Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Drop single point model recovery (#6262)

Browse Source 
* Pass rabit params in JVM package.
* Implement timeout using poll timeout parameter.
* Remove OOB data check.
This commit is contained in:
Jiaming Yuan
2020-10-21 15:27:03 +08:00
committed by GitHub
parent 81c37c28d5
commit b5c2a47b20
22 changed files with 63 additions and 2879 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
rabit/src/CMakeLists.txt
View File

@@ -1,31 +0,0 @@
option(DMLC_ROOT "Specify root of external dmlc core.")
add_library(allreduce_base "")
add_library(allreduce_mock "")
target_sources(
allreduce_base
PRIVATE
allreduce_base.cc
PUBLIC
${CMAKE_CURRENT_LIST_DIR}/allreduce_base.h
)
target_sources(
allreduce_mock
PRIVATE
allreduce_robust.cc
PUBLIC
${CMAKE_CURRENT_LIST_DIR}/allreduce_mock.h
)
target_include_directories(
allreduce_base
PUBLIC
${DMLC_ROOT}/include
${CMAKE_CURRENT_LIST_DIR}/../../include)
target_include_directories(
allreduce_mock
PUBLIC
${DMLC_ROOT}/include
${CMAKE_CURRENT_LIST_DIR}/../../include)

6
rabit/src/README.md
View File

@@ -1,6 +0,0 @@
Source Files of Rabit
====
* This folder contains the source files of rabit library
* The library headers are in folder [include](../include)
* The .h files in this folder are internal header files that are only used by rabit and will not be seen by users

39
rabit/src/allreduce_base.cc
View File

@@ -6,8 +6,9 @@
* \author Tianqi Chen, Ignacio Cano, Tianyi Zhou
*/
#define NOMINMAX
#include "rabit/base.h"
#include "rabit/internal/rabit-inl.h"
#include "allreduce_base.h"
#include <rabit/base.h>
#ifndef _WIN32
#include <netinet/tcp.h>
@@ -208,8 +209,8 @@ void AllreduceBase::SetParam(const char *name, const char *val) {
rabit_timeout = utils::StringToBool(val);
}
if (!strcmp(name, "rabit_timeout_sec")) {
timeout_sec = atoi(val);
utils::Assert(timeout_sec >= 0, "rabit_timeout_sec should be non negative second");
timeout_sec = std::chrono::seconds(atoi(val));
utils::Assert(timeout_sec.count() >= 0, "rabit_timeout_sec should be non negative second");
}
if (!strcmp(name, "rabit_enable_tcp_no_delay")) {
if (!strcmp(val, "true")) {
@@ -549,14 +550,7 @@ AllreduceBase::TryAllreduceTree(void *sendrecvbuf_,
// finish runing allreduce
if (finished) break;
// select must return
watcher.Poll();
// exception handling
for (int i = 0; i < nlink; ++i) {
// recive OOB message from some link
if (watcher.CheckExcept(links[i].sock)) {
return ReportError(&links[i], kGetExcept);
}
}
watcher.Poll(timeout_sec);
// read data from childs
for (int i = 0; i < nlink; ++i) {
if (i != parent_index && watcher.CheckRead(links[i].sock)) {
@@ -729,14 +723,7 @@ AllreduceBase::TryBroadcast(void *sendrecvbuf_, size_t total_size, int root) {
// finish running
if (finished) break;
// select
watcher.Poll();
// exception handling
for (int i = 0; i < nlink; ++i) {
// recive OOB message from some link
if (watcher.CheckExcept(links[i].sock)) {
return ReportError(&links[i], kGetExcept);
}
}
watcher.Poll(timeout_sec);
if (in_link == -2) {
// probe in-link
for (int i = 0; i < nlink; ++i) {
@@ -819,7 +806,7 @@ AllreduceBase::TryAllgatherRing(void *sendrecvbuf_, size_t total_size,
finished = false;
}
if (finished) break;
watcher.Poll();
watcher.Poll(timeout_sec);
if (read_ptr != stop_read && watcher.CheckRead(next.sock)) {
size_t size = stop_read - read_ptr;
size_t start = read_ptr % total_size;
@@ -831,7 +818,10 @@ AllreduceBase::TryAllgatherRing(void *sendrecvbuf_, size_t total_size,
read_ptr += static_cast<size_t>(len);
} else {
ReturnType ret = Errno2Return();
if (ret != kSuccess) return ReportError(&next, ret);
if (ret != kSuccess) {
auto err = ReportError(&next, ret);
return err;
}
}
}
if (write_ptr < read_ptr && write_ptr != stop_write) {
@@ -845,7 +835,10 @@ AllreduceBase::TryAllgatherRing(void *sendrecvbuf_, size_t total_size,
write_ptr += static_cast<size_t>(len);
} else {
ReturnType ret = Errno2Return();
if (ret != kSuccess) return ReportError(&prev, ret);
if (ret != kSuccess) {
auto err = ReportError(&prev, ret);
return err;
}
}
}
}
@@ -913,7 +906,7 @@ AllreduceBase::TryReduceScatterRing(void *sendrecvbuf_,
finished = false;
}
if (finished) break;
watcher.Poll();
watcher.Poll(timeout_sec);
if (read_ptr != stop_read && watcher.CheckRead(next.sock)) {
ReturnType ret = next.ReadToRingBuffer(reduce_ptr, stop_read);
if (ret != kSuccess) {

21
rabit/src/allreduce_base.h
View File

@@ -12,6 +12,8 @@
#ifndef RABIT_ALLREDUCE_BASE_H_
#define RABIT_ALLREDUCE_BASE_H_
#include <functional>
#include <future>
#include <vector>
#include <string>
#include <algorithm>
@@ -35,6 +37,7 @@ class Datatype {
}
namespace rabit {
namespace engine {
/*! \brief implementation of basic Allreduce engine */
class AllreduceBase : public IEngine {
public:
@@ -103,9 +106,11 @@ class AllreduceBase : public IEngine {
size_t slice_end, size_t size_prev_slice,
const char *_file = _FILE, const int _line = _LINE,
const char *_caller = _CALLER) override {
if (world_size == 1 || world_size == -1) return;
utils::Assert(TryAllgatherRing(sendrecvbuf_, total_size,
slice_begin, slice_end, size_prev_slice) == kSuccess,
if (world_size == 1 || world_size == -1) {
return;
}
utils::Assert(TryAllgatherRing(sendrecvbuf_, total_size, slice_begin,
slice_end, size_prev_slice) == kSuccess,
"AllgatherRing failed");
}
/*!
@@ -130,8 +135,8 @@ class AllreduceBase : public IEngine {
const char *_caller = _CALLER) override {
if (prepare_fun != nullptr) prepare_fun(prepare_arg);
if (world_size == 1 || world_size == -1) return;
utils::Assert(TryAllreduce(sendrecvbuf_,
type_nbytes, count, reducer) == kSuccess,
utils::Assert(TryAllreduce(sendrecvbuf_, type_nbytes, count, reducer) ==
kSuccess,
"Allreduce failed");
}
/*!
@@ -518,9 +523,9 @@ class AllreduceBase : public IEngine {
//---- data structure related to model ----
// call sequence counter, records how many calls we made so far
// from last call to CheckPoint, LoadCheckPoint
int seq_counter; // NOLINT
int seq_counter{0}; // NOLINT
// version number of model
int version_number; // NOLINT
int version_number {0}; // NOLINT
// whether the job is running in hadoop
bool hadoop_mode; // NOLINT
//---- local data related to link ----
@@ -571,7 +576,7 @@ class AllreduceBase : public IEngine {
// enable detailed logging
bool rabit_debug = false; // NOLINT
// by default, if rabit worker not recover in half an hour exit
int timeout_sec = 1800; // NOLINT
std::chrono::seconds timeout_sec{std::chrono::seconds{1800}}; // NOLINT
// flag to enable rabit_timeout
bool rabit_timeout = false; // NOLINT
// Enable TCP node delay

33
rabit/src/allreduce_mock.h
View File

@@ -13,11 +13,11 @@
#include <sstream>
#include "rabit/internal/engine.h"
#include "rabit/internal/timer.h"
#include "allreduce_robust.h"
#include "allreduce_base.h"
namespace rabit {
namespace engine {
class AllreduceMock : public AllreduceRobust {
class AllreduceMock : public AllreduceBase {
public:
// constructor
AllreduceMock() {
@@ -30,7 +30,7 @@ class AllreduceMock : public AllreduceRobust {
// destructor
~AllreduceMock() override = default;
void SetParam(const char *name, const char *val) override {
AllreduceRobust::SetParam(name, val);
AllreduceBase::SetParam(name, val);
// additional parameters
if (!strcmp(name, "rabit_num_trial")) num_trial_ = atoi(val);
if (!strcmp(name, "DMLC_NUM_ATTEMPT")) num_trial_ = atoi(val);
@@ -51,9 +51,8 @@ class AllreduceMock : public AllreduceRobust {
const char *_caller = _CALLER) override {
this->Verify(MockKey(rank, version_number, seq_counter, num_trial_), "AllReduce");
double tstart = utils::GetTime();
AllreduceRobust::Allreduce(sendrecvbuf_, type_nbytes,
count, reducer, prepare_fun, prepare_arg,
_file, _line, _caller);
AllreduceBase::Allreduce(sendrecvbuf_, type_nbytes, count, reducer,
prepare_fun, prepare_arg, _file, _line, _caller);
tsum_allreduce_ += utils::GetTime() - tstart;
}
void Allgather(void *sendrecvbuf, size_t total_size, size_t slice_begin,
@@ -62,16 +61,15 @@ class AllreduceMock : public AllreduceRobust {
const char *_caller = _CALLER) override {
this->Verify(MockKey(rank, version_number, seq_counter, num_trial_), "Allgather");
double tstart = utils::GetTime();
AllreduceRobust::Allgather(sendrecvbuf, total_size,
slice_begin, slice_end,
size_prev_slice, _file, _line, _caller);
AllreduceBase::Allgather(sendrecvbuf, total_size, slice_begin, slice_end,
size_prev_slice, _file, _line, _caller);
tsum_allgather_ += utils::GetTime() - tstart;
}
void Broadcast(void *sendrecvbuf_, size_t total_size, int root,
const char *_file = _FILE, const int _line = _LINE,
const char *_caller = _CALLER) override {
this->Verify(MockKey(rank, version_number, seq_counter, num_trial_), "Broadcast");
AllreduceRobust::Broadcast(sendrecvbuf_, total_size, root, _file, _line, _caller);
AllreduceBase::Broadcast(sendrecvbuf_, total_size, root, _file, _line, _caller);
}
int LoadCheckPoint(Serializable *global_model,
Serializable *local_model) override {
@@ -79,11 +77,11 @@ class AllreduceMock : public AllreduceRobust {
tsum_allgather_ = 0.0;
time_checkpoint_ = utils::GetTime();
if (force_local_ == 0) {
return AllreduceRobust::LoadCheckPoint(global_model, local_model);
return AllreduceBase::LoadCheckPoint(global_model, local_model);
} else {
DummySerializer dum;
ComboSerializer com(global_model, local_model);
return AllreduceRobust::LoadCheckPoint(&dum, &com);
return AllreduceBase::LoadCheckPoint(&dum, &com);
}
}
void CheckPoint(const Serializable *global_model,
@@ -92,18 +90,17 @@ class AllreduceMock : public AllreduceRobust {
double tstart = utils::GetTime();
double tbet_chkpt = tstart - time_checkpoint_;
if (force_local_ == 0) {
AllreduceRobust::CheckPoint(global_model, local_model);
AllreduceBase::CheckPoint(global_model, local_model);
} else {
DummySerializer dum;
ComboSerializer com(global_model, local_model);
AllreduceRobust::CheckPoint(&dum, &com);
AllreduceBase::CheckPoint(&dum, &com);
}
time_checkpoint_ = utils::GetTime();
double tcost = utils::GetTime() - tstart;
if (report_stats_ != 0 && rank == 0) {
std::stringstream ss;
ss << "[v" << version_number << "] global_size=" << global_checkpoint_.length()
<< ",local_size=" << (local_chkpt_[0].length() + local_chkpt_[1].length())
ss << "[v" << version_number << "] global_size="
<< ",check_tcost="<< tcost <<" sec"
<< ",allreduce_tcost=" << tsum_allreduce_ << " sec"
<< ",allgather_tcost=" << tsum_allgather_ << " sec"
@@ -116,7 +113,7 @@ class AllreduceMock : public AllreduceRobust {
void LazyCheckPoint(const Serializable *global_model) override {
this->Verify(MockKey(rank, version_number, seq_counter, num_trial_), "LazyCheckPoint");
AllreduceRobust::LazyCheckPoint(global_model);
AllreduceBase::LazyCheckPoint(global_model);
}
protected:
@@ -186,7 +183,7 @@ class AllreduceMock : public AllreduceRobust {
if (mock_map_.count(key) != 0) {
num_trial_ += 1;
// data processing frameworks runs on shared process
error_("[%d]@@@Hit Mock Error:%s ", rank, name);
throw dmlc::Error(std::to_string(rank) + "@@@Hit Mock Error: " + name);
}
}
};

169
rabit/src/allreduce_robust-inl.h
View File

@@ -1,169 +0,0 @@
/*!
* Copyright (c) 2014 by Contributors
* \file allreduce_robust-inl.h
* \brief implementation of inline template function in AllreduceRobust
*
* \author Tianqi Chen
*/
#ifndef RABIT_ALLREDUCE_ROBUST_INL_H_
#define RABIT_ALLREDUCE_ROBUST_INL_H_
#include <vector>
namespace rabit {
namespace engine {
/*!
* \brief run message passing algorithm on the allreduce tree
* the result is edge message stored in p_edge_in and p_edge_out
* \param node_value the value associated with current node
* \param p_edge_in used to store input message from each of the edge
* \param p_edge_out used to store output message from each of the edge
* \param func a function that defines the message passing rule
* Parameters of func:
* - node_value same as node_value in the main function
* - edge_in the array of input messages from each edge,
* this includes the output edge, which should be excluded
* - out_index array the index of output edge, the function should
* exclude the output edge when compute the message passing value
* Return of func:
* the function returns the output message based on the input message and node_value
*
* \tparam EdgeType type of edge message, must be simple struct
* \tparam NodeType type of node value
*/
template<typename NodeType, typename EdgeType>
inline AllreduceRobust::ReturnType
AllreduceRobust::MsgPassing(const NodeType &node_value,
std::vector<EdgeType> *p_edge_in,
std::vector<EdgeType> *p_edge_out,
EdgeType(*func)
(const NodeType &node_value,
const std::vector<EdgeType> &edge_in,
size_t out_index)) {
RefLinkVector &links = tree_links;
if (links.Size() == 0) return kSuccess;
// number of links
const int nlink = static_cast<int>(links.Size());
// initialize the pointers
for (int i = 0; i < nlink; ++i) {
links[i].ResetSize();
}
std::vector<EdgeType> &edge_in = *p_edge_in;
std::vector<EdgeType> &edge_out = *p_edge_out;
edge_in.resize(nlink);
edge_out.resize(nlink);
// stages in the process
// 0: recv messages from childs
// 1: send message to parent
// 2: recv message from parent
// 3: send message to childs
int stage = 0;
// if no childs, no need to, directly start passing message
if (nlink == static_cast<int>(parent_index != -1)) {
utils::Assert(parent_index == 0, "parent must be 0");
edge_out[parent_index] = func(node_value, edge_in, parent_index);
stage = 1;
}
// while we have not passed the messages out
while (true) {
// for node with no parent, directly do stage 3
if (parent_index == -1) {
utils::Assert(stage != 2 && stage != 1, "invalie stage id");
}
// poll helper
utils::PollHelper watcher;
bool done = (stage == 3);
for (int i = 0; i < nlink; ++i) {
watcher.WatchException(links[i].sock);
switch (stage) {
case 0:
if (i != parent_index && links[i].size_read != sizeof(EdgeType)) {
watcher.WatchRead(links[i].sock);
}
break;
case 1:
if (i == parent_index) {
watcher.WatchWrite(links[i].sock);
}
break;
case 2:
if (i == parent_index) {
watcher.WatchRead(links[i].sock);
}
break;
case 3:
if (i != parent_index && links[i].size_write != sizeof(EdgeType)) {
watcher.WatchWrite(links[i].sock);
done = false;
}
break;
default: utils::Error("invalid stage");
}
}
// finish all the stages, and write out message
if (done) break;
watcher.Poll();
// exception handling
for (int i = 0; i < nlink; ++i) {
// recive OOB message from some link
if (watcher.CheckExcept(links[i].sock)) {
return ReportError(&links[i], kGetExcept);
}
}
if (stage == 0) {
bool finished = true;
// read data from childs
for (int i = 0; i < nlink; ++i) {
if (i != parent_index) {
if (watcher.CheckRead(links[i].sock)) {
ReturnType ret = links[i].ReadToArray(&edge_in[i], sizeof(EdgeType));
if (ret != kSuccess) return ReportError(&links[i], ret);
}
if (links[i].size_read != sizeof(EdgeType)) finished = false;
}
}
// if no parent, jump to stage 3, otherwise do stage 1
if (finished) {
if (parent_index != -1) {
edge_out[parent_index] = func(node_value, edge_in, parent_index);
stage = 1;
} else {
for (int i = 0; i < nlink; ++i) {
edge_out[i] = func(node_value, edge_in, i);
}
stage = 3;
}
}
}
if (stage == 1) {
const int pid = this->parent_index;
utils::Assert(pid != -1, "MsgPassing invalid stage");
ReturnType ret = links[pid].WriteFromArray(&edge_out[pid], sizeof(EdgeType));
if (ret != kSuccess) return ReportError(&links[pid], ret);
if (links[pid].size_write == sizeof(EdgeType)) stage = 2;
}
if (stage == 2) {
const int pid = this->parent_index;
utils::Assert(pid != -1, "MsgPassing invalid stage");
ReturnType ret = links[pid].ReadToArray(&edge_in[pid], sizeof(EdgeType));
if (ret != kSuccess) return ReportError(&links[pid], ret);
if (links[pid].size_read == sizeof(EdgeType)) {
for (int i = 0; i < nlink; ++i) {
if (i != pid) edge_out[i] = func(node_value, edge_in, i);
}
stage = 3;
}
}
if (stage == 3) {
for (int i = 0; i < nlink; ++i) {
if (i != parent_index && links[i].size_write != sizeof(EdgeType)) {
ReturnType ret = links[i].WriteFromArray(&edge_out[i], sizeof(EdgeType));
if (ret != kSuccess) return ReportError(&links[i], ret);
}
}
}
}
return kSuccess;
}
} // namespace engine
} // namespace rabit
#endif // RABIT_ALLREDUCE_ROBUST_INL_H_

1590
rabit/src/allreduce_robust.cc
View File

File diff suppressed because it is too large Load Diff

665
rabit/src/allreduce_robust.h
View File

@@ -1,665 +0,0 @@
/*!
* Copyright (c) 2014 by Contributors
* \file allreduce_robust.h
* \brief Robust implementation of Allreduce
* using TCP non-block socket and tree-shape reduction.
*
* This implementation considers the failure of nodes
*
* \author Tianqi Chen, Ignacio Cano, Tianyi Zhou
*/
#ifndef RABIT_ALLREDUCE_ROBUST_H_
#define RABIT_ALLREDUCE_ROBUST_H_
#include <future>
#include <vector>
#include <string>
#include <algorithm>
#include "rabit/internal/engine.h"
#include "allreduce_base.h"
namespace rabit {
namespace engine {
/*! \brief implementation of fault tolerant all reduce engine */
class AllreduceRobust : public AllreduceBase {
public:
AllreduceRobust();
~AllreduceRobust() override = default;
// initialize the manager
bool Init(int argc, char* argv[]) override;
/*! \brief shutdown the engine */
bool Shutdown() override;
/*!
* \brief set parameters to the engine
* \param name parameter name
* \param val parameter value
*/
void SetParam(const char *name, const char *val) override;
/*!
* \brief perform immutable local bootstrap cache insertion
* \param key unique cache key
* \param buf buffer of allreduce/robust payload to copy
* \param buflen total number of bytes
* \return -1 if no recovery cache fetched otherwise 0
*/
int SetBootstrapCache(const std::string &key, const void *buf,
const size_t type_nbytes, const size_t count);
/*!
* \brief perform bootstrap cache lookup if nodes in fault recovery
* \param key unique cache key
* \param buf buffer for recv allreduce/robust payload
* \param buflen total number of bytes
*/
int GetBootstrapCache(const std::string &key, void *buf, const size_t type_nbytes,
const size_t count);
/*!
* \brief internal Allgather function, each node have a segment of data in the ring of sendrecvbuf,
* the data provided by current node k is [slice_begin, slice_end),
* the next node's segment must start with slice_end
* after the call of Allgather, sendrecvbuf_ contains all the contents including all segments
* use a ring based algorithm
*
* \param sendrecvbuf_ buffer for both sending and receiving data, it is a ring conceptually
* \param total_size total size of data to be gathered
* \param slice_begin beginning of the current slice
* \param slice_end end of the current slice
* \param size_prev_slice size of the previous slice i.e. slice of node (rank - 1) % world_size
* \param _file caller file name used to generate unique cache key
* \param _line caller line number used to generate unique cache key
* \param _caller caller function name used to generate unique cache key
*/
void Allgather(void *sendrecvbuf_, size_t total_size, size_t slice_begin,
size_t slice_end, size_t size_prev_slice,
const char *_file = _FILE, const int _line = _LINE,
const char *_caller = _CALLER) override;
/*!
* \brief perform in-place allreduce, on sendrecvbuf
* this function is NOT thread-safe
* \param sendrecvbuf_ buffer for both sending and recving data
* \param type_nbytes the unit number of bytes the type have
* \param count number of elements to be reduced
* \param reducer reduce function
* \param prepare_func Lazy preprocessing function, lazy prepare_fun(prepare_arg)
* will be called by the function before performing Allreduce, to intialize the data in sendrecvbuf_.
* If the result of Allreduce can be recovered directly, then prepare_func will NOT be called
* \param prepare_arg argument used to passed into the lazy preprocessing function
* \param prepare_arg argument used to passed into the lazy preprocessing function
* \param _file caller file name used to generate unique cache key
* \param _line caller line number used to generate unique cache key
* \param _caller caller function name used to generate unique cache key
*/
void Allreduce(void *sendrecvbuf_, size_t type_nbytes, size_t count,
ReduceFunction reducer, PreprocFunction prepare_fun = nullptr,
void *prepare_arg = nullptr, const char *_file = _FILE,
const int _line = _LINE,
const char *_caller = _CALLER) override;
/*!
* \brief broadcast data from root to all nodes
* \param sendrecvbuf_ buffer for both sending and recving data
* \param size the size of the data to be broadcasted
* \param root the root worker id to broadcast the data
* \param _file caller file name used to generate unique cache key
* \param _line caller line number used to generate unique cache key
* \param _caller caller function name used to generate unique cache key
*/
void Broadcast(void *sendrecvbuf_, size_t total_size, int root,
const char *_file = _FILE, const int _line = _LINE,
const char *_caller = _CALLER) override;
/*!
* \brief load latest check point
* \param global_model pointer to the globally shared model/state
* when calling this function, the caller need to gauranttees that global_model
* is the same in all nodes
* \param local_model pointer to local model, that is specific to current node/rank
* this can be NULL when no local model is needed
*
* \return the version number of check point loaded
* if returned version == 0, this means no model has been CheckPointed
* the p_model is not touched, user should do necessary initialization by themselves
*
* Common usage example:
* int iter = rabit::LoadCheckPoint(&model);
* if (iter == 0) model.InitParameters();
* for (i = iter; i < max_iter; ++i) {
* do many things, include allreduce
* rabit::CheckPoint(model);
* }
*
* \sa CheckPoint, VersionNumber
*/
int LoadCheckPoint(Serializable *global_model,
Serializable *local_model = nullptr) override;
/*!
* \brief checkpoint the model, meaning we finished a stage of execution
* every time we call check point, there is a version number which will increase by one
*
* \param global_model pointer to the globally shared model/state
* when calling this function, the caller need to gauranttees that global_model
* is the same in all nodes
* \param local_model pointer to local model, that is specific to current node/rank
* this can be NULL when no local state is needed
*
* NOTE: local_model requires explicit replication of the model for fault-tolerance, which will
* bring replication cost in CheckPoint function. global_model do not need explicit replication.
* So only CheckPoint with global_model if possible
*
* \sa LoadCheckPoint, VersionNumber
*/
void CheckPoint(const Serializable *global_model,
const Serializable *local_model = nullptr) override {
this->CheckPointImpl(global_model, local_model, false);
}
/*!
* \brief This function can be used to replace CheckPoint for global_model only,
* when certain condition is met(see detailed expplaination).
*
* This is a "lazy" checkpoint such that only the pointer to global_model is
* remembered and no memory copy is taken. To use this function, the user MUST ensure that:
* The global_model must remain unchanged util last call of Allreduce/Broadcast in current version finishs.
* In another words, global_model model can be changed only between last call of
* Allreduce/Broadcast and LazyCheckPoint in current version
*
* For example, suppose the calling sequence is:
* LazyCheckPoint, code1, Allreduce, code2, Broadcast, code3, LazyCheckPoint
*
* If user can only changes global_model in code3, then LazyCheckPoint can be used to
* improve efficiency of the program.
* \param global_model pointer to the globally shared model/state
* when calling this function, the caller need to gauranttees that global_model
* is the same in all nodes
* \sa LoadCheckPoint, CheckPoint, VersionNumber
*/
void LazyCheckPoint(const Serializable *global_model) override {
this->CheckPointImpl(global_model, nullptr, true);
}
/*!
* \brief explicitly re-init everything before calling LoadCheckPoint
* call this function when IEngine throw an exception out,
* this function is only used for test purpose
*/
void InitAfterException() override {
// simple way, shutdown all links
for (auto& link : all_links) {
if (link.sock.BadSocket()) {
link.sock.Close();
}
}
ReConnectLinks("recover");
}
protected:
// constant one byte out of band message to indicate error happening
// and mark for channel cleanup
static const char kOOBReset = 95;
// and mark for channel cleanup, after OOB signal
static const char kResetMark = 97;
// and mark for channel cleanup
static const char kResetAck = 97;
/*! \brief type of roles each node can play during recovery */
enum RecoverType {
/*! \brief current node have data */
kHaveData = 0,
/*! \brief current node request data */
kRequestData = 1,
/*! \brief current node only helps to pass data around */
kPassData = 2
};
enum SeqType {
/*! \brief apply to rabit seq code */
kSeq = 0,
/*! \brief apply to rabit cache seq code */
kCache = 1
};
/*!
* \brief summary of actions proposed in all nodes
* this data structure is used to make consensus decision
* about next action to take in the recovery mode
*/
struct ActionSummary {
// maximumly allowed sequence id
static const uint32_t kSpecialOp = (1 << 26);
// special sequence number for local state checkpoint
static const uint32_t kLocalCheckPoint = (1 << 26) - 2;
// special sequnce number for local state checkpoint ack signal
static const uint32_t kLocalCheckAck = (1 << 26) - 1;
//---------------------------------------------
// The following are bit mask of flag used in
//----------------------------------------------
// some node want to load check point
static const int kLoadCheck = 1;
// some node want to do check point
static const int kCheckPoint = 2;
// check point Ack, we use a two phase message in check point,
// this is the second phase of check pointing
static const int kCheckAck = 4;
// there are difference sequence number the nodes proposed
// this means we want to do recover execution of the lower sequence
// action instead of normal execution
static const int kDiffSeq = 8;
// there are nodes request load cache
static const int kLoadBootstrapCache = 16;
// constructor
ActionSummary() = default;
// constructor of action
explicit ActionSummary(int seqno_flag, int cache_flag = 0,
uint32_t minseqno = kSpecialOp, uint32_t maxseqno = kSpecialOp) {
seqcode_ = (minseqno << 5) | seqno_flag;
maxseqcode_ = (maxseqno << 5) | cache_flag;
}
// minimum number of all operations by default
// maximum number of all cache operations otherwise
inline uint32_t Seqno(SeqType t = SeqType::kSeq) const {
int code = t == SeqType::kSeq ? seqcode_ : maxseqcode_;
return code >> 5;
}
// whether the operation set contains a load_check
inline bool LoadCheck(SeqType t = SeqType::kSeq) const {
int code = t == SeqType::kSeq ? seqcode_ : maxseqcode_;
return (code & kLoadCheck) != 0;
}
// whether the operation set contains a load_cache
inline bool LoadCache(SeqType t = SeqType::kSeq) const {
int code = t == SeqType::kSeq ? seqcode_ : maxseqcode_;
return (code & kLoadBootstrapCache) != 0;
}
// whether the operation set contains a check point
inline bool CheckPoint(SeqType t = SeqType::kSeq) const {
int code = t == SeqType::kSeq ? seqcode_ : maxseqcode_;
return (code & kCheckPoint) != 0;
}
// whether the operation set contains a check ack
inline bool CheckAck(SeqType t = SeqType::kSeq) const {
int code = t == SeqType::kSeq ? seqcode_ : maxseqcode_;
return (code & kCheckAck) != 0;
}
// whether the operation set contains different sequence number
inline bool DiffSeq() const {
return (seqcode_ & kDiffSeq) != 0;
}
// returns the operation flag of the result
inline int Flag(SeqType t = SeqType::kSeq) const {
int code = t == SeqType::kSeq ? seqcode_ : maxseqcode_;
return code & 31;
}
// print flags in user friendly way
inline void PrintFlags(int rank, std::string prefix ) {
utils::HandleLogInfo("[%d] %s - |%lu|%d|%d|%d|%d| - |%lu|%d|\n", rank,
prefix.c_str(), Seqno(), CheckPoint(), CheckAck(),
LoadCache(), DiffSeq(), Seqno(SeqType::kCache),
LoadCache(SeqType::kCache));
}
// reducer for Allreduce, get the result ActionSummary from all nodes
inline static void Reducer(const void *src_, void *dst_,
int len, const MPI::Datatype &dtype) {
const ActionSummary *src = static_cast<const ActionSummary*>(src_);
ActionSummary *dst = reinterpret_cast<ActionSummary*>(dst_);
for (int i = 0; i < len; ++i) {
uint32_t min_seqno = Min(src[i].Seqno(), dst[i].Seqno());
uint32_t max_seqno = Max(src[i].Seqno(SeqType::kCache),
dst[i].Seqno(SeqType::kCache));
int action_flag = src[i].Flag() | dst[i].Flag();
// if any node is not requester set to 0 otherwise 1
int role_flag = src[i].Flag(SeqType::kCache) & dst[i].Flag(SeqType::kCache);
// if seqno is different in src and destination
int seq_diff_flag = src[i].Seqno() != dst[i].Seqno() ? kDiffSeq : 0;
// apply or to both seq diff flag as well as cache seq diff flag
dst[i] = ActionSummary(action_flag | seq_diff_flag,
role_flag, min_seqno, max_seqno);
}
}
private:
// internel sequence code min of rabit seqno
uint32_t seqcode_;
// internal sequence code max of cache seqno
uint32_t maxseqcode_;
};
/*! \brief data structure to remember result of Bcast and Allreduce calls*/
class ResultBuffer{
public:
// constructor
ResultBuffer() {
this->Clear();
}
// clear the existing record
inline void Clear() {
seqno_.clear(); size_.clear();
rptr_.clear(); rptr_.push_back(0);
data_.clear();
}
// allocate temporal space
inline void *AllocTemp(size_t type_nbytes, size_t count) {
size_t size = type_nbytes * count;
size_t nhop = (size + sizeof(uint64_t) - 1) / sizeof(uint64_t);
utils::Assert(nhop != 0, "cannot allocate 0 size memory");
// allocate addational nhop buffer size
data_.resize(rptr_.back() + nhop);
return BeginPtr(data_) + rptr_.back();
}
// push the result in temp to the
inline void PushTemp(int seqid, size_t type_nbytes, size_t count) {
size_t size = type_nbytes * count;
size_t nhop = (size + sizeof(uint64_t) - 1) / sizeof(uint64_t);
if (seqno_.size() != 0) {
utils::Assert(seqno_.back() < seqid, "PushTemp seqid inconsistent");
}
seqno_.push_back(seqid);
rptr_.push_back(rptr_.back() + nhop);
size_.push_back(size);
utils::Assert(data_.size() == rptr_.back(), "PushTemp inconsistent");
}
// return the stored result of seqid, if any
inline void* Query(int seqid, size_t *p_size) {
size_t idx = std::lower_bound(seqno_.begin(),
seqno_.end(), seqid) - seqno_.begin();
if (idx == seqno_.size() || seqno_[idx] != seqid) return nullptr;
*p_size = size_[idx];
return BeginPtr(data_) + rptr_[idx];
}
// drop last stored result
inline void DropLast() {
utils::Assert(seqno_.size() != 0, "there is nothing to be dropped");
seqno_.pop_back();
rptr_.pop_back();
size_.pop_back();
data_.resize(rptr_.back());
}
// the sequence number of last stored result
inline int LastSeqNo() const {
if (seqno_.size() == 0) return -1;
return seqno_.back();
}
private:
// sequence number of each
std::vector<int> seqno_;
// pointer to the positions
std::vector<size_t> rptr_;
// actual size of each buffer
std::vector<size_t> size_;
// content of the buffer
std::vector<uint64_t> data_;
};
/*!
* \brief internal consistency check function,
* use check to ensure user always call CheckPoint/LoadCheckPoint
* with or without local but not both, this function will set the approperiate settings
* in the first call of LoadCheckPoint/CheckPoint
*
* \param with_local whether the user calls CheckPoint with local model
*/
void LocalModelCheck(bool with_local);
/*!
* \brief internal implementation of checkpoint, support both lazy and normal way
*
* \param global_model pointer to the globally shared model/state
* when calling this function, the caller need to gauranttees that global_model
* is the same in all nodes
* \param local_model pointer to local model, that is specific to current node/rank
* this can be NULL when no local state is needed
* \param lazy_checkpt whether the action is lazy checkpoint
*
* \sa CheckPoint, LazyCheckPoint
*/
void CheckPointImpl(const Serializable *global_model,
const Serializable *local_model, bool lazy_checkpt);
/*!
* \brief reset the all the existing links by sending Out-of-Band message marker
* after this function finishes, all the messages received and sent
* before in all live links are discarded,
* This allows us to get a fresh start after error has happened
*
* TODO(tqchen): this function is not yet functioning was not used by engine,
* simple resetlink and reconnect strategy is used
*
* \return this function can return kSuccess or kSockError
* when kSockError is returned, it simply means there are bad sockets in the links,
* and some link recovery proceduer is needed
*/
ReturnType TryResetLinks();
/*!
* \brief if err_type indicates an error
* recover links according to the error type reported
* if there is no error, return true
* \param err_type the type of error happening in the system
* \return true if err_type is kSuccess, false otherwise
*/
bool CheckAndRecover(ReturnType err_type);
/*!
* \brief try to run recover execution for a request action described by flag and seqno,
* the function will keep blocking to run possible recovery operations before the specified action,
* until the requested result is received by a recovering procedure,
* or the function discovers that the requested action is not yet executed, and return false
*
* \param buf the buffer to store the result
* \param size the total size of the buffer
* \param flag flag information about the action \sa ActionSummary
* \param seqno sequence number of the action, if it is special action with flag set,
* seqno needs to be set to ActionSummary::kSpecialOp
*
* \return if this function can return true or false
* - true means buf already set to the
* result by recovering procedure, the action is complete, no further action is needed
* - false means this is the lastest action that has not yet been executed, need to execute the action
*/
bool RecoverExec(void *buf, size_t size, int flag,
int seqno = ActionSummary::kSpecialOp,
int cacheseqno = ActionSummary::kSpecialOp,
const char* caller = _CALLER);
/*!
* \brief try to load check point
*
* This is a collaborative function called by all nodes
* only the nodes with requester set to true really needs to load the check point
* other nodes acts as collaborative roles to complete this request
*
* \param requester whether current node is the requester
* \return this function can return kSuccess/kSockError/kGetExcept, see ReturnType for details
* \sa ReturnType
*/
ReturnType TryLoadCheckPoint(bool requester);
/*!
* \brief try to load cache
*
* This is a collaborative function called by all nodes
* only the nodes with requester set to true really needs to load the check point
* other nodes acts as collaborative roles to complete this request
* \param requester whether current node is the requester
* \return this function can return kSuccess/kSockError/kGetExcept, see ReturnType for details
* \sa ReturnType
*/
ReturnType TryRestoreCache(bool requester, const int min_seq = ActionSummary::kSpecialOp,
const int max_seq = ActionSummary::kSpecialOp);
/*!
* \brief try to get the result of operation specified by seqno
*
* This is a collaborative function called by all nodes
* only the nodes with requester set to true really needs to get the result
* other nodes acts as collaborative roles to complete this request
*
* \param buf the buffer to store the result, this parameter is only used when current node is requester
* \param size the total size of the buffer, this parameter is only used when current node is requester
* \param seqno sequence number of the operation, this is unique index of a operation in current iteration
* \param requester whether current node is the requester
* \return this function can return kSuccess/kSockError/kGetExcept, see ReturnType for details
* \sa ReturnType
*/
ReturnType TryGetResult(void *buf, size_t size, int seqno, bool requester);
/*!
* \brief try to decide the routing strategy for recovery
* \param role the current role of the node
* \param p_size used to store the size of the message, for node in state kHaveData,
* this size must be set correctly before calling the function
* for others, this surves as output parameter
* \param p_recvlink used to store the link current node should recv data from, if necessary
* this can be -1, which means current node have the data
* \param p_req_in used to store the resulting vector, indicating which link we should send the data to
*
* \return this function can return kSuccess/kSockError/kGetExcept, see ReturnType for details
* \sa ReturnType, TryRecoverData
*/
ReturnType TryDecideRouting(RecoverType role,
size_t *p_size,
int *p_recvlink,
std::vector<bool> *p_req_in);
/*!
* \brief try to finish the data recovery request,
* this function is used together with TryDecideRouting
* \param role the current role of the node
* \param sendrecvbuf_ the buffer to store the data to be sent/recived
* - if the role is kHaveData, this stores the data to be sent
* - if the role is kRequestData, this is the buffer to store the result
* - if the role is kPassData, this will not be used, and can be NULL
* \param size the size of the data, obtained from TryDecideRouting
* \param recv_link the link index to receive data, if necessary, obtained from TryDecideRouting
* \param req_in the request of each link to send data, obtained from TryDecideRouting
*
* \return this function can return kSuccess/kSockError/kGetExcept, see ReturnType for details
* \sa ReturnType, TryDecideRouting
*/
ReturnType TryRecoverData(RecoverType role,
void *sendrecvbuf_,
size_t size,
int recv_link,
const std::vector<bool> &req_in);
/*!
* \brief try to recover the local state, making each local state to be the result of itself
* plus replication of states in previous num_local_replica hops in the ring
*
* The input parameters must contain the valid local states available in current nodes,
* This function try ist best to "complete" the missing parts of local_rptr and local_chkpt
* If there is sufficient information in the ring, when the function returns, local_chkpt will
* contain num_local_replica + 1 checkpoints (including the chkpt of this node)
* If there is no sufficient information in the ring, this function the number of checkpoints
* will be less than the specified value
*
* \param p_local_rptr the pointer to the segment pointers in the states array
* \param p_local_chkpt the pointer to the storage of local check points
* \return this function can return kSuccess/kSockError/kGetExcept, see ReturnType for details
* \sa ReturnType
*/
ReturnType TryRecoverLocalState(std::vector<size_t> *p_local_rptr,
std::string *p_local_chkpt);
/*!
* \brief try to checkpoint local state, this function is called in normal executation phase
* of checkpoint that contains local state
o * the input state must exactly one saved state(local state of current node),
* after complete, this function will get local state from previous num_local_replica nodes and put them
* into local_chkpt and local_rptr
*
* It is also OK to call TryRecoverLocalState instead,
* TryRecoverLocalState makes less assumption about the input, and requires more communications
*
* \param p_local_rptr the pointer to the segment pointers in the states array
* \param p_local_chkpt the pointer to the storage of local check points
* \return this function can return kSuccess/kSockError/kGetExcept, see ReturnType for details
* \sa ReturnType, TryRecoverLocalState
*/
ReturnType TryCheckinLocalState(std::vector<size_t> *p_local_rptr,
std::string *p_local_chkpt);
/*!
* \brief perform a ring passing to receive data from prev link, and sent data to next link
* this allows data to stream over a ring structure
* sendrecvbuf[0:read_ptr] are already provided by current node
* current node will recv sendrecvbuf[read_ptr:read_end] from prev link
* current node will send sendrecvbuf[write_ptr:write_end] to next link
* write_ptr will wait till the data is readed before sending the data
* this function requires read_end >= write_end
*
* \param sendrecvbuf_ the place to hold the incoming and outgoing data
* \param read_ptr the initial read pointer
* \param read_end the ending position to read
* \param write_ptr the initial write pointer
* \param write_end the ending position to write
* \param read_link pointer to link to previous position in ring
* \param write_link pointer to link of next position in ring
*/
ReturnType RingPassing(void *senrecvbuf_,
size_t read_ptr,
size_t read_end,
size_t write_ptr,
size_t write_end,
LinkRecord *read_link,
LinkRecord *write_link);
/*!
* \brief run message passing algorithm on the allreduce tree
* the result is edge message stored in p_edge_in and p_edge_out
* \param node_value the value associated with current node
* \param p_edge_in used to store input message from each of the edge
* \param p_edge_out used to store output message from each of the edge
* \param func a function that defines the message passing rule
* Parameters of func:
* - node_value same as node_value in the main function
* - edge_in the array of input messages from each edge,
* this includes the output edge, which should be excluded
* - out_index array the index of output edge, the function should
* exclude the output edge when compute the message passing value
* Return of func:
* the function returns the output message based on the input message and node_value
*
* \tparam EdgeType type of edge message, must be simple struct
* \tparam NodeType type of node value
*/
template<typename NodeType, typename EdgeType>
inline ReturnType MsgPassing(const NodeType &node_value,
std::vector<EdgeType> *p_edge_in,
std::vector<EdgeType> *p_edge_out,
EdgeType(*func)
(const NodeType &node_value,
const std::vector<EdgeType> &edge_in,
size_t out_index));
//---- recovery data structure ----
// the round of result buffer, used to mode the result
int result_buffer_round_;
// result buffer of all reduce
ResultBuffer resbuf_;
// current cached allreduce/braodcast sequence number
int cur_cache_seq_;
// result buffer of cached all reduce
ResultBuffer cachebuf_;
// key of each cache entry
ResultBuffer lookupbuf_;
// last check point global model
std::string global_checkpoint_;
// lazy checkpoint of global model
const Serializable *global_lazycheck_;
// number of replica for local state/model
int num_local_replica_;
// number of default local replica
int default_local_replica_;
// flag to decide whether local model is used, -1: unknown, 0: no, 1:yes
int use_local_model_;
// number of replica for global state/model
int num_global_replica_;
// number of times recovery happens
int recover_counter_;
// --- recovery data structure for local checkpoint
// there is two version of the data structure,
// at one time one version is valid and another is used as temp memory
// pointer to memory position in the local model
// local model is stored in CSR format(like a sparse matrices)
// local_model[rptr[0]:rptr[1]] stores the model of current node
// local_model[rptr[k]:rptr[k+1]] stores the model of node in previous k hops
std::vector<size_t> local_rptr_[2];
// storage for local model replicas
std::string local_chkpt_[2];
// version of local checkpoint can be 1 or 0
int local_chkpt_version_;
// if checkpoint were loaded, used to distinguish results boostrap cache from seqno cache
bool checkpoint_loaded_;
// sidecar executing timeout task
std::future<bool> rabit_timeout_task_;
// flag to shutdown rabit_timeout_task before timeout
std::atomic<bool> shutdown_timeout_{false};
// error handler
void (* error_)(const char *fmt, ...) = utils::Error;
// assert handler
void (* assert_)(bool exp, const char *fmt, ...) = utils::Assert;
};
} // namespace engine
} // namespace rabit
// implementation of inline template function
#include "./allreduce_robust-inl.h"
#endif // RABIT_ALLREDUCE_ROBUST_H_

3
rabit/src/engine.cc
View File

@@ -12,14 +12,13 @@
#include <memory>
#include "rabit/internal/engine.h"
#include "allreduce_base.h"
#include "allreduce_robust.h"
namespace rabit {
namespace engine {
// singleton sync manager
#ifndef RABIT_USE_BASE
#ifndef RABIT_USE_MOCK
using Manager = AllreduceRobust;
using Manager = AllreduceBase;
#else
typedef AllreduceMock Manager;
#endif // RABIT_USE_MOCK