hoo2/PDS
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

HW2: RC4 - [Not tested] The final version

Browse Source
This commit is contained in:
hoo2
2025-01-09 00:30:16 +02:00
parent 1b271de2a0
commit f849e8a309
7 changed files with 269 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files
homework_2/include/config.h
+8 -3
View File
@@ -49,14 +49,19 @@ static constexpr size_t MAX_PIPELINE_SIZE = 64UL;
using distValue_t = uint32_t;
/*!
* Session option for each invocation of the executable
* Session option for each invocation of the executable.
*
* @note
* The values of the members are set from the command line.
*/
struct config_t {
size_t arraySize{DEFAULT_DATA_SIZE}; //!< The array size of the local data to sort.
size_t pipeline{1UL}; //!< Pipeline stages
bool exchangeOpt{false}; //!< Flag to request the exchange optimization
size_t pipeline{1UL}; //!< Pipeline stages (1 to disable)
bool validation{false}; //!< Request a full validation at the end, performed by process rank 0.
bool ndebug{false}; //!< Skips debug trap on DEBUG builds.
bool perf{false}; //!< Enable performance timing measurements and prints.
size_t perf{1}; //!< Enable performance timing measurements and prints and repeat
//!< the performs the sorting <perf> times to average the measurements
bool verbose{false}; //!< Flag to enable verbose output to stdout.
};
homework_2/include/distsort.hpp
+84 -18
View File
@@ -22,7 +22,15 @@
#include "utils.hpp"
extern Timing TfullSort, Texchange, Tminmax, TelbowSort; // make timers public
/*
* Exported timers
*/
extern Timing Ttotal;
extern Timing TfullSort;
extern Timing Texchange;
extern Timing Tminmax;
extern Timing TelbowSort;
/*!
* Enumerator for the different versions of the sorting method
@@ -167,6 +175,9 @@ void fullSort(RangeT& data, bool ascending) noexcept {
else {
__gnu_parallel::sort(data.begin(), data.end(), std::greater<>());
}
if (config.exchangeOpt)
updateMinMax(localStat, data);
}
/*!
@@ -231,6 +242,43 @@ void elbowSort(ShadowedDataT& data, bool ascending) noexcept {
elbowSortCore(data, ascending, std::greater<>());
}
/*!
* Predicate for exchange optimization. Returns true only if an exchange between partners is needed.
* In order to do that we exchange min and max statistics of the partner's data.
*
* @tparam StatT Statistics data type (for min-max)
*
* @param lstat [const StatT] Reference to the local statistic data
* @param rstat [StatT] Reference to the remote statistic data to fill
* @param part [mpi_id_t] The partner for the exchange
* @param tag [int] The tag to use for the exchange of stats
* @param keepSmall [bool] Flag to indicate if the local thread keeps the small ro the large values
* @return True if we need data exchange, false otherwise
*/
template<typename StatT>
bool needsExchange(const StatT& lstat, StatT& rstat, mpi_id_t part, int tag, bool keepSmall) {
timeCall(Texchange, mpi.exchange_it, lstat, rstat, part, tag);
return (keepSmall) ?
rstat.min < lstat.max // Lmin: rstat.min - Smax: lstat.max
: lstat.min < rstat.max; // Lmin: lstat.min - Smax: rstat.max
}
/*!
* Update stats utility
*
* @tparam RangeT A range type with random access iterator
* @tparam StatT Statistics data type (for min-max)
*
* @param stat [StatT] Reference to the statistic data to update
* @param data [const RangeT] Reference to the sequence to extract stats from
*/
template<typename RangeT, typename StatT>
void updateMinMax(StatT& stat, const RangeT& data) noexcept {
auto [min, max] = std::minmax_element(data.begin(), data.end());
stat.min = *min;
stat.max = *max;
}
/*!
* Takes two sequences and selects either the larger or the smaller items
@@ -276,7 +324,11 @@ void keepMinOrMax(ValueT* local, const ValueT* remote, size_t count, bool keepSm
size_t tagGenerator(size_t depth, size_t step, size_t stage = 0);
/*!
* A pipeline loop for mixing min-max process with mpi data exchange
* An exchange functionality to support both Bubbletonic and Bitonic sort algorithms.
*
* @note
* In case of pipeline request it switches to non-blocking MPI communication for
* pipelining min-max process with mpi data exchange
*
* @tparam ShadowedDataT A Shadowed buffer type with random access iterator.
*
@@ -289,28 +341,38 @@ size_t tagGenerator(size_t depth, size_t step, size_t stage = 0);
* The @c tag is increased inside the pipeline loop for each different data exchange
*/
template<typename ShadowedDataT>
void exchangePipeline(ShadowedDataT& data, mpi_id_t partner, bool keepSmall, int tag) {
void exchange(ShadowedDataT& data, mpi_id_t partner, bool keepSmall, int tag) {
using Value_t = typename ShadowedDataT::value_type;
// Init counters and pointers
size_t count = data.size() / config.pipeline;
Value_t* active = data.getActive().data();
Value_t* shadow = data.getShadow().data();
size_t count = data.size() / config.pipeline;
// Pipeline
Texchange.start();
mpi.exchange_start(active, shadow, count, partner, tag);
for (size_t stage = 0 ; stage < config.pipeline ; active += count, shadow += count) {
// Wait previous chunk
mpi.exchange_wait(); Texchange.stop();
if (++stage < config.pipeline) {
// Start next chunk if there is a next one
Texchange.start();
mpi.exchange_start(active + count, shadow + count, count, partner, ++tag);
if (config.pipeline > 1) {
// Pipeline case - use async MPI
Texchange.start();
mpi.exchange_start(active, shadow, count, partner, tag);
for (size_t stage = 0; stage < config.pipeline; active += count, shadow += count) {
// Wait previous chunk
mpi.exchange_wait();
Texchange.stop();
if (++stage < config.pipeline) {
// Start next chunk if there is a next one
Texchange.start();
mpi.exchange_start(active + count, shadow + count, count, partner, ++tag);
}
// process the arrived data
timeCall(Tminmax, keepMinOrMax, active, shadow, count, keepSmall);
}
// process the arrived data
}
else {
// No pipeline - use blocking MPI
timeCall(Texchange, mpi.exchange, active, shadow, count, partner, tag);
timeCall(Tminmax, keepMinOrMax, active, shadow, count, keepSmall);
}
if (config.exchangeOpt)
updateMinMax(localStat, data);
}
/*!
@@ -339,8 +401,10 @@ void distBubbletonic(ShadowedDataT& data, mpi_id_t Processes, mpi_id_t rank) {
isActive(part, Processes) ) {
// Exchange with partner, keep nim-or-max and sort - O(N)
int tag = static_cast<int>(tagGenerator(0, step));
exchangePipeline(data, part, ks, tag);
timeCall(TelbowSort, elbowSort, data, ascending<SortMode::Bubbletonic>(rank, Processes));
if (!config.exchangeOpt || needsExchange(localStat, remoteStat, part, tag++, ks)) {
exchange(data, part, ks, tag);
timeCall(TelbowSort, elbowSort, data, ascending<SortMode::Bubbletonic>(rank, Processes));
}
}
}
@@ -378,7 +442,9 @@ void distBitonic(ShadowedDataT& data, mpi_id_t Processes, mpi_id_t rank) {
auto ks = keepSmall<SortMode::Bitonic>(rank, part, depth);
// Exchange with partner, keep nim-or-max
int tag = static_cast<int>(tagGenerator(depth, step));
exchangePipeline(data, part, ks, tag);
if (!config.exchangeOpt || needsExchange(localStat, remoteStat, part, tag++, ks)) {
exchange(data, part, ks, tag);
}
}
// sort - O(N)
timeCall(TelbowSort, elbowSort, data, ascending<SortMode::Bitonic>(rank, depth));
homework_2/include/utils.hpp
+102 -9
View File
@@ -17,6 +17,22 @@
#include "config.h"
/*!
* Min-Max statistics data for exchange optimization
* @tparam Value_t The underlying data type of the sequence data
*/
template <typename Value_t>
struct Stat_t {
using value_type = Value_t; //!< meta-export the type
Value_t min{}; //!< The minimum value of the sequence
Value_t max{}; //!< The maximum value of the sequence
};
//! Application data selection alias
using distStat_t = Stat_t<distValue_t>;
extern distStat_t localStat, remoteStat; // Make stats public
/*
* MPI_<type> dispatcher mechanism
*/
@@ -78,6 +94,64 @@ struct MPI_t {
name_ = std::string (processor_name, name_len);
}
/*!
* Exchange one data object of type @c T with partner as part of the sorting network of both
* bubbletonic or bitonic sorting algorithms.
*
* This function matches a transmit and a receive in order for fully exchanged the data object
* between current node and partner.
*
* @tparam T The object type
*
* @param local [const T&] Reference to the local object to send
* @param remote [T&] Reference to the object to receive data from partner
* @param partner [mpi_id_t] The partner for the exchange
* @param tag [int] The tag to use for the MPI communication
*/
template<typename T>
void exchange_it(const T& local, T& remote, ID_t partner, int tag) {
if (tag < 0)
throw std::runtime_error("(MPI) exchange_it() [tag] - Out of bound");
MPI_Status status;
int err;
if ((err = MPI_Sendrecv(
&local, sizeof(T), MPI_BYTE, partner, tag,
&remote, sizeof(T), MPI_BYTE, partner, tag,
MPI_COMM_WORLD, &status
)) != MPI_SUCCESS)
mpi_throw(err, "(MPI) MPI_Sendrecv() [item] - ");
}
/*!
* Exchange data with partner as part of the sorting network of both bubbletonic or bitonic
* sorting algorithms.
*
* This function matches a transmit and a receive in order for fully exchanged data between
* current node and partner.
*
* @tparam T The inner valur type used in buffer
*
* @param ldata [const ValueT*] Pointer to local data to send
* @param rdata [ValueT*] Pointer to buffer to receive data from partner
* @param count [size_t] The number of data to exchange
* @param partner [mpi_id_t] The partner for the exchange
* @param tag [int] The tag to use for the MPI communication
*/
template<typename ValueT>
void exchange(const ValueT* ldata, ValueT* rdata, size_t count, ID_t partner, int tag) {
if (tag < 0)
throw std::runtime_error("(MPI) exchange_data() [tag] - Out of bound");
MPI_Datatype datatype = MPI_TypeMapper<ValueT>::getType();
MPI_Status status;
int err;
if ((err = MPI_Sendrecv(
ldata, count, datatype, partner, tag,
rdata, count, datatype, partner, tag,
MPI_COMM_WORLD, &status
)) != MPI_SUCCESS)
mpi_throw(err, "(MPI) MPI_Sendrecv() [data] - ");
}
/*!
* Initiate a data exchange data with partner using non-blocking Isend-Irecv, as part of the
@@ -353,33 +427,51 @@ struct Timing {
using milliseconds = std::chrono::milliseconds;
using seconds = std::chrono::seconds;
//! Setup measurement rounds
void init(size_t rounds) {
duration_.resize(rounds);
for (auto& d : duration_)
d = Tduration::zero();
}
//! tool to mark the starting point
Tpoint start() noexcept { return mark_ = std::chrono::steady_clock::now(); }
//! tool to mark the ending point
Tpoint stop() noexcept {
Tpoint now = std::chrono::steady_clock::now();
duration_ += dt(now, mark_);
duration_[current_] += dt(now, mark_);
return now;
}
//! Switch timing slot
void next() noexcept {
++current_;
current_ %= duration_.size();
}
Tduration& median() noexcept {
std::sort(duration_.begin(), duration_.end());
return duration_[duration_.size()/2];
}
//! A duration calculation utility
static Tduration dt(Tpoint t2, Tpoint t1) noexcept {
return std::chrono::duration_cast<Tduration>(t2 - t1);
}
//! Tool to print the time interval
void print_duration(const char *what, mpi_id_t rank) noexcept {
if (std::chrono::duration_cast<microseconds>(duration_).count() < 10000)
static void print_duration(const Tduration& duration, const char *what, mpi_id_t rank) noexcept {
if (std::chrono::duration_cast<microseconds>(duration).count() < 10000)
std::cout << "[Timing] (Rank " << rank << ") " << what << ": "
<< std::to_string(std::chrono::duration_cast<microseconds>(duration_).count()) << " [usec]\n";
else if (std::chrono::duration_cast<milliseconds>(duration_).count() < 10000)
<< std::to_string(std::chrono::duration_cast<microseconds>(duration).count()) << " [usec]\n";
else if (std::chrono::duration_cast<milliseconds>(duration).count() < 10000)
std::cout << "[Timing] (Rank " << rank << ") " << what << ": "
<< std::to_string(std::chrono::duration_cast<milliseconds>(duration_).count()) << " [msec]\n";
<< std::to_string(std::chrono::duration_cast<milliseconds>(duration).count()) << " [msec]\n";
else {
char stime[26]; // fit ulong
auto sec = std::chrono::duration_cast<seconds>(duration_).count();
auto msec = (std::chrono::duration_cast<milliseconds>(duration_).count() % 1000) / 10; // keep 2 digit
auto sec = std::chrono::duration_cast<seconds>(duration).count();
auto msec = (std::chrono::duration_cast<milliseconds>(duration).count() % 1000) / 10; // keep 2 digit
std::sprintf(stime, "%ld.%1ld", sec, msec);
std::cout << "[Timing] (Rank " << rank << ") " << what << ": " << stime << " [sec]\n";
}
@@ -387,8 +479,9 @@ struct Timing {
}
private:
size_t current_{0};
Tpoint mark_{};
Tduration duration_{};
std::vector<Tduration> duration_{1};
};
/*!