Commit 874b8980 authored by neverlord's avatar neverlord

benchmarking

parent 594858b8
...@@ -2,3 +2,6 @@ ...@@ -2,3 +2,6 @@
*.o *.o
*.dylib *.dylib
test test
*.png
*.dat
queue_test
all: all:
make -f libcppa.Makefile make -f libcppa.Makefile
make -C unit_testing make -C unit_testing
make -C queue_performances
clean: clean:
make -f libcppa.Makefile clean make -f libcppa.Makefile clean
make -C unit_testing clean make -C unit_testing clean
make -C queue_performances clean
...@@ -98,3 +98,10 @@ cppa/scheduling_hint.hpp ...@@ -98,3 +98,10 @@ cppa/scheduling_hint.hpp
src/actor_behavior.cpp src/actor_behavior.cpp
cppa/group.hpp cppa/group.hpp
src/group.cpp src/group.cpp
queue_performances/main.cpp
queue_performances/sutter_list.hpp
queue_performances/defines.hpp
queue_performances/cached_stack.hpp
queue_performances/blocking_cached_stack.hpp
queue_performances/blocking_cached_stack2.hpp
queue_performances/blocking_sutter_list.hpp
include ../Makefile.rules
#CXX = /opt/local/bin/g++-mp-4.5
#CXX = /opt/local/bin/g++-mp-4.6
#CXXFLAGS = -std=c++0x -pedantic -Wall -Wextra -g -O0 -I/opt/local/include/
#CXXFLAGS = -std=c++0x -pedantic -Wall -Wextra -O2 -I/opt/local/include/
#LIBS = -L/opt/local/lib -lboost_thread-mt -L../ -lcppa
INCLUDES = -I./ -I../
FLAGS = -DCACHE_LINE_SIZE=64
EXECUTABLE = ../queue_test
HEADERS = sutter_list.hpp
SOURCES = main.cpp
OBJECTS = $(SOURCES:.cpp=.o)
%.o : %.cpp $(HEADERS) $(HEADERS)
$(CXX) $(CXXFLAGS) $(INCLUDES) $(FLAGS) -c $< -o $@
$(EXECUTABLE) : $(OBJECTS) $(HEADERS)
$(CXX) $(LIBS) -L../ -lcppa $(OBJECTS) -o $(EXECUTABLE)
all : $(EXECUTABLE)
clean:
rm -f $(OBJECTS) $(EXECUTABLE)
#ifndef BLOCKING_CACHED_STACK_HPP
#define BLOCKING_CACHED_STACK_HPP
#include <atomic>
#include <boost/thread.hpp>
#include "defines.hpp"
// This class is intrusive.
template<typename T>
class blocking_cached_stack
{
// singly linked list, serves as cache
T* m_head;
char m_pad1[CACHE_LINE_SIZE - sizeof(T*)];
// modified by consumers
std::atomic<T*> m_stack;
char m_pad2[CACHE_LINE_SIZE - sizeof(std::atomic<T*>)];
// locked on enqueue/dequeue operations to/from an empty list
boost::mutex m_mtx;
boost::condition_variable m_cv;
typedef boost::unique_lock<boost::mutex> lock_type;
// read all elements of m_stack, convert them to FIFO order and store
// them in m_head
// precondition: m_head == nullptr
bool consume_stack()
{
T* e = m_stack.load();
while (e)
{
if (m_stack.compare_exchange_weak(e, 0))
{
// m_stack is now empty (m_stack == nullptr)
while (e)
{
T* next = e->next;
// enqueue to m_head
e->next = m_head;
m_head = e;
// next iteration
e = next;
}
return true;
}
}
// nothing to consume
return false;
}
void wait_for_data()
{
if (!m_head && !(m_stack.load()))
{
lock_type lock(m_mtx);
while (!(m_stack.load())) m_cv.wait(lock);
}
}
public:
blocking_cached_stack() : m_head(0)
{
m_stack = 0;
}
~blocking_cached_stack()
{
do
{
while (m_head)
{
T* next = m_head->next;
delete m_head;
m_head = next;
}
}
// repeat if m_stack is not empty
while (consume_stack());
}
void push(T* what)
{
T* e = m_stack.load();
for (;;)
{
what->next = e;
if (!e)
{
lock_type lock(m_mtx);
if (m_stack.compare_exchange_weak(e, what))
{
m_cv.notify_one();
return;
}
}
// compare_exchange_weak stores the
// new value to e if the operation fails
else if (m_stack.compare_exchange_weak(e, what)) return;
}
}
T* try_pop()
{
if (m_head || consume_stack())
{
T* result = m_head;
m_head = m_head->next;
return result;
}
return 0;
}
T* pop()
{
wait_for_data();
return try_pop();
}
};
#endif // BLOCKING_CACHED_STACK_HPP
#ifndef BLOCKING_CACHED_STACK2_HPP
#define BLOCKING_CACHED_STACK2_HPP
#include <atomic>
#include <boost/thread.hpp>
#include "defines.hpp"
// This class is intrusive.
template<typename T>
class blocking_cached_stack2
{
// singly linked list, serves as cache
T* m_head;
char m_pad1[CACHE_LINE_SIZE - sizeof(T*)];
// modified by consumers
std::atomic<T*> m_stack;
char m_pad2[CACHE_LINE_SIZE - sizeof(std::atomic<T*>)];
T* m_dummy;
char m_pad3[CACHE_LINE_SIZE - sizeof(T)];
// locked on enqueue/dequeue operations to/from an empty list
boost::mutex m_mtx;
boost::condition_variable m_cv;
typedef boost::unique_lock<boost::mutex> lock_type;
// read all elements of m_stack, convert them to FIFO order and store
// them in m_head
// precondition: m_head == nullptr
void consume_stack()
{
T* e = m_stack.load();
while (e)
{
// enqueue dummy instead of nullptr to reduce
// lock operations
if (m_stack.compare_exchange_weak(e, m_dummy))
{
// m_stack is now empty (m_stack == m_dummy)
// m_dummy marks always the end of the stack
while (e && e != m_dummy)
{
T* next = e->next;
// enqueue to m_head
e->next = m_head;
m_head = e;
// next iteration
e = next;
}
return;
}
}
// nothing to consume
}
void wait_for_data()
{
if (!m_head)
{
T* e = m_stack.load();
while (e == m_dummy)
{
if (m_stack.compare_exchange_weak(e, 0)) e = 0;
}
if (!e)
{
lock_type lock(m_mtx);
while (!(m_stack.load())) m_cv.wait(lock);
}
consume_stack();
}
}
void delete_head()
{
while (m_head)
{
T* next = m_head->next;
delete m_head;
m_head = next;
}
}
public:
blocking_cached_stack2() : m_head(0)
{
m_stack = 0;
m_dummy = new T;
}
~blocking_cached_stack2()
{
delete_head();
T* e = m_stack.load();
if (e && e != m_dummy)
{
consume_stack();
delete_head();
}
delete m_dummy;
}
void push(T* what)
{
T* e = m_stack.load();
for (;;)
{
what->next = e;
if (!e)
{
lock_type lock(m_mtx);
if (m_stack.compare_exchange_weak(e, what))
{
m_cv.notify_one();
return;
}
}
// compare_exchange_weak stores the
// new value to e if the operation fails
else if (m_stack.compare_exchange_weak(e, what)) return;
}
}
T* pop()
{
wait_for_data();
T* result = m_head;
m_head = m_head->next;
return result;
}
};
#endif // BLOCKING_CACHED_STACK2_HPP
#ifndef BLOCKING_SUTTER_LIST_HPP
#define BLOCKING_SUTTER_LIST_HPP
#include <atomic>
#include <boost/thread.hpp>
#include "defines.hpp"
// This implementation is a single-consumer version
// of an example implementation of Herb Sutter:
// http://drdobbs.com/cpp/211601363.
// T is any type
template<typename T>
class blocking_sutter_list
{
struct node
{
node(T* val = 0) : value(val), next(0) { }
T* value;
std::atomic<node*> next;
char pad[CACHE_LINE_SIZE - sizeof(T*)- sizeof(std::atomic<node*>)];
};
// one consumer at a time
node* m_first;
char m_pad1[CACHE_LINE_SIZE - sizeof(node*)];
// for one producers at a time
node* m_last;
char m_pad2[CACHE_LINE_SIZE - sizeof(node*)];
// shared among producers
std::atomic<bool> m_producer_lock;
char m_pad3[CACHE_LINE_SIZE - sizeof(std::atomic<bool>)];
// locked on enqueue/dequeue operations to/from an empty list
boost::mutex m_mtx;
boost::condition_variable m_cv;
typedef boost::unique_lock<boost::mutex> lock_type;
public:
blocking_sutter_list()
{
m_first = m_last = new node;
m_producer_lock = false;
}
~blocking_sutter_list()
{
while (m_first)
{
node* tmp = m_first;
m_first = tmp->next;
delete tmp;
}
}
// takes ownership of what
void push(T* what)
{
bool consumer_might_sleep = 0;
node* tmp = new node(what);
// acquire exclusivity
while (m_producer_lock.exchange(true))
{
boost::this_thread::yield();
}
// do we have to wakeup a sleeping consumer?
// this is a sufficient condition because m_last->value is 0
// if and only if m_head == m_tail
consumer_might_sleep = (m_last->value == 0);
// publish & swing last forward
m_last->next = tmp;
m_last = tmp;
// release exclusivity
m_producer_lock = false;
// wakeup consumer if needed
if (consumer_might_sleep)
{
lock_type lock(m_mtx);
m_cv.notify_one();
}
}
// polls the queue until an element was dequeued
T* pop()
{
node* first = m_first;
node* next = m_first->next;
if (!next)
{
lock_type lock(m_mtx);
while (!(next = m_first->next))
{
m_cv.wait(lock);
}
}
T* result = next->value; // take it out
next->value = 0; // of the node
// swing first forward
m_first = next;
// delete old dummy
delete first;
// done
return result;
}
};
#endif // BLOCKING_SUTTER_LIST_HPP
#ifndef CACHED_STACK_HPP
#define CACHED_STACK_HPP
#include <atomic>
#include <boost/thread.hpp>
#include "defines.hpp"
// This class is intrusive.
template<typename T>
class cached_stack
{
// singly linked list, serves as cache
T* m_head;
char m_pad1[CACHE_LINE_SIZE - sizeof(T*)];
// modified by consumers
std::atomic<T*> m_stack;
// read all elements of m_stack, convert them to FIFO order and store
// them in m_head
// precondition: m_head == nullptr
bool consume_stack()
{
T* e = m_stack.load();
while (e)
{
if (m_stack.compare_exchange_weak(e, 0))
{
// m_stack is now empty (m_stack == nullptr)
while (e)
{
T* next = e->next;
// enqueue to m_head
e->next = m_head;
m_head = e;
// next iteration
e = next;
}
return true;
}
}
// nothing to consume
return false;
}
public:
cached_stack() : m_head(0)
{
m_stack = 0;
}
~cached_stack()
{
do
{
while (m_head)
{
T* next = m_head->next;
delete m_head;
m_head = next;
}
}
// repeat if m_stack is not empty
while (consume_stack());
}
void push(T* what)
{
T* e = m_stack.load();
for (;;)
{
what->next = e;
// compare_exchange_weak stores the
// new value to e if the operation fails
if (m_stack.compare_exchange_weak(e, what)) return;
}
}
T* try_pop()
{
if (m_head || consume_stack())
{
T* result = m_head;
m_head = m_head->next;
return result;
}
return 0;
}
T* pop()
{
T* result = try_pop();
while (!result)
{
boost::this_thread::yield();
result = try_pop();
}
return result;
}
};
#endif // CACHED_STACK_HPP
#ifndef DEFINES_HPP
#define DEFINES_HPP
#ifndef CACHE_LINE_SIZE
// the cache line size of an i7 under Mac OS X
#define CACHE_LINE_SIZE 64
#endif
#endif // DEFINES_HPP
#include <vector>
#include <cstddef>
#include <sstream>
#include <iostream>
#include <stdexcept>
#include <boost/thread.hpp>
#include <boost/progress.hpp>
#include <boost/lexical_cast.hpp>
#include "sutter_list.hpp"
#include "cached_stack.hpp"
#include "blocking_sutter_list.hpp"
#include "blocking_cached_stack.hpp"
#include "blocking_cached_stack2.hpp"
using std::cout;
using std::cerr;
using std::endl;
namespace {
//const size_t num_messages = 1000000;
//const size_t num_producers = 10;
//const size_t num_messages_per_producer = num_messages / num_producers;
} // namespace <anonymous>
template<typename Queue, typename Allocator>
void producer(Queue& q, Allocator& a, size_t begin, size_t end)
{
for ( ; begin != end; ++begin)
{
q.push(a(begin));
}
}
template<typename Queue, typename Processor>
void consumer(Queue& q, Processor& p, size_t num_messages)
{
// vector<bool> scales better (in memory) than bool[num_messages]
std::vector<bool> received(num_messages);
for (size_t i = 0; i < num_messages; ++i) received[i] = false;
for (size_t i = 0; i < num_messages; ++i)
{
size_t value;
p(q.pop(), value);
if (value >= num_messages)
{
throw std::runtime_error("value out of bounds");
}
else if (received[value])
{
std::ostringstream oss;
oss << "ERROR: received element nr. " << value << " two times";
throw std::runtime_error(oss.str());
}
received[value] = true;
}
// done
}
void usage()
{
cout << "usage:" << endl
<< "queue_test [messages] [producer threads] "
<< "[list impl.] {format string}"
<< endl
<< " available implementations:" << endl
<< " - sutter_list" << endl
<< " - blocking_sutter_list" << endl
<< " - cached_stack" << endl
<< " - blocking_cached_stack" << endl
<< " - blocking_cached_stack2" << endl
<< endl
<< " possible format string variables: " << endl
<< " - $MESSAGES" << endl
<< " - $PRODUCERS" << endl
<< " - $TIME" << endl
<< endl
<< "example: ./queue_test 10000 10 cached_list \"$MESSAGES $TIME\""
<< endl;
}
template<typename Queue, typename Allocator, typename Processor>
double run_test(size_t num_messages, size_t num_producers,
Allocator element_allocator, Processor element_processor)
{
size_t num_messages_per_producer = num_messages / num_producers;
// measurement
boost::timer t0;
// locals
Queue list;
std::vector<boost::thread*> producer_threads(num_producers);
for (size_t i = 0; i < num_producers; ++i)
{
producer_threads[i] = new boost::thread(producer<Queue, Allocator>,
boost::ref(list),
boost::ref(element_allocator),
i * num_messages_per_producer,
(i+1) * num_messages_per_producer);
}
// run consumer in main thread
consumer(list, element_processor, num_messages);
// print result
return t0.elapsed();
}
struct cs_element
{
size_t value;
std::atomic<cs_element*> next;
cs_element(size_t val = 0) : value(val), next(0) { }
};
int main(int argc, char** argv)
{
if (argc < 4 || argc > 5)
{
usage();
return -1;
}
size_t num_messages = boost::lexical_cast<size_t>(argv[1]);
size_t num_producers = boost::lexical_cast<size_t>(argv[2]);
if (num_messages == 0 || num_producers == 0)
{
cerr << "invalid arguments" << endl;
return -2;
}
if ((num_messages % num_producers) != 0)
{
cerr << "(num_messages % num_producers) != 0" << endl;
return -3;
}
std::string format_string;
if (argc == 5)
{
format_string = argv[4];
}
else
{
format_string = "$MESSAGES $TIME";
}
std::string list_name = argv[3];
double elapsed_time;
if (list_name == "sutter_list")
{
elapsed_time = run_test<sutter_list<size_t>>(
num_messages,
num_producers,
[] (size_t value) -> size_t* {
return new size_t(value);
},
[] (size_t* value, size_t& storage) {
storage = *value;
delete value;
}
);
}
else if (list_name == "blocking_sutter_list")
{
elapsed_time = run_test<blocking_sutter_list<size_t>>(
num_messages,
num_producers,
[] (size_t value) -> size_t* {
return new size_t(value);
},
[] (size_t* value, size_t& storage) {
storage = *value;
delete value;
}
);
}
else if (list_name == "cached_stack")
{
elapsed_time = run_test<cached_stack<cs_element>>(
num_messages,
num_producers,
[] (size_t value) -> cs_element* {
return new cs_element(value);
},
[] (cs_element* e, size_t& storage) {
storage = e->value;
delete e;
}
);
}
else if (list_name == "blocking_cached_stack")
{
elapsed_time = run_test<blocking_cached_stack<cs_element>>(
num_messages,
num_producers,
[] (size_t value) -> cs_element* {
return new cs_element(value);
},
[] (cs_element* e, size_t& storage) {
storage = e->value;
delete e;
}
);
}
else if (list_name == "blocking_cached_stack2")
{
elapsed_time = run_test<blocking_cached_stack2<cs_element>>(
num_messages,
num_producers,
[] (size_t value) -> cs_element* {
return new cs_element(value);
},
[] (cs_element* e, size_t& storage) {
storage = e->value;
delete e;
}
);
}
else
{
cerr << "unknown list" << endl;
usage();
return -4;
}
// build output message
std::vector<std::pair<std::string, std::string>> replacements = {
{ "$MESSAGES", argv[1] },
{ "$PRODUCERS", argv[2] },
{ "$TIME", boost::lexical_cast<std::string>(elapsed_time) }
};
for (auto i = replacements.begin(); i != replacements.end(); ++i)
{
const std::string& needle = i->first;
const std::string& value = i->second;
std::string::size_type pos = format_string.find(needle);
if (pos != std::string::npos)
{
format_string.replace(pos, pos + needle.size(), value);
}
}
cout << format_string << endl;
// done
return 0;
}
#ifndef SUTTER_LIST_HPP
#define SUTTER_LIST_HPP
#include <atomic>
#include <boost/thread.hpp>
#include "defines.hpp"
// This implementation is a single-consumer version
// of an example implementation of Herb Sutter:
// http://drdobbs.com/cpp/211601363.
// T is any type
template<typename T>
class sutter_list
{
struct node
{
node(T* val = 0) : value(val), next(0) { }
T* value;
std::atomic<node*> next;
char pad[CACHE_LINE_SIZE - sizeof(T*)- sizeof(std::atomic<node*>)];
};
// one consumer at a time
node* m_first;
char m_pad1[CACHE_LINE_SIZE - sizeof(node*)];
// for one producers at a time
node* m_last;
char m_pad2[CACHE_LINE_SIZE - sizeof(node*)];
// shared among producers
std::atomic<bool> m_producer_lock;
public:
sutter_list()
{
m_first = m_last = new node;
m_producer_lock = false;
}
~sutter_list()
{
while (m_first)
{
node* tmp = m_first;
m_first = tmp->next;
delete tmp;
}
}
// takes ownership of what
void push(T* what)
{
node* tmp = new node(what);
// acquire exclusivity
while (m_producer_lock.exchange(true))
{
boost::this_thread::yield();
}
// publish & swing last forward
m_last->next = tmp;
m_last = tmp;
// release exclusivity
m_producer_lock = false;
}
// returns nullptr on failure
T* try_pop()
{
// no critical section; only one consumer allowed
node* first = m_first;
node* next = m_first->next;
if (next)
{
// queue is not empty
T* result = next->value; // take it out
next->value = 0; // of the node
// swing first forward
m_first = next;
// delete old dummy
delete first;
// done
return result;
}
// queue was empty
return 0;
}
// polls the queue until an element was dequeued
T* pop()
{
T* result = try_pop();
while (!result)
{
boost::this_thread::yield();
result = try_pop();
}
return result;
}
};
#endif // SUTTER_LIST_HPP
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment