510
Intel MIC架構下COI框架介紹
開始介紹之前先寫一下曆史背景，為了最大限度地提高計算速度，單一地提高一個核的主頻以提高計算速率的方法已經不再適用。所以向量機、超標量計算機等紛紛出現，並行計算也再度成為了一個熱門的方向。現有的並行計算架構主要有兩個：GPGPU（通用GPU）以及Intel的MIC（Many Integrated Core）架構。通用GPU加速主要是利用GPU本身具有多線程的特性，將計算密集型任務遷移到GPU上並將計算任務劃分到多個線程內同時進行計算，再將計算結果傳回已達到提高計算速率的效果。主要的用到的技術就是CUDA。但是CUDA有一個很大的缺點就是：要編寫CUDA的代碼工作量很大，通常情況下需要將原有的在CPU下可以運行的代碼做大量的改動。而這一點正是Intel MIC架構的優勢，通常情況下可以在CPU上直接運行的代碼也可以在MIC卡上直接運行。如果想了解更多MIC的相關知識可以持續關注我的博客，我將在接下來對MIC進行更詳細的介紹。
MIC編程主要有native模式、對稱模式以及offload模式。native模式就是程序隻在MIC卡上運行，並沒有將CPU的運行效能全部運行起來；對稱模式就是CPU和MIC卡運行的程序完全相同，可以看成是對等的節點；而offload模式則是以CPU上運行程序為主，但是將一些計算密集型的任務offload到MIC上進行運算。從代碼的角度來看，在MIC上運行的並不是一套完整的代碼，而隻是一些代碼片段。但是使用過offload模式進行大規模數值模式程序編寫的人應該都有體會，offload的方式有這麼兩個缺點：1、數據傳輸會是一個瓶頸，很多時候不是運算不夠快而是數據的傳輸跟不上；2、offload模式通常隻適用於一些較為扁平的數據結構的操作。博主手頭的程序都是大量使用了麵向對象特性的程序，offload模式使用起來很麻煩，對於程序的結構也不能很好的維護。而針對這些問題，本著提高傳輸效率以及增強編碼人員自由度的目的，Intel推出了COI（Coprocessor
Offload Infrastructure）。
-
COI基本概念
COI是MIC架構下的分載模式的一個庫，mic雖然提供了通過簡單的編譯製導語句（#pragma offload target）的方式來將部fen代碼和數據分載到Xeon Phi上進行計算。但這樣的方式太過簡單，且自由度較小。但如果使用COI，用戶可以獲取更大的自由度，包括控製CPU和MIC的同步，控製MIC卡上程序的創建和退出；起止端的異步操作；起止端的數據緩衝。為開發更加靈活的MIC程序提供了便利。
利用COI編寫的程序在編譯的時候就會生成兩個可執行文件。一個在CPU上執行，一個在MIC上執行。啟動程序隻需要調用CPU端的可執行文件即可。係統就會在需要的時候把MIC端可執行文件和所需要的庫發送到MIC端並執行MIC端的程序。接下來開始詳細介紹COI。
術語說明：因為COI可以實現CPU到MIC的分載，同時也可以實現MIC到CPU的分載，所以在下文中使用source和sink來表示起止端。
-
基本概念
Enumeration：COIEngine，COISysinfo
列舉出硬件的信息：如MIC卡的數量，線程數，核數，cache等等。示例代碼如下：
<span >COIENGINE engine;
COIFUNCTION func[1];
const char* SINK_NAME = "coi_simple_sink_mic";
// Make sure there is an Intel(r) Xeon Phi(tm) device available
//
CHECK_RESULT(
COIEngineGetCount(COI_ISA_MIC, &num_engines));
printf("%u engines available\n", num_engines);
// If there isn't at least one engine, there is something wrong
//
if (num_engines < 1)
{
printf("ERROR: Need at least 1 engine\n");
return -1;
}</span>
Process Management:COIProcess
COIProcess是指source端在sink端創建的一個進程在source端創建的一個進程在source端的一個句柄。source端負責該進程的創建，開啟以及銷毀等工作。其主要的作用有：
- 抽象sink端的進程的運行
- 提供開啟和停止遠程進程的各種API已經load動態鏈接庫
- 提供在遠端查詢函數並執行函數的功能
具體示例代碼如下：
COIRESULT result = COI_ERROR;
COIPROCESS proc;
COIENGINE engine;
result = COIProcessCreateFromFile(
engine, // The engine to create the process on.
SINK_NAME, // The local path to the sink side binary to launch.
0, NULL, // argc and argv for the sink process.
false, NULL, // Environment variables to set for the sink process.
true, NULL, // Enable the proxy but don't specify a proxy root path.
0, // The amount of memory to pre-allocate
// and register for use with COIBUFFERs.
NULL, // Path to search for dependencies
&proc // The resulting process handle.
);
if (result != COI_SUCCESS)
{
printf("COIProcessCreateFromFile result %s\n",
COIResultGetName(result));
return -1;
}
printf("Sink process created, press enter to destroy it.\n");
getchar();
// Destroy the process
//
result = COIProcessDestroy(
proc, // Process handle to be destroyed
-1, // Wait indefinitely until main() (on sink side) returns
false, // Don't force to exit. Let it finish executing
// functions enqueued and exit gracefully
&sink_return, // Don't care about the exit result.
&exit_reason
);
if (result != COI_SUCCESS)
{
printf("COIProcessDestroy result %s\n", COIResultGetName(result));
return -1;
}
執行流：COIPipeline
COIPipeline類似於RPC（Remote Procedure Call）的機製，可以講一係列指令序列插入到COIPipeline中，這些指令可以順序地在sink端執行，這裏的指令序列主要是要在遠程調用的函數序列。其主要有以下幾個重要的性質：
- 在COIPipeline中插入的函數會在sink端按序執行。
- COIPipeline就是一種遠程調用的機製。因為可以在遠端調用完整函數，所以有了比單一offload更多的自由度。
- COIPipeline在插入函數時除了插入該函數需要的參數外，還可以傳遞一塊buffer
- 從source到sink端的數據傳輸使用的是SCIF
示例代碼如下：（source端）
CHECK_RESULT(
COIProcessCreateFromFile(
engine, // The engine to create the process on.
SINK_NAME, // The local path to the sink side binary to launch.
0, NULL, // argc and argv for the sink process.
false, NULL, // Environment variables to set for the sink
// process.
true, NULL, // Enable the proxy but don't specify a proxy root
// path.
0, // The amount of memory to pre-allocate
// and register for use with COIBUFFERs.
NULL, // Path to search for dependencies
&proc // The resulting process handle.
));
printf("Created sink process %s\n", SINK_NAME);
// Pipeline:
// After a sink side process is created, multiple pipelines can be created
// to that process. Pipelines are queues where functions(represented by
// COIFUNCTION) to be executed on sink side can be enqueued.
// The following call creates a pipeline associated with process created
// earlier.
CHECK_RESULT(
COIPipelineCreate(
proc, // Process to associate the pipeline with
NULL, // Do not set any sink thread affinity for the pipeline
0, // Use the default stack size for the pipeline thread
&pipeline // Handle to the new pipeline
));
printf("Created pipeline\n");
// Retrieve handle to function belonging to sink side process
const char* func_name = "Foo";
CHECK_RESULT(
COIProcessGetFunctionHandles(
proc, // Process to query for the function
1, // The number of functions to look up
&func_name, // The name of the function to look up
func // A handle to the function
));
printf("Got handle to sink function %s\n", func_name);
const char *misc_data = "Hello COI";
int strlength = (int)strlen(misc_data) + 1;
// Enough to hold the return value
char* return_value = (char*) malloc(strlength);
if (return_value == NULL) {
fprintf(stderr, "failed to allocate return value\n");
return -1;
}
// Enqueue the function for execution
// Pass in misc_data and return value pointer to run function
// Get an event to wait on until the run function completion
CHECK_RESULT(
COIPipelineRunFunction(
pipeline, func[0], // Pipeline handle and function handle
0, NULL, NULL, // Buffers and access flags
0, NULL, // Input dependencies
misc_data, strlength, // Misc Data to pass to the function
return_value, strlength, // Return values that will be passed back
&completion_event // Event to signal when it completes
));
printf("Called sink function %s(\"%s\" [%d bytes])\n",
func_name, misc_data, strlength);
// Now wait indefinitely for the function to complete
CHECK_RESULT(
COIEventWait(
1, // Number of events to wait for
&completion_event, // Event handles
-1, // Wait indefinitely
true, // Wait for all events
NULL, NULL // Number of events signaled
// and their indices
));
printf("Function returned \"%s\"\n", return_value);
sink端：
// main is automatically called whenever the source creates a process.
// However, once main exits, the process that was created exits.
int main(int argc, char** argv)
{
UNUSED_ATTR COIRESULT result;
UNREFERENCED_PARAM (argc);
UNREFERENCED_PARAM (argv);
// Functions enqueued on the sink side will not start executing until
// you call COIPipelineStartExecutingRunFunctions(). This call is to
// synchronize any initialization required on the sink side
result = COIPipelineStartExecutingRunFunctions();
assert(result == COI_SUCCESS);
// This call will wait until COIProcessDestroy() gets called on the source
// side. If COIProcessDestroy is called without force flag set, this call
// will make sure all the functions enqueued are executed and does all
// clean up required to exit gracefully.
COIProcessWaitForShutdown();
return 0;
}
// Prototype of run function that can be retrieved on the source side.
// Copies misc data to return pointer.
COINATIVELIBEXPORT
void Foo (uint32_t in_BufferCount,
void** in_ppBufferPointers,
uint64_t* in_pBufferLengths,
void* in_pMiscData,
uint16_t in_MiscDataLength,
void* in_pReturnValue,
uint16_t in_ReturnValueLength)
{
UNREFERENCED_PARAM(in_BufferCount);
UNREFERENCED_PARAM(in_ppBufferPointers);
UNREFERENCED_PARAM(in_pBufferLengths);
UNREFERENCED_PARAM(in_pMiscData);
UNREFERENCED_PARAM(in_MiscDataLength);
assert (in_MiscDataLength>=in_ReturnValueLength);
if(in_pMiscData!=NULL && in_pReturnValue!=NULL)
{
memcpy(in_pReturnValue, in_pMiscData, in_ReturnValueLength);
}
}
source端的COIEventWait暫時可以不關心，接下來會詳細介紹。通過上述代碼可以發現，在source端，將對應的函數插入COIPipeline隊列之後，該函數將會在sink端執行。但前提是該函數必須在sink端的代碼中被申明。不過被插入到在COIPipeline中的函數並不是會立即在sink端執行。首先是要在sink端調用COIPipelineStartExecutingRunFunctions（）之後才能被執行。另外如果被插入的執行函數有相關的buffer，那麼函數也必須在buffer可用的時候才能執行，此外通過COIEvent也可以來控製函數的執行。更多的內容會在後麵講到。
COIBuffer
COIBuffer用於管理在遠程設備上的數據。
Buffer可以通過傳入執行函數給遠程端點，也可以直接使用讀寫API。
COI的runtime（運行時環境）來管理buffer
- Buffer用於在source和sink端的數據傳輸，buffer可以使source和sink端的讀寫異步，隱藏掉通信延遲。
- Buffer可能是位於device也可能是位於host端的物理內存中。
- Buffer實際是利用SCIF的內存窗口實現的
- 數據的傳輸實際利用的是readfrom/writeto API
- map操作可以訪問到buffer對應的區域，而不需要將其數據移到host上
COIBuffer的操作比較複雜，更多的細節會在後續的博客中涉及到。
COIEvent
COIEvent可以用來創建依賴關係，從而使source和sink端進行同步，可以通過創建事件然後等待事件來達到同步，有點類似於MPI中的MPI_Barrier函數。一個函數的執行可以有一個先導COIEvent，隻有當該event被消費之後，位於COIPipeline中的對應函數才能被執行（該函數此時必須位於COIPipeline隊列的首位）。同時，當該函數被執行結束之後，也可以產生一個事件，遠端的程序可以通過等待對應的時候來進行同步操作。具體看示例代碼：
source端：
// This tutorial demonstrates:
// 1. Registering a User event
// 2. Pass the event to the sink side
// 3. Signaling the event from the sink side and using it
// to synchronize on the source side.
// It first enqueues a run function with a registered user event (which
// is not signaled) as input dependency. Then a second function is enqueued
// on a different pipeline that signals the user event.
// User events are one shot events. Once they are signaled they
// can't be signaled again. You have to register them again to enable
// signaling.
</span><span ><span >...
//Create two pipelines
CHECK_RESULT(
COIPipelineCreate(
proc, // Process to associate the pipeline with
NULL, // Do not set any sink thread affinity for the pipeline
0, // Use the default stack size for the pipeline thread
&pipeline[0] // Handle to the new pipeline
));
CHECK_RESULT(
COIPipelineCreate(
proc, // Process to associate the pipeline with
NULL, // Do not set any sink thread affinity for the pipeline
0, // Use the default stack size for the pipeline thread
&pipeline[1] // Handle to the new pipeline
));
printf("Created sink process %s and two pipelines\n", SINK_NAME);
// Retrieve handle to functions belonging to sink side process
const char* names[] = {"Return2","SignalUserEvent"};
CHECK_RESULT(
COIProcessGetFunctionHandles(
proc, // Process to query for the function
2, // The number of functions to query
names, // The name of the function
func // A handle to the function
));
printf("Got handles to functions %s and %s\n", names[0], names[1]);
uint64_t return_value = 0;
COIEVENT user_event;
// Register this event so that it can be signaled
CHECK_RESULT(
COIEventRegisterUserEvent(&user_event));
printf("Registered user event\n");
// Now pass this registered user event as an input dependency to the run
// function. This run function will not be started until the user event
// is signaled.
CHECK_RESULT(
COIPipelineRunFunction(
pipeline[0], func[0], // Pipeline handle and function
// handle
0, NULL, NULL, // Buffers and access flags to
// pass to the function
1, &user_event, // Input dependencies
NULL, 0, // Misc data to pass to
// the function
&return_value, sizeof(return_value), // Return value passed back
// from the function
&completion_event // Event to signal when
// the function is complete
));
printf("Enqueued sink function %s depending on user event\n", names[0]);
// Sleep for 2 sec which is enough for run function to be started on sink
// side
#ifndef _WIN32
sleep(2);
#else
Sleep(2000);
#endif
// Now try waiting for the completion_event. It should return
// COI_TIME_OUT_REACHED (as the event isn't signaled)
if(COIEventWait(1, &completion_event, 0, true, NULL, NULL) !=
COI_TIME_OUT_REACHED)
{
printf("Error: Did not execute as expected\n");
return -1;
}
printf("As expected, event wait timed out\n");
// User event handles can be passed down to run function as misc
// data (or via buffers) and on sink side can be type-casted back to
// COIEVENT object to signal them.
CHECK_RESULT(
COIPipelineRunFunction(
pipeline[1], func[1], // Pipeline handle and function
// handle
0, NULL, NULL, // Buffers and access flags to
// pass to the function
0, NULL, // Input dependencies
&user_event, sizeof(user_event), // Misc data to pass to
// the function
NULL,0, // Return value passed back
// from the function
NULL // Event to signal when
// the function is complete
));
printf("Enqueued sink function %s passing user event as misc arg\n",
names[1]);
// Wait until the user event is signaled
CHECK_RESULT(
COIEventWait(
1, // Number of events to wait for
&user_event, // Event handles
-1, // Wait indefinitely
true, // Wait for all events
NULL, NULL // Number of events signaled
// and their indices
));
printf("Successfully waited for user event (signaled sink side)\n");
// Once a user event is signaled the first run function will be able to
// proceed. Wait until the function finishes (-1 wait indefinite)
CHECK_RESULT(
COIEventWait(
1, // Number of events to wait for
&completion_event, // Event handles
-1, // Wait indefinitely
true, // Wait for all events
NULL, NULL // Number of events signaled
// and their indices
));
printf("Sink function %s completed since user event signaled\n", names[0]);
// Unregister the event to cleanup
CHECK_RESULT(
COIEventUnregisterUserEvent(user_event));
MIC端：
// main is automatically called whenever the source creates a process.
// However, once main exits, the process that was created exits.
int main(int argc, char** argv)
{
UNUSED_ATTR COIRESULT result;
UNREFERENCED_PARAM (argc);
UNREFERENCED_PARAM (argv);
// Functions enqueued on the sink side will not start executing until
// you call COIPipelineStartExecutingRunFunctions()
// This call is to synchronize any initialization required on the sink side
result = COIPipelineStartExecutingRunFunctions();
assert(result == COI_SUCCESS);
// This call will wait until COIProcessDestroy() gets called on the source
// side. If COIProcessDestroy is called without force flag set, this call
// will make sure all the functions enqueued are executed and does all
// clean up required to exit gracefully.
COIProcessWaitForShutdown();
return 0;
}
// Prototype of run functions that can be retrieved on the sink side
// This Function just returns 2
COINATIVELIBEXPORT
void Return2(uint32_t in_BufferCount,
void** in_ppBufferPointers,
uint64_t* in_pBufferLengths,
void* in_pMiscData,
uint16_t in_MiscDataLength,
void* in_pReturnValue,
uint16_t in_ReturnValueLength)
{
UNREFERENCED_PARAM(in_BufferCount);
UNREFERENCED_PARAM(in_ppBufferPointers);
UNREFERENCED_PARAM(in_pBufferLengths);
UNREFERENCED_PARAM(in_pMiscData);
UNREFERENCED_PARAM(in_MiscDataLength);
if (sizeof(uint64_t) <= in_ReturnValueLength)
{
*(uint64_t*)(in_pReturnValue) = 2;
}
}
//Assumes a user_event is passed as Misc_data and signals it
COINATIVELIBEXPORT
void SignalUserEvent(uint32_t in_BufferCount,
void** in_ppBufferPointers,
uint64_t* in_pBufferLengths,
void* in_pMiscData,
uint16_t in_MiscDataLength,
void* in_pReturnValue,
uint16_t in_ReturnValueLength)
{
UNREFERENCED_PARAM(in_BufferCount);
UNREFERENCED_PARAM(in_ppBufferPointers);
UNREFERENCED_PARAM(in_pBufferLengths);
UNREFERENCED_PARAM(in_MiscDataLength);
UNREFERENCED_PARAM(in_pReturnValue);
UNREFERENCED_PARAM(in_ReturnValueLength);
COIEVENT user_event;
assert(in_pMiscData != NULL);
assert(in_MiscDataLength >= sizeof(user_event));
memcpy(&user_event, in_pMiscData, sizeof(user_event));
COIEventSignalUserEvent(user_event);
}
可以看到，雖然在source端的代碼中func[0]被先於func1[1]插入COIPipeline中，但是func[0]存在一個輸入依賴——user_event,而user_event被當作func[1]的參數傳輸到sink端，而隻有當user_event在sink端被消費（singnaled）之後，func[0]方可以執行。
最後更新：2017-04-03 05:40:13