the cobalt memory docs , we set max_cobalt_cpu_usage limit to 250M, max_cobalt_gpu_usage to 150M, and found the actual max memory used by cobalt is about 370M when play the 4K videos for about 12 hours, it exceeded 250M a lot, so what kind of memory(eg. malloc/new/mmap etc.) is counted into max_cobalt_cpu_usage? How to set the max_cobalt_cpu_usage and keep the actual memory usage within the max_cobalt_cpu_usage?
PS.The memory statistical approach is as follows:
1>Move focus to the cobalt icon, drop the memory cache, record the memory free value(memoryBegin) from /proc/meminfo;
2>Enter cobalt youtube, and keep playing 4K videos for about 12 hours;
3>During playing the videos, record the memory free value(memoryEnd) from /proc/meminfo every 10s(drop the memory cache before record the memory);
4>Find the minimum memoryEnd as memoryEndMin;
5>memoryBegin - memoryEndMin = 370M;
what kind of memory(eg. malloc/new/mmap etc.) is counted into max_cobalt_cpu_usage?
The value returned from SbSystemGetUsedCPUMemory() is used to determine the current cpu memory usage. Please ensure that this returns the expected values.
Note that the cache's won't dynamically re-size to keep within the limits. The initial conditions must be set to account for the maximum memory limit that is required to be set.
You can manually reduce the sizes of your caches (see memory_tuning.md document) or you can use --reduce_cpu_memory_by=Xmb which will allow the AutoSet memory settings to reduce their memory usage.
For more information about memory use, use --memory_tracker as a command line argument.
I see the following when I run Intel VTune on my workload:
Memory Bound 50.8%
I read the Intel doc, which says (Intel doc):
Memory Bound measures a fraction of slots where pipeline could be stalled due to demand load or store instructions. This accounts mainly for incomplete in-flight memory demand loads that coincide with execution starvation in addition to less common cases where stores could imply back-pressure on the pipeline.
Does that mean that roughly half of the instructions in my app are stalled waiting for memory, or is it more subtle than that?
The pipeline slots concept used by VTune is explain e.g. here: https://software.intel.com/en-us/top-down-microarchitecture-analysis-method-win.
In short pipeline slot represents the hardware resources needed to process one uOp. So for 4-wide CPUs (most Intel processors) we can execute 4 Ops each cycle and the total number of slots will be measured as 4 * CPU_CLK_UNHALTED.THREAD by VTune.
The Memory Bound metric is built on CYCLE_ACTIVITY.STALLS_MEM_ANY event which gives you directly stalls due to memory. Taking into account out-of-order. Basically only if CPU is stalled and at the same time it has in-flight loads the counter is incremented. If there are loads in-flight but CPU is kept busy it is not accounted as memory stall.
So Memory Bound metric provides quite accurate estimation on how much the workload is bound by memory performance issues. The value of 50% means that half of the time was wasted waiting for data from memory.
A slot is an execution port of the pipeline. In general in the VTune documentation, a stall could either mean "not retired" or "not dispatched for execution". In this case, it refers to the number of cycles in which zero uops were dispatched.
According to the VTune include configuration files, Memory Bound is calculated as follows:
Memory_Bound = Memory_Bound_Fraction * BackendBound
Memory_Bound_Fraction is basically the fraction of slots mentioned in the documentation. However, according to the top-down method discussed in the optimization manual, the memory bound metric is relative to the backend bound metric. So this is why it is multiplied by BackendBound.
I'll focus on the first term of the formula, Memory_Bound_Fraction. The formula for the second term, BackendBound, is actually complicated.
Memory_Bound_Fraction is calculated as follows:
Memory_Bound_Fraction = (CYCLE_ACTIVITY.STALLS_MEM_ANY + RESOURCE_STALLS.SB) * NUM_OF_PORTS / Backend_Bound_Cycles * NUM_OF_PORTS
NUM_OF_PORTS is the number of execution ports of the microarchitecture of the target CPU. This can be simplified to:
Memory_Bound_Fraction = CYCLE_ACTIVITY.STALLS_MEM_ANY + RESOURCE_STALLS.SB / Backend_Bound_Cycles
CYCLE_ACTIVITY.STALLS_MEM_ANY and RESOURCE_STALLS.SB are performance events. Backend_Bound_Cycles is calculated as follows:
Backend_Bound_Cycles = CYCLE_ACTIVITY.STALLS_TOTAL + UOPS_EXECUTED.CYCLES_GE_1_UOP_EXEC - Few_Uops_Executed_Threshold - Frontend_RS_Empty_Cycles + RESOURCE_STALLS.SB
Few_Uops_Executed_Threshold is either UOPS_EXECUTED.CYCLES_GE_2_UOP_EXEC or UOPS_EXECUTED.CYCLES_GE_3_UOP_EXEC depending on some other metric. Frontend_RS_Empty_Cycles is either RS_EVENTS.EMPTY_CYCLES or zero depending on some metric.
I realize this answer still needs a lot of additional explanation and BackendBound needs to be expanded. But this early edit makes the answer accurate.
I tested the speed of memcpy() noticing the speed drops dramatically at i*4KB. The result is as follow: the Y-axis is the speed(MB/second) and the X-axis is the size of buffer for memcpy(), increasing from 1KB to 2MB. Subfigure 2 and Subfigure 3 detail the part of 1KB-150KB and 1KB-32KB.
Environment：
CPU : Intel(R) Xeon(R) CPU E5620 # 2.40GHz
OS : 2.6.35-22-generic #33-Ubuntu
GCC compiler flags : -O3 -msse4 -DINTEL_SSE4 -Wall -std=c99
I guess it must be related to caches, but I can't find a reason from the following cache-unfriendly cases:
Why is my program slow when looping over exactly 8192 elements?
Why is transposing a matrix of 512x512 much slower than transposing a matrix of 513x513?
Since the performance degradation of these two cases are caused by unfriendly loops which read scattered bytes into the cache, wasting the rest of the space of a cache line.
Here is my code:
void memcpy_speed(unsigned long buf_size, unsigned long iters){
struct timeval start, end;
unsigned char * pbuff_1;
unsigned char * pbuff_2;
pbuff_1 = malloc(buf_size);
pbuff_2 = malloc(buf_size);
gettimeofday(&start, NULL);
for(int i = 0; i < iters; ++i){
memcpy(pbuff_2, pbuff_1, buf_size);
}
gettimeofday(&end, NULL);
printf("%5.3f\n", ((buf_size*iters)/(1.024*1.024))/((end.tv_sec - \
start.tv_sec)*1000*1000+(end.tv_usec - start.tv_usec)));
free(pbuff_1);
free(pbuff_2);
}
UPDATE
Considering suggestions from #usr, #ChrisW and #Leeor, I redid the test more precisely and the graph below shows the results. The buffer size is from 26KB to 38KB, and I tested it every other 64B(26KB, 26KB+64B, 26KB+128B, ......, 38KB). Each test loops 100,000 times in about 0.15 second. The interesting thing is the drop not only occurs exactly in 4KB boundary, but also comes out in 4*i+2 KB, with a much less falling amplitude.
PS
#Leeor offered a way to fill the drop, adding a 2KB dummy buffer between pbuff_1 and pbuff_2. It works, but I am not sure about Leeor's explanation.
Memory is usually organized in 4k pages (although there's also support for larger sizes). The virtual address space your program sees may be contiguous, but it's not necessarily the case in physical memory. The OS, which maintains a mapping of virtual to physical addresses (in the page map) would usually try to keep the physical pages together as well but that's not always possible and they may be fractured (especially on long usage where they may be swapped occasionally).
When your memory stream crosses a 4k page boundary, the CPU needs to stop and go fetch a new translation - if it already saw the page, it may be cached in the TLB, and the access is optimized to be the fastest, but if this is the first access (or if you have too many pages for the TLBs to hold on to), the CPU will have to stall the memory access and start a page walk over the page map entries - that's relatively long as each level is in fact a memory read by itself (on virtual machines it's even longer as each level may need a full pagewalk on the host).
Your memcpy function may have another issue - when first allocating memory, the OS would just build the pages to the pagemap, but mark them as unaccessed and unmodified due to internal optimizations. The first access may not only invoke a page walk, but possibly also an assist telling the OS that the page is going to be used (and stores into, for the target buffer pages), which would take an expensive transition to some OS handler.
In order to eliminate this noise, allocate the buffers once, perform several repetitions of the copy, and calculate the amortized time. That, on the other hand, would give you "warm" performance (i.e. after having the caches warmed up) so you'll see the cache sizes reflect on your graphs. If you want to get a "cold" effect while not suffering from paging latencies, you might want to flush the caches between iteration (just make sure you don't time that)
EDIT
Reread the question, and you seem to be doing a correct measurement. The problem with my explanation is that it should show a gradual increase after 4k*i, since on every such drop you pay the penalty again, but then should enjoy the free ride until the next 4k. It doesn't explain why there are such "spikes" and after them the speed returns to normal.
I think you are facing a similar issue to the critical stride issue linked in your question - when your buffer size is a nice round 4k, both buffers will align to the same sets in the cache and thrash each other. Your L1 is 32k, so it doesn't seem like an issue at first, but assuming the data L1 has 8 ways it's in fact a 4k wrap-around to the same sets, and you have 2*4k blocks with the exact same alignment (assuming the allocation was done contiguously) so they overlap on the same sets. It's enough that the LRU doesn't work exactly as you expect and you'll keep having conflicts.
To check this, i'd try to malloc a dummy buffer between pbuff_1 and pbuff_2, make it 2k large and hope that it breaks the alignment.
EDIT2:
Ok, since this works, it's time to elaborate a little. Say you assign two 4k arrays at ranges 0x1000-0x1fff and 0x2000-0x2fff. set 0 in your L1 will contain the lines at 0x1000 and 0x2000, set 1 will contain 0x1040 and 0x2040, and so on. At these sizes you don't have any issue with thrashing yet, they can all coexist without overflowing the associativity of the cache. However, everytime you perform an iteration you have a load and a store accessing the same set - i'm guessing this may cause a conflict in the HW. Worse - you'll need multiple iteration to copy a single line, meaning that you have a congestion of 8 loads + 8 stores (less if you vectorize, but still a lot), all directed at the same poor set, I'm pretty sure there's are a bunch of collisions hiding there.
I also see that Intel optimization guide has something to say specifically about that (see 3.6.8.2):
4-KByte memory aliasing occurs when the code accesses two different
memory locations with a 4-KByte offset between them. The 4-KByte
aliasing situation can manifest in a memory copy routine where the
addresses of the source buffer and destination buffer maintain a
constant offset and the constant offset happens to be a multiple of
the byte increment from one iteration to the next.
...
loads have to wait until stores have been retired before they can
continue. For example at offset 16, the load of the next iteration is
4-KByte aliased current iteration store, therefore the loop must wait
until the store operation completes, making the entire loop
serialized. The amount of time needed to wait decreases with larger
offset until offset of 96 resolves the issue (as there is no pending
stores by the time of the load with same address).
I expect it's because:
When the block size is a 4KB multiple, then malloc allocates new pages from the O/S.
When the block size is not a 4KB multiple, then malloc allocates a range from its (already allocated) heap.
When the pages are allocated from the O/S then they are 'cold': touching them for the first time is very expensive.
My guess is that, if you do a single memcpy before the first gettimeofday then that will 'warm' the allocated memory and you won't see this problem. Instead of doing an initial memcpy, even writing one byte into each allocated 4KB page might be enough to pre-warm the page.
Usually when I want a performance test like yours I code it as:
// Run in once to pre-warm the cache
runTest();
// Repeat
startTimer();
for (int i = count; i; --i)
runTest();
stopTimer();
// use a larger count if the duration is less than a few seconds
// repeat test 3 times to ensure that results are consistent
Since you are looping many times, I think arguments about pages not being mapped are irrelevant. In my opinion what you are seeing is the effect of hardware prefetcher not willing to cross page boundary in order not to cause (potentially unnecessary) page faults.
The title might be more specific than my actual problem is, although I believe answering this question would solve a more general problem, which is: how to decrease the effect of high latency (~700 cycle) that comes from random (but coalesced) global memory access in GPUs.
In general if one accesses the global memory with coalesced load (eg. I read 128 consecutive bytes), but with very large distance (256KB-64MB) between coalesced accesses, one gets a high TLB (Translation Lookaside Buffer) miss rate. This high TLB miss rate is due to the limited number (~512) and size (~4KB) of the memory pages used in the TLB lookup table.
I suppose the high TLB miss rate because of the fact that virtual memory is used by NVIDIA, the fact that I get high (98%) Global Memory Replay Overhead and low throughput (45GB/s, with a K20c) in the profiler and the fact that partition camping is not an issue since Fermi.
Is it possible to avoid high TLB miss rate somehow? Would 3D texture cache help if I'm accessing a (X x Y x Z) cube coalesced along X dimension and with a X*Y "stride" along the Z dimension?
Any comment on this topic is appreciated.
Constraints: 1) global data can not be reordered/transposed; 2) kernel is communication bound.
You can only avoid TLB misses by changing your memory access pattern. A different layout of your data in memory can help with this.
A 3D texture will not improve your situation, as it trades improved spatial locality in two additional dimensions against reduced spatial locality in the third dimension. Thus you would unnecessarily read data of neighbors along the Y axis.
What you can do however is mitigate the impact of the resulting latency on throughput. In order to hide t = 700 cycles of latency at a global memory bandwidth of b = 250GB/s, you need to have memory transactions for b / t = 175 KB of data in flight at any time (or 12.5 KB for each of the 14 SMX). With a fully loaded memory interface and a high ratio of TLB misses, you will however find that latency gets closer to 2000 cycles, requiring roughly 32 KB of transactions in flight per sm.
As each word of a memory read transaction in flight requires one register where the value will be stored once it arrives, hiding memory latency has to be balances against register pressure. Keeping 32 KB of data in flight requires 8192 registers, or 12.5% of the total registers available on an SMX.
(Note that for above rough estimates I have neglected the difference between KiB and KB).