Merge rabit

This commit is contained in:
fis
2020-08-18 03:52:33 +08:00
81 changed files with 11230 additions and 0 deletions

3
rabit/test/.gitignore vendored Normal file
View File

@@ -0,0 +1,3 @@
*.mpi
*_test
*_recover

77
rabit/test/Makefile Normal file
View File

@@ -0,0 +1,77 @@
RABIT_BUILD_DMLC = 0
ifeq ($(RABIT_BUILD_DMLC),1)
DMLC=../dmlc-core
else
DMLC=../../dmlc-core
endif
MPICXX=../mpich/bin/mpicxx
export LDFLAGS= -L../lib -pthread -lm
export CFLAGS = -Wall -O3 -Wno-unknown-pragmas
export CC = gcc
export CXX = g++
#----------------------------
# Settings for power and arm arch
#----------------------------
ARCH := $(shell uname -a)
ifneq (,$(filter $(ARCH), armv6l armv7l powerpc64le ppc64le aarch64))
CFLAGS += -march=native
else
CFLAGS += -msse2
endif
ifndef WITH_FPIC
WITH_FPIC = 1
endif
ifeq ($(WITH_FPIC), 1)
CFLAGS += -fPIC
endif
CFLAGS += -I../include -I $(DMLC)/include -std=c++11
# specify tensor path
BIN = speed_test model_recover local_recover lazy_recover
OBJ = $(RABIT_OBJ) speed_test.o model_recover.o local_recover.o lazy_recover.o
MPIBIN = speed_test.mpi
.PHONY: clean all lib mpi
.PHONY: lib all
all: $(BIN)
lib:
cd ..;make clean;make;cd -
.PHONY: mpi
mpi:
cd ..;make mpi;cd -
# programs
speed_test.o: speed_test.cc ../include/rabit/*.h lib mpi
model_recover.o: model_recover.cc ../include/rabit/*.h lib
local_recover.o: local_recover.cc ../include/rabit/*.h lib
lazy_recover.o: lazy_recover.cc ../include/rabit/*.h lib
# we can link against MPI version to get use MPI
speed_test: speed_test.o $(RABIT_OBJ)
speed_test.mpi: speed_test.o $(MPIOBJ)
model_recover: model_recover.o $(RABIT_OBJ)
local_recover: local_recover.o $(RABIT_OBJ)
lazy_recover: lazy_recover.o $(RABIT_OBJ)
$(BIN) :
$(CXX) $(CFLAGS) -o $@ $(filter %.cpp %.o %.c %.cc, $^) ../lib/librabit_mock.a $(LDFLAGS)
$(OBJ) :
$(CXX) -c $(CFLAGS) -o $@ $(firstword $(filter %.cpp %.c %.cc, $^) )
$(MPIBIN) :
$(MPICXX) $(CFLAGS) -I../mpich/include -shared -o $@ $(filter %.cpp %.o %.c %.cc, $^) \
../lib/librabit_mpi.so $(LDFLAGS) -L../mpich/lib -Wl,-rpath,../mpich/lib -lmpi
clean:
$(RM) $(OBJ) $(BIN) $(MPIBIN) $(MPIOBJ) *~ ../src/*~

18
rabit/test/README.md Normal file
View File

@@ -0,0 +1,18 @@
Testcases of Rabit
====
This folder contains internal testcases to test correctness and efficiency of rabit API
The example running scripts for testcases are given by test.mk
* type ```make -f test.mk testcasename``` to run certain testcase
Helper Scripts
====
* test.mk contains Makefile documentation of all testcases
* keepalive.sh helper bash to restart a program when it dies abnormally
List of Programs
====
* speed_test: test the running speed of rabit API
* test_local_recover: test recovery of local state when error happens
* test_model_recover: test recovery of global state when error happens

View File

@@ -0,0 +1,32 @@
find_package(GTest REQUIRED)
add_executable(
unit_tests
test_io.cc
test_utils.cc
allreduce_robust_test.cc
allreduce_base_test.cc
allreduce_mock_test.cc
test_main.cpp)
target_link_libraries(
unit_tests PRIVATE
GTest::GTest GTest::Main
rabit_base rabit_mock rabit)
target_include_directories(unit_tests PUBLIC
"$<BUILD_INTERFACE:${rabit_SOURCE_DIR}/include>"
"$<BUILD_INTERFACE:${DMLC_ROOT}/include>")
set_target_properties(unit_tests
PROPERTIES
CXX_STANDARD 11
CXX_STANDARD_REQUIRED ON
RUNTIME_OUTPUT_DIRECTORY ${rabit_BINARY_DIR}
RUNTIME_OUTPUT_DIRECTORY_DEBUG ${rabit_BINARY_DIR}
RUNTIME_OUTPUT_DIRECTORY_RELEASE ${rabit_BINARY_DIR})
add_test(
NAME TestRabitLib
COMMAND unit_tests
WORKING_DIRECTORY ${rabit_BINARY_DIR})

1
rabit/test/cpp/README.md Normal file
View File

@@ -0,0 +1 @@
Unittests for Rabit

View File

@@ -0,0 +1,66 @@
#define RABIT_CXXTESTDEFS_H
#include <gtest/gtest.h>
#include <string>
#include <iostream>
#include "../../src/allreduce_base.h"
TEST(allreduce_base, init_task)
{
rabit::engine::AllreduceBase base;
std::string rabit_task_id = "rabit_task_id=1";
char cmd[rabit_task_id.size()+1];
std::copy(rabit_task_id.begin(), rabit_task_id.end(), cmd);
cmd[rabit_task_id.size()] = '\0';
char* argv[] = {cmd};
base.Init(1, argv);
EXPECT_EQ(base.task_id, "1");
}
TEST(allreduce_base, init_with_cache_on)
{
rabit::engine::AllreduceBase base;
std::string rabit_task_id = "rabit_task_id=1";
char cmd[rabit_task_id.size()+1];
std::copy(rabit_task_id.begin(), rabit_task_id.end(), cmd);
cmd[rabit_task_id.size()] = '\0';
std::string rabit_bootstrap_cache = "rabit_bootstrap_cache=1";
char cmd2[rabit_bootstrap_cache.size()+1];
std::copy(rabit_bootstrap_cache.begin(), rabit_bootstrap_cache.end(), cmd2);
cmd2[rabit_bootstrap_cache.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd3[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd3);
cmd3[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd2, cmd3};
base.Init(3, argv);
EXPECT_EQ(base.task_id, "1");
EXPECT_EQ(base.rabit_bootstrap_cache, 1);
EXPECT_EQ(base.rabit_debug, 1);
}
TEST(allreduce_base, init_with_ring_reduce)
{
rabit::engine::AllreduceBase base;
std::string rabit_task_id = "rabit_task_id=1";
char cmd[rabit_task_id.size()+1];
std::copy(rabit_task_id.begin(), rabit_task_id.end(), cmd);
cmd[rabit_task_id.size()] = '\0';
std::string rabit_reduce_ring_mincount = "rabit_reduce_ring_mincount=1";
char cmd2[rabit_reduce_ring_mincount.size()+1];
std::copy(rabit_reduce_ring_mincount.begin(), rabit_reduce_ring_mincount.end(), cmd2);
cmd2[rabit_reduce_ring_mincount.size()] = '\0';
char* argv[] = {cmd, cmd2};
base.Init(2, argv);
EXPECT_EQ(base.task_id, "1");
EXPECT_EQ(base.reduce_ring_mincount, 1);
}

View File

@@ -0,0 +1,66 @@
#define RABIT_CXXTESTDEFS_H
#include <gtest/gtest.h>
#include <string>
#include <iostream>
#include "../../src/allreduce_base.h"
TEST(allreduce_base, init_task)
{
rabit::engine::AllreduceBase base;
std::string rabit_task_id = "rabit_task_id=1";
char cmd[rabit_task_id.size()+1];
std::copy(rabit_task_id.begin(), rabit_task_id.end(), cmd);
cmd[rabit_task_id.size()] = '\0';
char* argv[] = {cmd};
base.Init(1, argv);
EXPECT_EQ(base.task_id, "1");
}
TEST(allreduce_base, init_with_cache_on)
{
rabit::engine::AllreduceBase base;
std::string rabit_task_id = "rabit_task_id=1";
char cmd[rabit_task_id.size()+1];
std::copy(rabit_task_id.begin(), rabit_task_id.end(), cmd);
cmd[rabit_task_id.size()] = '\0';
std::string rabit_bootstrap_cache = "rabit_bootstrap_cache=1";
char cmd2[rabit_bootstrap_cache.size()+1];
std::copy(rabit_bootstrap_cache.begin(), rabit_bootstrap_cache.end(), cmd2);
cmd2[rabit_bootstrap_cache.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd3[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd3);
cmd3[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd2, cmd3};
base.Init(3, argv);
EXPECT_EQ(base.task_id, "1");
EXPECT_EQ(base.rabit_bootstrap_cache, 1);
EXPECT_EQ(base.rabit_debug, 1);
}
TEST(allreduce_base, init_with_ring_reduce)
{
rabit::engine::AllreduceBase base;
std::string rabit_task_id = "rabit_task_id=1";
char cmd[rabit_task_id.size()+1];
std::copy(rabit_task_id.begin(), rabit_task_id.end(), cmd);
cmd[rabit_task_id.size()] = '\0';
std::string rabit_reduce_ring_mincount = "rabit_reduce_ring_mincount=1";
char cmd2[rabit_reduce_ring_mincount.size()+1];
std::copy(rabit_reduce_ring_mincount.begin(), rabit_reduce_ring_mincount.end(), cmd2);
cmd2[rabit_reduce_ring_mincount.size()] = '\0';
char* argv[] = {cmd, cmd2};
base.Init(2, argv);
EXPECT_EQ(base.task_id, "1");
EXPECT_EQ(base.reduce_ring_mincount, 1);
}

View File

@@ -0,0 +1,36 @@
#define RABIT_CXXTESTDEFS_H
#include <gtest/gtest.h>
#include <string>
#include <iostream>
#include "../../src/allreduce_mock.h"
TEST(allreduce_mock, mock_allreduce)
{
rabit::engine::AllreduceMock m;
std::string mock_str = "mock=0,0,0,0";
char cmd[mock_str.size()+1];
std::copy(mock_str.begin(), mock_str.end(), cmd);
cmd[mock_str.size()] = '\0';
char* argv[] = {cmd};
m.Init(1, argv);
m.rank = 0;
EXPECT_THROW(m.Allreduce(nullptr,0,0,nullptr,nullptr,nullptr), dmlc::Error);
}
TEST(allreduce_mock, mock_broadcast)
{
rabit::engine::AllreduceMock m;
std::string mock_str = "mock=0,1,2,0";
char cmd[mock_str.size()+1];
std::copy(mock_str.begin(), mock_str.end(), cmd);
cmd[mock_str.size()] = '\0';
char* argv[] = {cmd};
m.Init(1, argv);
m.rank = 0;
m.version_number=1;
m.seq_counter=2;
EXPECT_THROW(m.Broadcast(nullptr,0,0), dmlc::Error);
}

View File

@@ -0,0 +1,52 @@
#define RABIT_CXXTESTDEFS_H
#include <gtest/gtest.h>
#include <string>
#include <iostream>
#include <dmlc/logging.h>
#include "../../src/allreduce_mock.h"
TEST(allreduce_mock, mock_allreduce)
{
rabit::engine::AllreduceMock m;
std::string mock_str = "mock=0,0,0,0";
char cmd[mock_str.size()+1];
std::copy(mock_str.begin(), mock_str.end(), cmd);
cmd[mock_str.size()] = '\0';
char* argv[] = {cmd};
m.Init(1, argv);
m.rank = 0;
EXPECT_THROW({m.Allreduce(nullptr,0,0,nullptr,nullptr,nullptr);}, dmlc::Error);
}
TEST(allreduce_mock, mock_broadcast)
{
rabit::engine::AllreduceMock m;
std::string mock_str = "mock=0,1,2,0";
char cmd[mock_str.size()+1];
std::copy(mock_str.begin(), mock_str.end(), cmd);
cmd[mock_str.size()] = '\0';
char* argv[] = {cmd};
m.Init(1, argv);
m.rank = 0;
m.version_number=1;
m.seq_counter=2;
EXPECT_THROW({m.Broadcast(nullptr,0,0);}, dmlc::Error);
}
TEST(allreduce_mock, mock_gather)
{
rabit::engine::AllreduceMock m;
std::string mock_str = "mock=3,13,22,0";
char cmd[mock_str.size()+1];
std::copy(mock_str.begin(), mock_str.end(), cmd);
cmd[mock_str.size()] = '\0';
char* argv[] = {cmd};
m.Init(1, argv);
m.rank = 3;
m.version_number=13;
m.seq_counter=22;
EXPECT_THROW({m.Allgather(nullptr,0,0,0,0);}, dmlc::Error);
}

View File

@@ -0,0 +1,233 @@
#define RABIT_CXXTESTDEFS_H
#include <gtest/gtest.h>
#include <chrono>
#include <string>
#include <iostream>
#include "../../src/allreduce_robust.h"
inline void mockerr(const char *fmt, ...) {EXPECT_STRCASEEQ(fmt, "[%d] exit due to time out %d s\n");}
inline void mockassert(bool val, const char *fmt, ...) {}
rabit::engine::AllreduceRobust::ReturnType err_type(rabit::engine::AllreduceRobust::ReturnTypeEnum::kSockError);
rabit::engine::AllreduceRobust::ReturnType succ_type(rabit::engine::AllreduceRobust::ReturnTypeEnum::kSuccess);
TEST(allreduce_robust, sync_error_timeout)
{
rabit::engine::AllreduceRobust m;
std::string rabit_timeout = "rabit_timeout=1";
char cmd[rabit_timeout.size()+1];
std::copy(rabit_timeout.begin(), rabit_timeout.end(), cmd);
cmd[rabit_timeout.size()] = '\0';
std::string rabit_timeout_sec = "rabit_timeout_sec=1";
char cmd1[rabit_timeout_sec.size()+1];
std::copy(rabit_timeout_sec.begin(), rabit_timeout_sec.end(), cmd1);
cmd1[rabit_timeout_sec.size()] = '\0';
char* argv[] = {cmd,cmd1};
m.Init(2, argv);
m.rank = 0;
m.rabit_bootstrap_cache = 1;
m._error = mockerr;
m._assert = mockassert;
EXPECT_EQ(m.CheckAndRecover(err_type), false);
std::this_thread::sleep_for(std::chrono::milliseconds(1500));
EXPECT_EQ(m.rabit_timeout_task.get(), false);
}
TEST(allreduce_robust, sync_error_reset)
{
rabit::engine::AllreduceRobust m;
std::string rabit_timeout = "rabit_timeout=1";
char cmd[rabit_timeout.size()+1];
std::copy(rabit_timeout.begin(), rabit_timeout.end(), cmd);
cmd[rabit_timeout.size()] = '\0';
std::string rabit_timeout_sec = "rabit_timeout_sec=1";
char cmd1[rabit_timeout_sec.size()+1];
std::copy(rabit_timeout_sec.begin(), rabit_timeout_sec.end(), cmd1);
cmd1[rabit_timeout_sec.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd2[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd2);
cmd2[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd1,cmd2};
m.Init(3, argv);
m.rank = 0;
m._assert = mockassert;
EXPECT_EQ(m.CheckAndRecover(err_type), false);
std::this_thread::sleep_for(std::chrono::milliseconds(100));
EXPECT_EQ(m.CheckAndRecover(succ_type), true);
EXPECT_EQ(m.rabit_timeout_task.get(), true);
m.Shutdown();
}
TEST(allreduce_robust, sync_success_error_timeout)
{
rabit::engine::AllreduceRobust m;
std::string rabit_timeout = "rabit_timeout=1";
char cmd[rabit_timeout.size()+1];
std::copy(rabit_timeout.begin(), rabit_timeout.end(), cmd);
cmd[rabit_timeout.size()] = '\0';
std::string rabit_timeout_sec = "rabit_timeout_sec=1";
char cmd1[rabit_timeout_sec.size()+1];
std::copy(rabit_timeout_sec.begin(), rabit_timeout_sec.end(), cmd1);
cmd1[rabit_timeout_sec.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd2[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd2);
cmd2[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd1,cmd2};
m.Init(3, argv);
m.rank = 0;
m.rabit_bootstrap_cache = 1;
m._assert = mockassert;
m._error = mockerr;
EXPECT_EQ(m.CheckAndRecover(succ_type), true);
std::this_thread::sleep_for(std::chrono::milliseconds(100));
EXPECT_EQ(m.CheckAndRecover(err_type), false);
std::this_thread::sleep_for(std::chrono::milliseconds(1500));
EXPECT_EQ(m.rabit_timeout_task.get(), false);
}
TEST(allreduce_robust, sync_success_error_success)
{
rabit::engine::AllreduceRobust m;
std::string rabit_timeout = "rabit_timeout=1";
char cmd[rabit_timeout.size()+1];
std::copy(rabit_timeout.begin(), rabit_timeout.end(), cmd);
cmd[rabit_timeout.size()] = '\0';
std::string rabit_timeout_sec = "rabit_timeout_sec=1";
char cmd1[rabit_timeout_sec.size()+1];
std::copy(rabit_timeout_sec.begin(), rabit_timeout_sec.end(), cmd1);
cmd1[rabit_timeout_sec.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd2[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd2);
cmd2[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd1,cmd2};
m.Init(3, argv);
m.rank = 0;
m.rabit_bootstrap_cache = 1;
m._assert = mockassert;
EXPECT_EQ(m.CheckAndRecover(succ_type), true);
std::this_thread::sleep_for(std::chrono::milliseconds(10));
EXPECT_EQ(m.CheckAndRecover(err_type), false);
std::this_thread::sleep_for(std::chrono::milliseconds(10));
EXPECT_EQ(m.CheckAndRecover(succ_type), true);
std::this_thread::sleep_for(std::chrono::milliseconds(1100));
EXPECT_EQ(m.rabit_timeout_task.get(), true);
m.Shutdown();
}
TEST(allreduce_robust, sync_error_no_reset_timeout)
{
rabit::engine::AllreduceRobust m;
std::string rabit_timeout = "rabit_timeout=1";
char cmd[rabit_timeout.size()+1];
std::copy(rabit_timeout.begin(), rabit_timeout.end(), cmd);
cmd[rabit_timeout.size()] = '\0';
std::string rabit_timeout_sec = "rabit_timeout_sec=1";
char cmd1[rabit_timeout_sec.size()+1];
std::copy(rabit_timeout_sec.begin(), rabit_timeout_sec.end(), cmd1);
cmd1[rabit_timeout_sec.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd2[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd2);
cmd2[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd1,cmd2};
m.Init(3, argv);
m.rank = 0;
m.rabit_bootstrap_cache = 1;
m._assert = mockassert;
m._error = mockerr;
auto start = std::chrono::system_clock::now();
EXPECT_EQ(m.CheckAndRecover(err_type), false);
std::this_thread::sleep_for(std::chrono::milliseconds(1100));
EXPECT_EQ(m.CheckAndRecover(err_type), false);
m.rabit_timeout_task.wait();
auto end = std::chrono::system_clock::now();
std::chrono::duration<double> diff = end-start;
EXPECT_EQ(m.rabit_timeout_task.get(), false);
// expect second error don't overwrite/reset timeout task
EXPECT_LT(diff.count(), 2);
}
TEST(allreduce_robust, no_timeout_shut_down)
{
rabit::engine::AllreduceRobust m;
std::string rabit_timeout = "rabit_timeout=1";
char cmd[rabit_timeout.size()+1];
std::copy(rabit_timeout.begin(), rabit_timeout.end(), cmd);
cmd[rabit_timeout.size()] = '\0';
std::string rabit_timeout_sec = "rabit_timeout_sec=1";
char cmd1[rabit_timeout_sec.size()+1];
std::copy(rabit_timeout_sec.begin(), rabit_timeout_sec.end(), cmd1);
cmd1[rabit_timeout_sec.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd2[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd2);
cmd2[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd1,cmd2};
m.Init(3, argv);
m.rank = 0;
EXPECT_EQ(m.CheckAndRecover(succ_type), true);
std::this_thread::sleep_for(std::chrono::milliseconds(10));
m.Shutdown();
}
TEST(allreduce_robust, shut_down_before_timeout)
{
rabit::engine::AllreduceRobust m;
std::string rabit_timeout = "rabit_timeout=1";
char cmd[rabit_timeout.size()+1];
std::copy(rabit_timeout.begin(), rabit_timeout.end(), cmd);
cmd[rabit_timeout.size()] = '\0';
std::string rabit_timeout_sec = "rabit_timeout_sec=1";
char cmd1[rabit_timeout_sec.size()+1];
std::copy(rabit_timeout_sec.begin(), rabit_timeout_sec.end(), cmd1);
cmd1[rabit_timeout_sec.size()] = '\0';
std::string rabit_debug = "rabit_debug=1";
char cmd2[rabit_debug.size()+1];
std::copy(rabit_debug.begin(), rabit_debug.end(), cmd2);
cmd2[rabit_debug.size()] = '\0';
char* argv[] = {cmd, cmd1,cmd2};
m.Init(3, argv);
m.rank = 0;
rabit::engine::AllreduceRobust::LinkRecord a;
m.err_link = &a;
EXPECT_EQ(m.CheckAndRecover(err_type), false);
std::this_thread::sleep_for(std::chrono::milliseconds(10));
m.Shutdown();
}

18
rabit/test/cpp/test_io.cc Normal file
View File

@@ -0,0 +1,18 @@
/*!
* Copyright (c) 2019 by Contributors
*/
#include <gtest/gtest.h>
#include <rabit/internal/io.h>
#include <vector>
namespace rabit {
TEST(MemoryFixSizeBuffer, Seek) {
size_t constexpr kSize { 64 };
std::vector<int32_t> memory( kSize );
utils::MemoryFixSizeBuffer buf(memory.data(), memory.size());
buf.Seek(utils::MemoryFixSizeBuffer::SeekEnd);
size_t end = buf.Tell();
ASSERT_EQ(end, kSize);
}
} // namespace rabit

View File

@@ -0,0 +1,8 @@
#include "gtest/gtest.h"
int main(int argc, char** argv)
{
::testing::InitGoogleTest(&argc, argv);
::testing::FLAGS_gtest_death_test_style = "threadsafe";
return RUN_ALL_TESTS();
}

View File

@@ -0,0 +1,6 @@
#include <gtest/gtest.h>
#include <rabit/internal/utils.h>
TEST(Utils, Assert) {
EXPECT_THROW({rabit::utils::Assert(false, "foo");}, dmlc::Error);
}

125
rabit/test/lazy_recover.cc Normal file
View File

@@ -0,0 +1,125 @@
// this is a test case to test whether rabit can recover model when
// facing an exception
#include <rabit/rabit.h>
#include <cstdio>
#include <cstdlib>
#include <cmath>
using namespace rabit;
// dummy model
class Model : public rabit::Serializable {
public:
// iterations
std::vector<float> data;
// load from stream
virtual void Load(rabit::Stream *fi) {
fi->Read(&data);
}
/*! \brief save the model to the stream */
virtual void Save(rabit::Stream *fo) const {
fo->Write(data);
}
virtual void InitModel(size_t n) {
data.clear();
data.resize(n, 1.0f);
}
};
inline void TestMax(Model *model, int ntrial, int iter) {
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
const int z = iter + 111;
std::vector<float> ndata(model->data.size());
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % z + model->data[i];
}
rabit::Allreduce<op::Max>(&ndata[0], ndata.size());
for (size_t i = 0; i < ndata.size(); ++i) {
float rmax = (i * 1) % z + model->data[i];
for (int r = 0; r < nproc; ++r) {
rmax = std::max(rmax, (float)((i * (r+1)) % z) + model->data[i]);
}
utils::Check(rmax == ndata[i], "[%d] TestMax check failurem i=%lu, rmax=%f, ndata=%f", rank, i, rmax, ndata[i]);
}
}
inline void TestSum(Model *model, int ntrial, int iter) {
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
const int z = 131 + iter;
std::vector<float> ndata(model->data.size());
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % z + model->data[i];
}
Allreduce<op::Sum>(&ndata[0], ndata.size());
for (size_t i = 0; i < ndata.size(); ++i) {
float rsum = model->data[i] * nproc;
for (int r = 0; r < nproc; ++r) {
rsum += (float)((i * (r+1)) % z);
}
utils::Check(fabsf(rsum - ndata[i]) < 1e-5 ,
"[%d] TestSum check failure, local=%g, allreduce=%g", rank, rsum, ndata[i]);
}
model->data = ndata;
}
inline void TestBcast(size_t n, int root, int ntrial, int iter) {
int rank = rabit::GetRank();
std::string s; s.resize(n);
for (size_t i = 0; i < n; ++i) {
s[i] = char(i % 126 + 1);
}
std::string res;
if (root == rank) {
res = s;
rabit::Broadcast(&res, root);
} else {
rabit::Broadcast(&res, root);
}
utils::Check(res == s, "[%d] TestBcast fail", rank);
}
int main(int argc, char *argv[]) {
if (argc < 3) {
printf("Usage: <ndata> <config>\n");
return 0;
}
int n = atoi(argv[1]);
rabit::Init(argc, argv);
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
std::string name = rabit::GetProcessorName();
Model model;
srand(0);
int ntrial = 0;
for (int i = 1; i < argc; ++i) {
int n;
if (sscanf(argv[i], "rabit_num_trial=%d", &n) == 1) ntrial = n;
}
int iter = rabit::LoadCheckPoint(&model);
if (iter == 0) {
model.InitModel(n);
printf("[%d] reload-trail=%d, init iter=%d\n", rank, ntrial, iter);
} else {
printf("[%d] reload-trail=%d, init iter=%d\n", rank, ntrial, iter);
}
for (int r = iter; r < 3; ++r) {
TestMax(&model, ntrial, r);
printf("[%d] !!!TestMax pass, iter=%d\n", rank, r);
int step = std::max(nproc / 3, 1);
for (int i = 0; i < nproc; i += step) {
TestBcast(n, i, ntrial, r);
}
printf("[%d] !!!TestBcast pass, iter=%d\n", rank, r);
TestSum(&model, ntrial, r);
printf("[%d] !!!TestSum pass, iter=%d\n", rank, r);
rabit::LazyCheckPoint(&model);
printf("[%d] !!!CheckPoint pass, iter=%d\n", rank, r);
}
rabit::Finalize();
return 0;
}

137
rabit/test/local_recover.cc Normal file
View File

@@ -0,0 +1,137 @@
// this is a test case to test whether rabit can recover model when
// facing an exception
#include <rabit/rabit.h>
#include <cstdio>
#include <cstdlib>
#include <cmath>
using namespace rabit;
// dummy model
class Model : public rabit::Serializable {
public:
// iterations
std::vector<float> data;
// load from stream
virtual void Load(rabit::Stream *fi) {
fi->Read(&data);
}
/*! \brief save the model to the stream */
virtual void Save(rabit::Stream *fo) const {
fo->Write(data);
}
virtual void InitModel(size_t n, float v) {
data.clear();
data.resize(n, v);
}
};
inline void TestMax(Model *model, Model *local, int ntrial, int iter) {
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
const int z = iter + 111;
std::vector<float> ndata(model->data.size());
rabit::Allreduce<op::Max>(&ndata[0], ndata.size(),
[&]() {
// use lambda expression to prepare the data
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % z + local->data[i];
}
});
for (size_t i = 0; i < ndata.size(); ++i) {
float rmax = (i * 1) % z + model->data[i];
for (int r = 0; r < nproc; ++r) {
rmax = std::max(rmax, (float)((i * (r+1)) % z) + model->data[i] + r);
}
utils::Check(rmax == ndata[i], "[%d] TestMax check failure", rank);
}
model->data = ndata;
local->data = ndata;
for (size_t i = 0; i < ndata.size(); ++i) {
local->data[i] = ndata[i] + rank;
}
}
inline void TestSum(Model *model, Model *local, int ntrial, int iter) {
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
const int z = 131 + iter;
std::vector<float> ndata(model->data.size());
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % z + local->data[i];
}
Allreduce<op::Sum>(&ndata[0], ndata.size());
for (size_t i = 0; i < ndata.size(); ++i) {
float rsum = 0.0f;
for (int r = 0; r < nproc; ++r) {
rsum += (float)((i * (r+1)) % z) + model->data[i] + r;
}
utils::Check(fabsf(rsum - ndata[i]) < 1e-5 ,
"[%d] TestSum check failure, local=%g, allreduce=%g", rank, rsum, ndata[i]);
}
model->data = ndata;
for (size_t i = 0; i < ndata.size(); ++i) {
local->data[i] = ndata[i] + rank;
}
}
inline void TestBcast(size_t n, int root, int ntrial, int iter) {
int rank = rabit::GetRank();
std::string s; s.resize(n);
for (size_t i = 0; i < n; ++i) {
s[i] = char(i % 126 + 1);
}
std::string res;
if (root == rank) {
res = s;
rabit::Broadcast(&res, root);
} else {
rabit::Broadcast(&res, root);
}
utils::Check(res == s, "[%d] TestBcast fail", rank);
}
int main(int argc, char *argv[]) {
if (argc < 3) {
printf("Usage: <ndata>\n");
return 0;
}
int n = atoi(argv[1]);
rabit::Init(argc, argv);
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
std::string name = rabit::GetProcessorName();
Model model, local;
srand(0);
int ntrial = 0;
for (int i = 1; i < argc; ++i) {
int n;
if (sscanf(argv[i], "repeat=%d", &n) == 1) ntrial = n;
}
int iter = rabit::LoadCheckPoint(&model, &local);
if (iter == 0) {
model.InitModel(n, 1.0f);
local.InitModel(n, 1.0f + rank);
printf("[%d] reload-trail=%d, init iter=%d\n", rank, ntrial, iter);
} else {
printf("[%d] reload-trail=%d, init iter=%d\n", rank, ntrial, iter);
}
for (int r = iter; r < 3; ++r) {
TestMax(&model, &local, ntrial, r);
printf("[%d] !!!TestMax pass, iter=%d\n", rank, r);
int step = std::max(nproc / 3, 1);
for (int i = 0; i < nproc; i += step) {
TestBcast(n, i, ntrial, r);
}
printf("[%d] !!!TestBcast pass, iter=%d\n", rank, r);
TestSum(&model, &local, ntrial, r);
printf("[%d] !!!TestSum pass, iter=%d\n", rank, r);
rabit::CheckPoint(&model, &local);
printf("[%d] !!!CheckPoint pass, iter=%d\n", rank, r);
}
rabit::Finalize();
return 0;
}

32
rabit/test/local_recover.py Executable file
View File

@@ -0,0 +1,32 @@
#!/usr/bin/env python3
from __future__ import print_function
from builtins import range
import sys
sys.path.append('../python')
import rabit
import numpy as np
rabit.init(lib='mock')
rank = rabit.get_rank()
n = 10
nround = 3
data = np.ones(n) * rank
version, model, local = rabit.load_checkpoint(True)
if version == 0:
model = np.zeros(n)
local = np.ones(n)
else:
print('[%d] restart from version %d' % (rank, version))
for i in range(version, nround):
res = rabit.allreduce(data + model+local, rabit.SUM)
print('[%d] iter=%d: %s' % (rank, i, str(res)))
model = res
local[:] = i
rabit.checkpoint(model, local)
rabit.finalize()

157
rabit/test/model_recover.cc Normal file
View File

@@ -0,0 +1,157 @@
// this is a test case to test whether rabit can recover model when
// facing an exception
#include <rabit/rabit.h>
#include <cstdio>
#include <cstdlib>
#include <cmath>
using namespace rabit;
// dummy model
class Model : public rabit::Serializable {
public:
// iterations
std::vector<float> data;
// load from stream
virtual void Load(rabit::Stream *fi) {
fi->Read(&data);
}
/*! \brief save the model to the stream */
virtual void Save(rabit::Stream *fo) const {
fo->Write(data);
}
virtual void InitModel(size_t n) {
data.clear();
data.resize(n, 1.0f);
}
};
inline void TestMax(Model *model, int iter) {
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
const int z = iter + 111;
std::vector<float> ndata(model->data.size());
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % z + model->data[i];
}
rabit::Allreduce<op::Max>(&ndata[0], ndata.size());
for (size_t i = 0; i < ndata.size(); ++i) {
float rmax = (i * 1) % z + model->data[i];
for (int r = 0; r < nproc; ++r) {
rmax = std::max(rmax, (float)((i * (r+1)) % z) + model->data[i]);
}
utils::Check(rmax == ndata[i], "[%d] TestMax check failurem i=%lu, rmax=%f, ndata=%f", rank, i, rmax, ndata[i]);
}
model->data = ndata;
}
inline void TestSum(Model *model, int iter) {
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
const int z = 131 + iter;
std::vector<float> ndata(model->data.size());
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % z + model->data[i];
}
Allreduce<op::Sum>(&ndata[0], ndata.size());
for (size_t i = 0; i < ndata.size(); ++i) {
float rsum = model->data[i] * nproc;
for (int r = 0; r < nproc; ++r) {
rsum += (float)((i * (r+1)) % z);
}
utils::Check(fabsf(rsum - ndata[i]) < 1e-5 ,
"[%d] TestSum check failure, local=%g, allreduce=%g", rank, rsum, ndata[i]);
}
model->data = ndata;
}
inline void TestAllgather(Model *model, int iter) {
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
const int z = 131 + iter;
std::vector<float> ndata(model->data.size() * nproc);
size_t beginSlice = rank * model->data.size();
for (size_t i = 0; i < model->data.size(); ++i) {
ndata[beginSlice + i] = (i * (rank+1)) % z + model->data[i];
}
Allgather(&ndata[0], ndata.size(), beginSlice,
model->data.size(), model->data.size());
for (size_t i = 0; i < ndata.size(); ++i) {
int curRank = i / model->data.size();
int remainder = i % model->data.size();
float data = (remainder * (curRank+1)) % z + model->data[remainder];
utils::Check(fabsf(data - ndata[i]) < 1e-5 ,
"[%d] TestAllgather check failure, local=%g, allgatherring=%g", rank, data, ndata[i]);
}
model->data = ndata;
}
inline void TestBcast(size_t n, int root) {
int rank = rabit::GetRank();
std::string s; s.resize(n);
for (size_t i = 0; i < n; ++i) {
s[i] = char(i % 126 + 1);
}
std::string res;
if (root == rank) {
res = s;
}
rabit::Broadcast(&res, root);
utils::Check(res == s, "[%d] TestBcast fail", rank);
}
int main(int argc, char *argv[]) {
if (argc < 3) {
printf("Usage: <ndata> <config>\n");
return 0;
}
int n = atoi(argv[1]);
rabit::Init(argc, argv);
int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
std::string name = rabit::GetProcessorName();
int max_rank = rank;
rabit::Allreduce<op::Max>(&max_rank, 1);
utils::Check(max_rank == nproc - 1, "max rank is world size-1");
Model model;
srand(0);
int ntrial = 0;
for (int i = 1; i < argc; ++i) {
int n;
if (sscanf(argv[i], "rabit_num_trial=%d", &n) == 1) ntrial = n;
}
int iter = rabit::LoadCheckPoint(&model);
if (iter == 0) {
model.InitModel(n);
}
printf("[%d] reload-trail=%d, init iter=%d\n", rank, ntrial, iter);
for (int r = iter; r < 3; ++r) {
TestMax(&model, r);
printf("[%d] !!!TestMax pass, iter=%d\n", rank, r);
int step = std::max(nproc / 3, 1);
for (int i = 0; i < nproc; i += step) {
TestBcast(n, i);
}
printf("[%d] !!!TestBcast pass, iter=%d\n", rank, r);
TestSum(&model, r);
printf("[%d] !!!TestSum pass, iter=%d\n", rank, r);
TestAllgather(&model, r);
printf("[%d] !!!TestAllgather pass, iter=%d\n", rank, r);
rabit::CheckPoint(&model);
printf("[%d] !!!Checkpoint pass, iter=%d\n", rank, r);
}
rabit::Finalize();
return 0;
}

View File

@@ -0,0 +1,34 @@
import os
import argparse
import sys
def main():
parser = argparse.ArgumentParser(description='TODO')
parser.add_argument('-ho', '--host_dir', required=True)
parser.add_argument('-s', '--submit_script', required=True)
parser.add_argument('-rex', '--rabit_exec', required=True)
parser.add_argument('-mpi', '--mpi_exec', required=True)
args = parser.parse_args()
ndata = [10**4, 10**5, 10**6, 10**7]
nrepeat = [10**4, 10**3, 10**2, 10]
machines = [2,4,8,16,31]
executables = [args.rabit_exec, args.mpi_exec]
for executable in executables:
sys.stderr.write('Executable %s' % executable)
sys.stderr.flush()
for i, data in enumerate(ndata):
for machine in machines:
host_file = os.path.join(args.host_dir, 'hosts%d' % machine)
cmd = 'python %s %d %s %s %d %d' % (args.submit_script, machine, host_file, executable, data, nrepeat[i])
sys.stderr.write('data=%d, repeat=%d, machine=%d\n' % (data, nrepeat[i], machine))
sys.stderr.flush()
os.system(cmd)
sys.stderr.write('\n')
sys.stderr.flush()
if __name__ == "__main__":
main()

99
rabit/test/speed_test.cc Normal file
View File

@@ -0,0 +1,99 @@
// This program is used to test the speed of rabit API
#include <rabit/rabit.h>
#include <rabit/internal/timer.h>
#include <cstdio>
#include <cstdlib>
#include <cmath>
#include <time.h>
using namespace rabit;
double max_tdiff, sum_tdiff, bcast_tdiff, tot_tdiff;
inline void TestMax(size_t n) {
int rank = rabit::GetRank();
std::vector<float> ndata(n);
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % 111;
}
double tstart = utils::GetTime();
rabit::Allreduce<op::Max>(&ndata[0], ndata.size());
max_tdiff += utils::GetTime() - tstart;
}
inline void TestSum(size_t n) {
int rank = rabit::GetRank();
const int z = 131;
std::vector<float> ndata(n);
for (size_t i = 0; i < ndata.size(); ++i) {
ndata[i] = (i * (rank+1)) % z;
}
double tstart = utils::GetTime();
rabit::Allreduce<op::Sum>(&ndata[0], ndata.size());
sum_tdiff += utils::GetTime() - tstart;
}
inline void TestBcast(size_t n, int root) {
int rank = rabit::GetRank();
std::string s; s.resize(n);
for (size_t i = 0; i < n; ++i) {
s[i] = char(i % 126 + 1);
}
std::string res;
res.resize(n);
if (root == rank) {
res = s;
}
double tstart = utils::GetTime();
rabit::Broadcast(&res[0], res.length(), root);
bcast_tdiff += utils::GetTime() - tstart;
}
inline void PrintStats(const char *name, double tdiff, int n, int nrep, size_t size) {
int nproc = rabit::GetWorldSize();
double tsum = tdiff;
rabit::Allreduce<op::Sum>(&tsum, 1);
double tavg = tsum / nproc;
double tsqr = tdiff - tavg;
tsqr *= tsqr;
rabit::Allreduce<op::Sum>(&tsqr, 1);
double tstd = sqrt(tsqr / nproc);
if (rabit::GetRank() == 0) {
rabit::TrackerPrintf("%s: mean=%g, std=%g sec\n", name, tavg, tstd);
double ndata = n;
ndata *= nrep * size;
if (n != 0) {
rabit::TrackerPrintf("%s-speed: %g MB/sec\n", name, (ndata / tavg) / 1024 / 1024 );
}
}
}
int main(int argc, char *argv[]) {
if (argc < 3) {
printf("Usage: <ndata> <nrepeat>\n");
return 0;
}
srand(0);
int n = atoi(argv[1]);
int nrep = atoi(argv[2]);
utils::Check(nrep >= 1, "need to at least repeat running once");
rabit::Init(argc, argv);
//int rank = rabit::GetRank();
int nproc = rabit::GetWorldSize();
std::string name = rabit::GetProcessorName();
max_tdiff = sum_tdiff = bcast_tdiff = 0;
double tstart = utils::GetTime();
for (int i = 0; i < nrep; ++i) {
TestMax(n);
TestSum(n);
TestBcast(n, rand() % nproc);
}
tot_tdiff = utils::GetTime() - tstart;
// use allreduce to get the sum and std of time
PrintStats("max_tdiff", max_tdiff, n, nrep, sizeof(float));
PrintStats("sum_tdiff", sum_tdiff, n, nrep, sizeof(float));
PrintStats("bcast_tdiff", bcast_tdiff, n, nrep, sizeof(char));
PrintStats("tot_tdiff", tot_tdiff, 0, nrep, sizeof(float));
rabit::Finalize();
return 0;
}

37
rabit/test/test.mk Normal file
View File

@@ -0,0 +1,37 @@
RABIT_BUILD_DMLC = 0
ifeq ($(RABIT_BUILD_DMLC),1)
DMLC=../dmlc-core
else
DMLC=../../dmlc-core
endif
# this is a makefile used to show testcases of rabit
.PHONY: all
all: model_recover_10_10k model_recover_10_10k_die_same model_recover_10_10k_die_hard local_recover_10_10k lazy_recover_10_10k_die_hard lazy_recover_10_10k_die_same ringallreduce_10_10k pylocal_recover_10_10k
# this experiment test recovery with actually process exit, use keepalive to keep program alive
model_recover_10_10k:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 --local-num-attempt=20 model_recover 10000 mock=0,0,1,0 mock=1,1,1,0 rabit_bootstrap_cache=true rabit_debug=true rabit_reduce_ring_mincount=1 rabit_timeout=true rabit_timeout_sec=5
model_recover_10_10k_die_same:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 --local-num-attempt=20 model_recover 10000 mock=0,0,1,0 mock=1,1,1,0 mock=0,1,1,0 mock=4,1,1,0 mock=9,1,1,0 rabit_bootstrap_cache=1
model_recover_10_10k_die_hard:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 --local-num-attempt=20 model_recover 10000 mock=0,0,1,0 mock=1,1,1,0 mock=1,1,1,1 mock=0,1,1,0 mock=4,1,1,0 mock=9,1,1,0 mock=8,1,2,0 mock=4,1,3,0 rabit_bootstrap_cache=1
local_recover_10_10k:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 --local-num-attempt=20 local_recover 10000 mock=0,0,1,0 mock=1,1,1,0 mock=0,1,1,0 mock=4,1,1,0 mock=9,1,1,0 mock=1,1,1,1
pylocal_recover_10_10k:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 --local-num-attempt=20 local_recover.py 10000 mock=0,0,1,0 mock=1,1,1,0 mock=0,1,1,0 mock=4,1,1,0 mock=9,1,1,0 mock=1,1,1,1
lazy_recover_10_10k_die_hard:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 --local-num-attempt=20 lazy_recover 10000 mock=0,0,1,0 mock=1,1,1,0 mock=1,1,1,1 mock=0,1,1,0 mock=4,1,1,0 mock=9,1,1,0 mock=8,1,2,0 mock=4,1,3,0
lazy_recover_10_10k_die_same:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 --local-num-attempt=20 lazy_recover 10000 mock=0,0,1,0 mock=1,1,1,0 mock=0,1,1,0 mock=4,1,1,0 mock=9,1,1,0
ringallreduce_10_10k:
python $(DMLC)/tracker/dmlc-submit --cluster local --num-workers=10 model_recover 10000 rabit_reduce_ring_mincount=10