I have encountered an interesting issue where a PERCPU_ARRAY created on one system with 2 processors creates an array with 2 per-CPU elements and on another system with 2 processors, an array with 128 per-CPU elements. The latter was rather unexpected to me!
The way I discovered this behavior is that a program that allocated an array for the number of CPUs (using get_nprocs_conf(3)) and then read in the PERCPU_ARRAY into it (using bpf_map_lookup_elem()) ended up writing past the end of the array and crashing.
I would like to find out what is the proper way to determine in a program that reads BPF maps the number of elements in a PERCPU_ARRAY used on a system.
Failing that, I think the second best approach is to pick a buffer for reading in that is "large enough." Here, the problem is similar: what is that number and is there way to learn it at runtime?
The question comes from reading the source of bpftool, which figures this out:
unsigned int get_possible_cpus(void)
{
int cpus = libbpf_num_possible_cpus();
if (cpus < 0) {
p_err("Can't get # of possible cpus: %s", strerror(-cpus));
exit(-1);
}
return cpus;
}
int libbpf_num_possible_cpus(void)
{
static const char *fcpu = "/sys/devices/system/cpu/possible";
static int cpus;
int err, n, i, tmp_cpus;
bool *mask;
/* ---8<--- snip */
}
So that's how they do it!
Related
I am using OpenCL (Xcode, Intel GPU), and I am trying to implement a kernel that calculates moving averages and deviations. I want to pass several double arrays of varying lengths to the kernel. Is this possible to implement, or do I need to pad smaller arrays with zeroes so all the arrays are the same size?
I am new to OpenCL and GPGPU, so please forgive my ignorance of any nomenclature.
You can pass to the kernel any buffer, the kernel does not need to use it all.
For example, if your kernel reduces a buffer you can query at run time the amount of work items (items to reduce) using get_global_size(0). And then call the kernel with the proper parameters.
An example (unoptimized):
__kernel reduce_step(__global float* data)
{
int id = get_global_id(0);
int size = get_global_size(0);
int size2 = size/2;
int size2p = (size+1)/2;
if(id<size2) //Only reduce up to size2, the odd element will remain in place
data[id] += data[id+size2p];
}
Then you can call it like this.
void reduce_me(std::vector<cl_float>& data){
size_t size = data.size();
//Copy to a buffer already created, equal or bigger size than data size
// ... TODO, check sizes of buffer or change the buffer set to the kernel args.
queue.enqueueWriteBuffer(buffer,CL_FALSE,0,sizeof(cl_float)*size,data.data());
//Reduce until 1024
while(size > 1024){
queue.enqueueNDRangeKernel(reduce_kernel,cl::NullRange,cl::NDRange(size),cl::NullRange);
size /= 2;
}
//Read out and trim
queue.enqueueReadBuffer(buffer,CL_TRUE,0,sizeof(cl_float)*size,data.data());
data.resize(size);
}
I would like to program in threading building blocks with tasks. But how does one do the debugging in practice?
In general the print method is a solid technique for debugging programs.
In my experience with MPI parallelization, the right way to do logging is that each thread print its debugging information in its own file (say "debug_irank" with irank the rank in the MPI_COMM_WORLD) so that the logical errors can be found.
How can something similar be achieved with TBB? It is not clear how to access the thread number in the thread pool as this is obviously something internal to tbb.
Alternatively, one could add an additional index specifying the rank when a task is generated but this makes the code rather complicated since the whole program has to take care of that.
First, get the program working with 1 thread. To do this, construct a task_scheduler_init as the first thing in main, like this:
#include "tbb/tbb.h"
int main() {
tbb::task_scheduler_init init(1);
...
}
Be sure to compile with the macro TBB_USE_DEBUG set to 1 so that TBB's checking will be enabled.
If the single-threaded version works, but the multi-threaded version does not, consider using Intel Inspector to spot race conditions. Be sure to compile with TBB_USE_THREADING_TOOLS so that Inspector gets enough information.
Otherwise, I usually first start by adding assertions, because the machine can check assertions much faster than I can read log messages. If I am really puzzled about why an assertion is failing, I use printfs and task ids (not thread ids). Easiest way to create a task id is to allocate one by post-incrementing a tbb::atomic<size_t> and storing the result in the task.
If I'm having a really bad day and the printfs are changing program behavior so that the error does not show up, I use "delayed printfs". Stuff the printf arguments in a circular buffer, and run printf on the records later after the failure is detected. Typically for the buffer, I use an array of structs containing the format string and a few word-size values, and make the array size a power of two. Then an atomic increment and mask suffices to allocate slots. E.g., something like this:
const size_t bufSize = 1024;
struct record {
const char* format;
void *arg0, *arg1;
};
tbb::atomic<size_t> head;
record buf[bufSize];
void recf(const char* fmt, void* a, void* b) {
record* r = &buf[head++ & bufSize-1];
r->format = fmt;
r->arg0 = a;
r->arg1 = b;
}
void recf(const char* fmt, int a, int b) {
record* r = &buf[head++ & bufSize-1];
r->format = fmt;
r->arg0 = (void*)a;
r->arg1 = (void*)b;
}
The two recf routines record the format and the values. The casting is somewhat abusive, but on most architectures you can print the record correctly in practice with printf(r->format, r->arg0, r->arg1) even if the the 2nd overload of recf created the record.
~
~
I try to parse many Google protocol buffer messages from a binary file generated by calling SerializeToString. I first load all Bytes into a heap memory by calling new function. I also have two arrays to store the Bytes begin address of a message in the heap memory and the Bytes count of the message.
Then I begin to parse message by calling ParseFromString.I want to quicken the procedure by using multi-thread.
In each thread, I pass the start index and end index of address array and Byte count array.
In parent process. the main code is:
struct ParsePara
{
char* str_buffer;
size_t* buffer_offset;
size_t* binary_string_length_array;
size_t start_idx;
size_t end_idx;
Flight_Ticket_Info* ticket_info_buffer_array;
};
//Flight_Ticket_Info is class of message
//offset_size is the count of message
ticket_array = new Flight_Ticket_Info[offset_size];
const int max_thread_count = 6;
pthread_t pthread_id_vec[max_thread_count];
CTimer thread_cost;
thread_cost.start();
vector<ParsePara*> para_vec;
const size_t each_count = ceil(float(offset_size) / max_thread_count);
for (size_t k = 0;k < max_thread_count;k++)
{
size_t start_idx = each_count * k;
size_t end_idx = each_count * (k+1);
if (start_idx >= offset_size)
break;
if (end_idx >= offset_size)
end_idx = offset_size;
ParsePara* cand_para_ptr = new ParsePara();
if (!cand_para_ptr)
{
_ERROR_EXIT(0,"[Malloc memory fail.]");
}
cand_para_ptr->str_buffer = m_valdata;//heap memory for storing Bytes of message
cand_para_ptr->buffer_offset = offset_array;//begin address of each message
cand_para_ptr->start_idx = start_idx;
cand_para_ptr->end_idx = end_idx;
cand_para_ptr->ticket_info_buffer_array = ticket_array;//array to store message
cand_para_ptr->binary_string_length_array = binary_length_array;//Bytes count of each message
para_vec.push_back(cand_para_ptr);
}
for(size_t k = 0 ;k < para_vec.size();k++)
{
int ret = pthread_create(&pthread_id_vec[k],NULL,parserFlightTicketForMultiThread,para_vec[k]);
if (0 != ret)
{
_ERROR_EXIT(0,"[Error] [create thread fail]");
}
}
for (size_t k = 0;k < para_vec.size();k++)
{
pthread_join(pthread_id_vec[k],NULL);
}
In each thread the thread function is:
void* parserFlightTicketForMultiThread(void* void_para_ptr)
{
ParsePara* para_ptr = (ParsePara*) void_para_ptr;
parserFlightTicketForMany(para_ptr->str_buffer,para_ptr->ticket_info_buffer_array,para_ptr->buffer_offset,
para_ptr->start_idx,para_ptr->end_idx,para_ptr->binary_string_length_array);
}
void parserFlightTicketForMany(const char* str_buffer,Flight_Ticket_Info* ticket_info_buffer_array,
size_t* buffer_offset,const size_t start_idx,const size_t end_idx,size_t* binary_string_length_array)
{
printf("start_idx:%d,end_idx:%d\n",start_idx,end_idx);
for (size_t k = start_idx;k < end_idx;k++)
{
if (k % 100000 == 0)
cout << k << endl;
size_t cand_offset = buffer_offset[k];
size_t binary_length = binary_string_length_array[k];
ticket_info_buffer_array[k].ParseFromString(string(&str_buffer[cand_offset],binary_length-1));
}
printf("done %ld %ld\n",start_idx,end_idx);
}
But multi-thread cost is more than one thread.
one thread cost is:40455623ms
My computer is 8 core and six thread cost is:131586865ms
Anyone can help me? thank you!
Some possible problems -- you'll have to experiment to determine which:
Protobuf parsing speed is often limited by memory bandwidth rather than CPU time, especially with a large input data set. In that case, more threads won't help, since all the cores are sharing bandwidth to main memory. Indeed, having multiple cores fighting over memory bandwidth could make the overall operation slower. Note that the biggest consumer of memory is not the input bytes but rather the parsed data objects -- that is, the output of parsing -- which are many times larger than the encoded data. To improve this problem, consider writing the parsing loop so that it fully-processes each message immediately after parsing, before moving on to the text message. That way, instead of allocating k protobuf objects, you only need to allocate one protobuf object per thread, and repeatedly reuse the same object for parsing. This way the object will (probably) stay in the core's private L1 cache and avoid consuming memory bandwidth; only the input bytes will be read over the main bus.
How are you loading data into RAM? Did you read() into a large array or did you mmap()? In the latter case the data is read from disk lazily -- it won't happen until you actually attempt to parse it. Even in the read() case, it could be that the data has been swapped out, creating similar effects. Either way, your threads are now not just fighting for memory bandwidth, but disk bandwidth, which is of course much slower. Having six threads reading separate parts of a big file will definitely be slower overall than having one thread read the whole file, because the operating system optimizes for sequential access.
Protobuf allocates memory during parsing. Many memory allocators take a lock while allocating new memory. Since all your threads are allocating tons and tons of objects in a tight loop, they will contend for this lock. Make sure you are using a thread-friendly memory allocator, such as Google's tcmalloc. Note that repeatedly reusing the same protobuf object in a parse-consume loop rather than allocating lots of different objects will also help immensely here, because the protobuf object will automatically reuse memory for sub-objects.
There may be a bug in your code and it might not be doing what you expect at all when multithreaded. For example, a bug might be causing all the threads to process the same data, rather than different data, and it could be that the data they're choosing happens to be bigger. Make sure you are testing that the results of your code are exactly the same when you run single-threaded vs. multi-threaded.
In short, if you want multiple cores to make your code faster, you have to think about not just what each core is doing, but what data is going in and out of each core, and how much the cores have to talk to each other. Ideally you want each core to operate all on its own without talking to anyone or anything; then you get maximum parallelism. That's not usually possible, of course, but the closer you can get to that, the better.
BTW, a random optimization for you:
ParseFromString(string(&str_buffer[cand_offset],binary_length-1))
Replace that with:
ParseFromArray(&str_buffer[cand_offset],binary_length-1)
Creating at std::string makes a copy of the data, which wastes time (and memory bandwidth). (This doesn't explain why threading is slow, though.)
I would like calculate the memory usage for single process. So after a little bit of research I came across over smaps and statm.
First of all what is smaps and statm? What is the difference?
statm has a field RSS and in smaps I sum up all RSS values. But those values are different for the same process. I know that statm measures in pages. For comparison purposes I converted that value in kb as in smaps. But those values are not equal.
Why do these two values differ, even though they represent the rss value for the same process?
statm
232214 80703 7168 27 0 161967 0 (measured in pages, pages size is 4096)
smaps
Rss 1956
My aim is to calculate the memory usage for a single process. I am interested in two values. USS and PSS.
Can I gain those two values by just using smaps? Is that value correct?
Also, I would like to return that value as percentage.
I think statm is an approximated simplification of smaps, which is more expensive to get. I came to this conclusion after I looked at the source:
smaps
The information you see in smaps is defined in /fs/proc/task_mmu.c:
static int show_smap(struct seq_file *m, void *v, int is_pid)
{
(...)
struct mm_walk smaps_walk = {
.pmd_entry = smaps_pte_range,
.mm = vma->vm_mm,
.private = &mss,
};
memset(&mss, 0, sizeof mss);
walk_page_vma(vma, &smaps_walk);
show_map_vma(m, vma, is_pid);
seq_printf(m,
(...)
"Rss: %8lu kB\n"
(...)
mss.resident >> 10,
The information in mss is used by walk_page_vma defined in /mm/pagewalk.c. However, the mss member resident is not filled in walk_page_vma - instead, walk_page_vma calls callback specified in smaps_walk:
.pmd_entry = smaps_pte_range,
.private = &mss,
like this:
if (walk->pmd_entry)
err = walk->pmd_entry(pmd, addr, next, walk);
So what does our callback, smaps_pte_range in /fs/proc/task_mmu.c, do?
It calls smaps_pte_entry and smaps_pmd_entry in some circumstances, out of which both call statm_account(), which in turn... upgrades resident size! All of these functions are defined in the already linked task_mmu.c so I didn't post relevant code snippets as they can be easily seen in the linked sources.
PTE stands for Page Table Entry and PMD is Page Middle Directory. So basically we iterate through the page entries associated with given process and update RAM usage depending on the circumstances.
statm
The information you see in statm is defined in /fs/proc/array.c:
int proc_pid_statm(struct seq_file *m, struct pid_namespace *ns,
struct pid *pid, struct task_struct *task)
{
unsigned long size = 0, resident = 0, shared = 0, text = 0, data = 0;
struct mm_struct *mm = get_task_mm(task);
if (mm) {
size = task_statm(mm, &shared, &text, &data, &resident);
mmput(mm);
}
seq_put_decimal_ull(m, 0, size);
seq_put_decimal_ull(m, ' ', resident);
seq_put_decimal_ull(m, ' ', shared);
seq_put_decimal_ull(m, ' ', text);
seq_put_decimal_ull(m, ' ', 0);
seq_put_decimal_ull(m, ' ', data);
seq_put_decimal_ull(m, ' ', 0);
seq_putc(m, '\n');
return 0;
}
This time, resident is filled by task_statm. This one has two implementations, one in /fs/proc/task_mmu.c and second in /fs/proc/task_nomm.c. Since they're almost surely mutually exclusive, I'll focus on the implementation in task_mmu.c (which also contained task_smaps). In this implementation we see that
unsigned long task_statm(struct mm_struct *mm,
unsigned long *shared, unsigned long *text,
unsigned long *data, unsigned long *resident)
{
*shared = get_mm_counter(mm, MM_FILEPAGES);
(...)
*resident = *shared + get_mm_counter(mm, MM_ANONPAGES);
return mm->total_vm;
}
it queries some counters, namely, MM_FILEPAGES and MM_ANONPAGES. These counters are modified during different operations on memory such as do_wp_page defined at /mm/memory.c. All of the modifications seem to be done by the files located in /mm/ and there seem to be quite a lot of them, so I didn't include them here.
Conclusion
smaps does complicated iteration through all referenced memory regions and updates resident size using the collected information. statm uses data that was already calculated by someone else.
The most important part is that while smaps collects the data each time in an independent manner, statm uses counters that get incremented or decremented during process life cycle. There are a lot of places that need to do the bookkeeping, and perhaps some places don't upgrade the counters like they should. That's why IMO statm is inferior to smaps, even if it takes fewer CPU cycles to complete.
Please note that this is the conclusion I drew based on common sense, but I might be wrong - perhaps there are no internal inconsistencies in counter decrementing and incrementing, and instead, they might count some pages differently than smaps. At this point I believe it'd be wise to take it to some experienced kernel maintainers.
I would like to apply a reduce on this piece of my kernel code (1 dimensional data):
__local float sum = 0;
int i;
for(i = 0; i < length; i++)
sum += //some operation depending on i here;
Instead of having just 1 thread that performs this operation, I would like to have n threads (with n = length) and at the end having 1 thread to make the total sum.
In pseudo code, I would like to able to write something like this:
int i = get_global_id(0);
__local float sum = 0;
sum += //some operation depending on i here;
barrier(CLK_LOCAL_MEM_FENCE);
if(i == 0)
res = sum;
Is there a way?
I have a race condition on sum.
To get you started you could do something like the example below (see Scarpino). Here we also take advantage of vector processing by using the OpenCL float4 data type.
Keep in mind that the kernel below returns a number of partial sums: one for each local work group, back to the host. This means that you will have to carry out the final sum by adding up all the partial sums, back on the host. This is because (at least with OpenCL 1.2) there is no barrier function that synchronizes work-items in different work-groups.
If summing the partial sums on the host is undesirable, you can get around this by launching multiple kernels. This introduces some kernel-call overhead, but in some applications the extra penalty is acceptable or insignificant. To do this with the example below you will need to modify your host code to call the kernel repeatedly and then include logic to stop executing the kernel after the number of output vectors falls below the local size (details left to you or check the Scarpino reference).
EDIT: Added extra kernel argument for the output. Added dot product to sum over the float 4 vectors.
__kernel void reduction_vector(__global float4* data,__local float4* partial_sums, __global float* output)
{
int lid = get_local_id(0);
int group_size = get_local_size(0);
partial_sums[lid] = data[get_global_id(0)];
barrier(CLK_LOCAL_MEM_FENCE);
for(int i = group_size/2; i>0; i >>= 1) {
if(lid < i) {
partial_sums[lid] += partial_sums[lid + i];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0) {
output[get_group_id(0)] = dot(partial_sums[0], (float4)(1.0f));
}
}
I know this is a very old post, but from everything I've tried, the answer from Bruce doesn't work, and the one from Adam is inefficient due to both global memory use and kernel execution overhead.
The comment by Jordan on the answer from Bruce is correct that this algorithm breaks down in each iteration where the number of elements is not even. Yet it is essentially the same code as can be found in several search results.
I scratched my head on this for several days, partially hindered by the fact that my language of choice is not C/C++ based, and also it's tricky if not impossible to debug on the GPU. Eventually though, I found an answer which worked.
This is a combination of the answer by Bruce, and that from Adam. It copies the source from global memory into local, but then reduces by folding the top half onto the bottom repeatedly, until there is no data left.
The result is a buffer containing the same number of items as there are work-groups used (so that very large reductions can be broken down), which must be summed by the CPU, or else call from another kernel and do this last step on the GPU.
This part is a little over my head, but I believe, this code also avoids bank switching issues by reading from local memory essentially sequentially. ** Would love confirmation on that from anyone that knows.
Note: The global 'AOffset' parameter can be omitted from the source if your data begins at offset zero. Simply remove it from the kernel prototype and the fourth line of code where it's used as part of an array index...
__kernel void Sum(__global float * A, __global float *output, ulong AOffset, __local float * target ) {
const size_t globalId = get_global_id(0);
const size_t localId = get_local_id(0);
target[localId] = A[globalId+AOffset];
barrier(CLK_LOCAL_MEM_FENCE);
size_t blockSize = get_local_size(0);
size_t halfBlockSize = blockSize / 2;
while (halfBlockSize>0) {
if (localId<halfBlockSize) {
target[localId] += target[localId + halfBlockSize];
if ((halfBlockSize*2)<blockSize) { // uneven block division
if (localId==0) { // when localID==0
target[localId] += target[localId + (blockSize-1)];
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
blockSize = halfBlockSize;
halfBlockSize = blockSize / 2;
}
if (localId==0) {
output[get_group_id(0)] = target[0];
}
}
https://pastebin.com/xN4yQ28N
You can use new work_group_reduce_add() function for sum reduction inside single work group if you have support for OpenCL C 2.0 features
A simple and fast way to reduce data is by repeatedly folding the top half of the data into the bottom half.
For example, please use the following ridiculously simple CL code:
__kernel void foldKernel(__global float *arVal, int offset) {
int gid = get_global_id(0);
arVal[gid] = arVal[gid]+arVal[gid+offset];
}
With the following Java/JOCL host code (or port it to C++ etc):
int t = totalDataSize;
while (t > 1) {
int m = t / 2;
int n = (t + 1) / 2;
clSetKernelArg(kernelFold, 0, Sizeof.cl_mem, Pointer.to(arVal));
clSetKernelArg(kernelFold, 1, Sizeof.cl_int, Pointer.to(new int[]{n}));
cl_event evFold = new cl_event();
clEnqueueNDRangeKernel(commandQueue, kernelFold, 1, null, new long[]{m}, null, 0, null, evFold);
clWaitForEvents(1, new cl_event[]{evFold});
t = n;
}
The host code loops log2(n) times, so it finishes quickly even with huge arrays. The fiddle with "m" and "n" is to handle non-power-of-two arrays.
Easy for OpenCL to parallelize well for any GPU platform (i.e. fast).
Low memory, because it works in place
Works efficiently with non-power-of-two data sizes
Flexible, e.g. you can change kernel to do "min" instead of "+"