DisCostiC v2.0.0
Distributed Cost in Cluster: a cross-architecture simulator for massively parallel MPI programs
Loading...
Searching...
No Matches
DisCosTiC.cpp File Reference
#include <iostream>
#include <list>
#include <algorithm>
#include <cmath>
#include <mpi.h>
#include <iomanip>
#include <numeric>
#include "../include/ConfigParser.hpp"
#include "../include/NetworkConfigParser.hpp"
#include "../include/YAMLParser.hpp"
#include "../visualization/TimeRankOP.hpp"
#include "../visualization/ItacVtf.hpp"
#include "../include/GridInit.hpp"
#include "../include/CollectiveModel.hpp"
#include <mutex>
+ Include dependency graph for DisCosTiC.cpp:

Go to the source code of this file.

Macros

#define USE_CHROMEVIZ
 
#define USE_ITACVIZ
 

Enumerations

enum  communication_mode { LOGGP = 0 , SIMPLELB = 1 }
 
enum  communication_type { INTRACHIP = 0 , INTERCHIP = 1 , INTERNODE = 2 , INTERCLUSTER = 3 }
 
enum  time { START = 0 , END = 1 }
 
enum  bound_type { COMPUTE = 0 , MEMORY = 1 }
 

Functions

DisCosTiC::DisCosTiC_OP copy (DisCosTiC::DisCosTiC_OP op_og, DisCosTiC::DisCosTiC_OP op_copy)
 
double * finalize (double a, double b, double c, double d, double e, double arr[5])
 
static std::set< int > ncm_overlap_ranks (AST *ast, int comm_kind, int system_number, int local_node, int local_socket, int peer_node)
 
int main (int argc, char **argv)
 

Variables

int bound
 
int scaling_cores
 
int bytes_to_send
 
int virtual_rank = 0
 
int system_number = 0
 
int task_per_node = 0
 
int node = 0
 
int cc_numa_domain_per_socket = 0
 
int cores_per_socket = 0
 
int cc_numa_domain = 0
 
int socket = 0
 
int primary_processes = 0
 
int secondary_processes = 0
 
int heteregeneous_mode = 0
 
int primary_cores_per_socket = 0
 
int secondary_cores_per_socket = 0
 
int primary_task_per_node = 0
 
int secondary_task_per_node = 0
 
std::string arch_name = ""
 

Detailed Description

DISCOSTIC_HPP \author : Ayesha Afzal ayesh.nosp@m.a.af.nosp@m.zal@f.nosp@m.au.d.nosp@m.e \copyright © 2024 HPC, FAU Erlangen-Nuremberg. All rights reserved.

Definition in file DisCosTiC.cpp.

Macro Definition Documentation

◆ USE_CHROMEVIZ

#define USE_CHROMEVIZ

Definition at line 27 of file DisCosTiC.cpp.

◆ USE_ITACVIZ

#define USE_ITACVIZ

Definition at line 28 of file DisCosTiC.cpp.

Enumeration Type Documentation

◆ bound_type

enum bound_type
Enumerator
COMPUTE 
MEMORY 

Definition at line 51 of file DisCosTiC.cpp.

◆ communication_mode

Enumerator
LOGGP 
SIMPLELB 

Definition at line 31 of file DisCosTiC.cpp.

◆ communication_type

Enumerator
INTRACHIP 
INTERCHIP 
INTERNODE 
INTERCLUSTER 

Definition at line 37 of file DisCosTiC.cpp.

◆ time

enum time
Enumerator
START 
END 

Definition at line 45 of file DisCosTiC.cpp.

Function Documentation

◆ copy()

◆ finalize()

double * finalize ( double a,
double b,
double c,
double d,
double e,
double arr[5] )

Definition at line 98 of file DisCosTiC.cpp.

Referenced by main().

+ Here is the caller graph for this function:

◆ main()

int main ( int argc,
char ** argv )

< generate AST for Oneway_PositiveDisplacement communication pattern example

rootGrabber: get root nodes of DisCosTiC::AST_OP_TYPE type (i.e., exec in our case ) of all ranks

it parses and print user-defined parameters or args of configuration file (config.cfg).

< number of synchronization induced by MPI routines

< print version number of toolkit

< print current date and time

define and initialize next available time of computation, receive and send for all CPU nodes and ranks

Starting Time

initialize root operations for each rank (only once per run!). These sorted operations order by type retrieve from test case application added to the queue (that has operations of all processes and their earliest start times). (each operation = time, rank, next operation).

Start of parallelization code runs for all processes that are not the master rank

the switch on type of each operation: whenever a blocking or non-blocking dependency or something is satisfied, the rank is added to the list of ranks that completed someting (finishedRankList)

operations of all ranks with their starting times and counter over operation statuses

< get rid of baseline comp time

< get rid of baseline comp time and add original computing time modeled by analytical roofline model

injected long disturbance at given rank and given timestep

< message is only received after G is charged

add new sorted operations order by type to the queue

Start of Master Ranks code runs for the master rank

end of parallelization

Definition at line 137 of file DisCosTiC.cpp.

References DisCosTiC::DisCosTiC_OP::algorithm, allRanksTime, UserInterface::ChromeTraceViz::arc, arch_name, UserInterface::ChromeTraceViz::args(), DisCosTiC::BCAST_RING, DisCosTiC::BLOCKING, bound, DisCosTiC::DisCosTiC_OP::bufSize, bytes_to_send, DisCosTiC::CollectiveModel::calculateCollectiveTime(), cc_numa_domain, cc_numa_domain_per_socket, CFG_args, UserInterface::ChromeTraceViz::closeFile(), UserInterface::ItacVtf::closeFile(), DisCosTiC::COLL_ALLREDUCE, DisCosTiC::COLL_BCAST, DisCosTiC::COLL_REDUCE, DisCosTiC::COLLECTIVE, UserInterface::ItacVtf::collective(), DisCosTiC::DisCosTiC_OP::collectiveNumRanks, DisCosTiC::DisCosTiC_OP::collectiveRoot, DisCosTiC::DisCosTiC_OP::collectiveType, DisCosTiC::COMP, UserInterface::ItacVtf::comp(), UserInterface::TimeRankOP::comp(), DisCosTiC::COMPADD, UserInterface::ChromeTraceViz::completeEvents(), copy(), UserInterface::YAMLParser::cores_per_chip, UserInterface::YAMLParser::cores_per_numa_domain, cores_per_socket, END, finalize(), UserInterface::ChromeTraceViz::flowEventBegin(), UserInterface::ChromeTraceViz::flowEventEnd(), DisCosTiC::Benchmark::GetNumCores(), DisCosTiC::Benchmark::GetNumNetworks(), DisCosTiC::Grid::getNumOps(), DisCosTiC::getRange(), DisCosTiC::Grid::getSortedRootOps(), DisCosTiC::Grid::getTypeSortedOps(), UserInterface::ConfigParser::getValue(), DisCosTiC::Grid_Init::graphVec, help, heteregeneous_mode, INTERCHIP, INTERCLUSTER, INTERNODE, INTRACHIP, DisCosTiC::DisCosTiC_OP::label, LOGGP, MEMORY, DisCosTiC::DisCosTiC_OP::mode, ncm_overlap_ranks(), DisCosTiC::DisCosTiC_OP::network, DisCosTiC::DisCosTiC_OP::node, node, DisCosTiC::NONBLOCKING, DisCosTiC::Benchmark::numOperations, DisCosTiC::DisCosTiC_OP::numOpsInQueue, DisCosTiC::Benchmark::Oneway_PositiveDisplacement(), UserInterface::TimeRankOP::orecv(), UserInterface::TimeRankOP::osend(), primary_cores_per_socket, primary_processes, primary_task_per_node, print_OpPropertiesNonPointerT, AST::printNCMRanks(), DisCosTiC::DisCosTiC_OP::rank, UserInterface::NetworkConfigParser::readData(), DisCosTiC::RECV, UserInterface::ItacVtf::recv(), scaling_cores, secondary_cores_per_socket, secondary_processes, secondary_task_per_node, DisCosTiC::SEND, UserInterface::ItacVtf::send(), UserInterface::NetworkConfigParser::setData(), SIMPLELB, slowRankTime, socket, START, DisCosTiC::DisCosTiC_OP::starttime, DisCosTiC::DisCosTiC_OP::syncstart, system_number, DisCosTiC::Benchmark::systemsize, DisCosTiC::DisCosTiC_OP::tag, DisCosTiC::DisCosTiC_OP::target, task_per_node, DisCosTiC::DisCosTiC_OP::time, DisCosTiC::DisCosTiC_OP::type, verboseCompFinalPrint, verboseCompInitPrint, verboseCompPrint, verboseRecvFinalPrint, verboseRecvInitPrint, verboseSendFinalPrint, verboseSendInitPrint, version, and virtual_rank.

+ Here is the call graph for this function:

◆ ncm_overlap_ranks()

static std::set< int > ncm_overlap_ranks ( AST * ast,
int comm_kind,
int system_number,
int local_node,
int local_socket,
int peer_node )
static

Definition at line 109 of file DisCosTiC.cpp.

References INTERCHIP, INTERCLUSTER, AST::interClusterRanks, INTERNODE, AST::intraAndInterNodeRanks, and system_number.

Referenced by main().

+ Here is the caller graph for this function:

Variable Documentation

◆ arch_name

std::string arch_name = ""

Definition at line 77 of file DisCosTiC.cpp.

Referenced by main(), and DisCosTiC::Benchmark::Oneway_PositiveDisplacement().

◆ bound

int bound
extern

Definition at line 9 of file NodeLvlScg.cpp.

Referenced by estimation(), and main().

◆ bytes_to_send

int bytes_to_send
extern

Definition at line 37 of file AST.hpp.

Referenced by main().

◆ cc_numa_domain

int cc_numa_domain = 0

Definition at line 68 of file DisCosTiC.cpp.

Referenced by main().

◆ cc_numa_domain_per_socket

int cc_numa_domain_per_socket = 0

Definition at line 66 of file DisCosTiC.cpp.

Referenced by main().

◆ cores_per_socket

int cores_per_socket = 0

◆ heteregeneous_mode

int heteregeneous_mode = 0

Definition at line 72 of file DisCosTiC.cpp.

Referenced by DisCosTiC::Benchmark::Benchmark(), AST::execNodeLVL(), and main().

◆ node

int node = 0

Definition at line 65 of file DisCosTiC.cpp.

Referenced by main().

◆ primary_cores_per_socket

int primary_cores_per_socket = 0

Definition at line 73 of file DisCosTiC.cpp.

Referenced by AST::insertInNetwork(), and main().

◆ primary_processes

◆ primary_task_per_node

int primary_task_per_node = 0

Definition at line 75 of file DisCosTiC.cpp.

Referenced by AST::insertInNetwork(), and main().

◆ scaling_cores

int scaling_cores
extern

Definition at line 36 of file AST.hpp.

Referenced by main().

◆ secondary_cores_per_socket

int secondary_cores_per_socket = 0

Definition at line 74 of file DisCosTiC.cpp.

Referenced by AST::insertInNetwork(), and main().

◆ secondary_processes

int secondary_processes = 0

◆ secondary_task_per_node

int secondary_task_per_node = 0

Definition at line 76 of file DisCosTiC.cpp.

Referenced by AST::insertInNetwork(), and main().

◆ socket

int socket = 0

Definition at line 69 of file DisCosTiC.cpp.

Referenced by main().

◆ system_number

◆ task_per_node

int task_per_node = 0

◆ virtual_rank

int virtual_rank = 0

Definition at line 62 of file DisCosTiC.cpp.

Referenced by main().