Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Refactoring: Handle cache line size consistently (Addresses issue #245) #253

Merged
merged 7 commits into from
Oct 12, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 1 addition & 5 deletions src/libraries/JANA/Engine/JMailbox.h
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@

#include <queue>
#include <mutex>
#include <JANA/Utils/JCpuInfo.h>
#include <JANA/Services/JLoggingService.h>

/// JMailbox is a threadsafe event queue designed for communication between Arrows.
Expand All @@ -32,11 +33,6 @@
/// 3. Triple mutex trick to give push() priority?


#ifndef CACHE_LINE_BYTES
#define CACHE_LINE_BYTES 64
#endif


template <typename T>
class JMailbox {

Expand Down
8 changes: 2 additions & 6 deletions src/libraries/JANA/Engine/JSubeventMailbox.h
Original file line number Diff line number Diff line change
Expand Up @@ -10,11 +10,7 @@
#include <mutex>
#include <JANA/Services/JLoggingService.h>
#include <JANA/JEvent.h>

#ifndef CACHE_LINE_BYTES
#define CACHE_LINE_BYTES 64
#endif

#include <JANA/Utils/JCpuInfo.h>

template <typename SubeventT>
struct SubeventWrapper {
Expand All @@ -37,7 +33,7 @@ class JSubeventMailbox {

private:

struct alignas(CACHE_LINE_BYTES) LocalMailbox {
struct alignas(JANA2_CACHE_LINE_BYTES) LocalMailbox {
std::mutex mutex;
std::deque<std::shared_ptr<JEvent>> ready;
std::map<std::shared_ptr<JEvent>, size_t> in_progress;
Expand Down
4 changes: 2 additions & 2 deletions src/libraries/JANA/JApplication.h
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@
#include <vector>
#include <string>
#include <iostream>
#include <atomic>

class JApplication;
class JEventProcessor;
Expand All @@ -25,12 +26,11 @@ extern JApplication* japp;
#include <JANA/Services/JParameterManager.h>
#include <JANA/Services/JLoggingService.h>
#include <JANA/Status/JComponentSummary.h>
#include <JANA/Utils/JResourcePool.h>
#include <JANA/Status/JPerfSummary.h>


//////////////////////////////////////////////////////////////////////////////////////////////////
/// JANA application class (singleton).
/// JANA application class
///
/// The JApplication class serves as a central access point for getting to most things
/// in the JANA application. It owns the JThreadManager, JParameterManager, etc.
Expand Down
14 changes: 12 additions & 2 deletions src/libraries/JANA/Utils/JCpuInfo.h
Original file line number Diff line number Diff line change
@@ -1,9 +1,19 @@

// Copyright 2020, Jefferson Science Associates, LLC.
// Subject to the terms in the LICENSE file found in the top-level directory.

#pragma once

#ifndef JANA2_CACHE_LINE_BYTES
#define JANA2_CACHE_LINE_BYTES 64
#endif
// The cache line size is 64 for ifarm1402. gcc won't allow larger than 128
// You can find the cache line size in /sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size
/*! The cache line size is useful for creating a buffer to make sure that a variable accessed by multiple threads does not share the cache line it is on.
This is useful for variables that may be written-to by one of the threads, because the thread will acquire locked access to the entire cache line.
This blocks other threads from operating on the other data stored on the cache line. Note that it is also important to align the shared data as well.
See http://www.drdobbs.com/parallel/eliminate-false-sharing/217500206?pgno=4 for more details. */


#include <thread>

namespace JCpuInfo {
Expand All @@ -20,4 +30,4 @@ namespace JCpuInfo {

bool PinThreadToCpu(std::thread *thread, size_t cpu_id);

}
}
4 changes: 2 additions & 2 deletions src/libraries/JANA/Utils/JEventPool.h
Original file line number Diff line number Diff line change
Expand Up @@ -7,13 +7,13 @@
#define JANA2_JEVENTPOOL_H

#include <JANA/JEvent.h>
#include <JANA/Utils/JCpuInfo.h>
#include <JANA/JFactoryGenerator.h>
#include <JANA/Services/JComponentManager.h>

class JEventPool {
private:

struct alignas(64) LocalPool {
struct alignas(JANA2_CACHE_LINE_BYTES) LocalPool {
std::mutex mutex;
std::vector<std::shared_ptr<JEvent>> events;
};
Expand Down
27 changes: 8 additions & 19 deletions src/libraries/JANA/Utils/JResourcePool.h
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
#include <iostream>
#include <type_traits>
#include <memory>
#include <JANA/Utils/JCpuInfo.h>
#include <algorithm>
#include <iterator>

Expand Down Expand Up @@ -88,18 +89,6 @@ template <typename DType> class JResourcePool
std::size_t Get_PoolSize(void) const;
std::size_t Get_NumObjectsAllThreads(void) const{return dObjectCounter;}

static constexpr unsigned int Get_CacheLineSize(void)
{
/// Returns the cache line size for the processor of the target platform.
/*! The cache line size is useful for creating a buffer to make sure that a variable accessed by multiple threads does not share the cache line it is on.
This is useful for variables that may be written-to by one of the threads, because the thread will acquire locked access to the entire cache line.
This blocks other threads from operating on the other data stored on the cache line. Note that it is also important to align the shared data as well.
See http://www.drdobbs.com/parallel/eliminate-false-sharing/217500206?pgno=4 for more details. */

//cache line size is 64 for ifarm1402, gcc won't allow larger than 128
//the cache line size is in /sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size
return 64; //units are in bytes
}

private:

Expand All @@ -119,10 +108,10 @@ template <typename DType> class JResourcePool
void Recycle_Resources_StaticPool(std::vector<DType*>& sResources);
void Recycle_Resource_StaticPool(DType* sResource);

alignas(Get_CacheLineSize()) std::size_t dDebugLevel = 0;
alignas(JANA2_CACHE_LINE_BYTES) std::size_t dDebugLevel = 0;

//static class members have external linkage: same instance shared between every translation unit (would be globally, put only private access)
alignas(Get_CacheLineSize()) static std::atomic<bool> dPoolLock;
alignas(JANA2_CACHE_LINE_BYTES) static std::atomic<bool> dPoolLock;
alignas(Get_CacheLineSize()) static std::vector<DType*> mResourcePool;
alignas(Get_CacheLineSize()) static std::size_t dMaxPoolSize;
alignas(Get_CacheLineSize()) static std::size_t dPoolCounter; //must be accessed within a lock due to how it's used in destructor: freeing all resources
Expand All @@ -147,11 +136,11 @@ template <typename DType> class JSharedPtrRecycler

//STATIC MEMBER DEFINITIONS
//Since these are part of a template, these statics will only be defined once, no matter how much this header is included
template <typename DType> std::atomic<bool> JResourcePool<DType>::dPoolLock{0};
template <typename DType> std::vector<DType*> JResourcePool<DType>::mResourcePool = {};
template <typename DType> std::size_t JResourcePool<DType>::dMaxPoolSize{10};
template <typename DType> std::size_t JResourcePool<DType>::dPoolCounter{0};
template <typename DType> std::atomic<std::size_t> JResourcePool<DType>::dObjectCounter{0};
template <typename DType> alignas(JANA2_CACHE_LINE_BYTES) std::atomic<bool> JResourcePool<DType>::dPoolLock{0};
template <typename DType> alignas(JANA2_CACHE_LINE_BYTES) std::vector<DType*> JResourcePool<DType>::mResourcePool = {};
template <typename DType> alignas(JANA2_CACHE_LINE_BYTES) std::size_t JResourcePool<DType>::dMaxPoolSize{10};
template <typename DType> alignas(JANA2_CACHE_LINE_BYTES) std::size_t JResourcePool<DType>::dPoolCounter{0};
template <typename DType> alignas(JANA2_CACHE_LINE_BYTES) std::atomic<std::size_t> JResourcePool<DType>::dObjectCounter{0};

//CONSTRUCTORS
template <typename DType> JResourcePool<DType>::JResourcePool(std::size_t sMaxPoolSize, std::size_t sDebugLevel) : JResourcePool()
Expand Down