Distributed optimizations #5557

ShvetsKS · 2020-04-17T16:18:53Z

This PR consists changes to optimize distributed mode of training.

depthwise changes: Possibility to build 'smaller' nodes was introduced for distributed mode (before this only left nodes were build) for more balanced work. Than 'subtraction trick' is applied for 'larger' node.
lossguide changes: similar to above.
number of rabit::Allreduce calls is the same as before due to structured histograms adding.
Timers were added to track most time consuming parts for distributed also.

mortgage 3.5Gb

6 workers x 8 threads	loss	depth
Master	76.6	68.5
This PR	48.7	44

higgs 10m, 2.7Gb
running line: python .../dmlc-core/tracker/dmlc-submit --cluster=local --num-workers=4 --worker-cores=12 python distr_higgs10m.py

4 workers x 12 threads	depth
Master	877.83
This PR	223.66

thvasilo · 2020-04-21T20:27:43Z

Hello @ShvetsKS could you provide some details as to what the plan is here?

SmirnovEgorRu · 2020-04-21T21:36:51Z

@thvasilo,
The changes are close to this PR #4824. Currently we have different behavior for single-node and distributed modes in terms of what histograms we compute explicitly by BuildHist function and what nodes are built by SubtractionTrick.

Now for single-node it's more optimal, because we select smaller set of rows for BuildHist function and SubtractionTrick for larger, it brings performance gain. For distributed mode we just use BuildHist for left siblings always, in general it shows worse performance.

The PR introduces similar logic to distributed mode.

src/tree/updater_quantile_hist.cc

trivialfis · 2020-04-27T21:17:21Z

A genuine question, do you have difficulty to read the hist code after many optimizations (I see that you are really good at spotting optimization opportunity so I may be not as skilled)? I see that specializing for each case is good for performance, but could you try to arrange the source code into finer modules and create an easier to follow structure? Maybe some of these can help others to understand the code?

policy based http://loki-lib.sourceforge.net/index.php?n=Main.Policy-basedDesign
strategy https://en.wikipedia.org/wiki/Strategy_pattern

trivialfis · 2020-04-27T21:18:44Z

Also, if you are specializing over something, then there needs to be a new test case for your specialization.

src/tree/updater_quantile_hist.cc

SmirnovEgorRu · 2020-04-30T04:10:50Z

src/tree/updater_quantile_hist.cc

-  common::BlockedSpace2d space(nodes_for_explicit_hist_build_.size(), [&](size_t node) {
-    return nbins;
-  }, 1024);
+  if (!isDistributed) {


This "if-else" is tooo large.
Need to split this in different methods or classes or apply another refactoring approach.

src/tree/updater_quantile_hist.cc

SmirnovEgorRu · 2020-04-30T04:15:26Z

src/tree/updater_quantile_hist.cc

  BuildLocalHistograms(gmat, gmatb, p_tree, gpair_h);
  SyncHistograms(starting_index, sync_count, p_tree);
 }


-void QuantileHistMaker::Builder::AddHistRows(int *starting_index, int *sync_count) {
+void QuantileHistMaker::Builder::AddHistRows(int *starting_index, int *sync_count,


Unfortunately this function became massive and needed to be refactored.
I think we can add firstly left and secondary right nodes - in simpler manner.

For example, you don't need to create tmp vector and sort it. It's enough to just have a loop which handle all required things.

unfortunately it should be sorted, without sorting I get wrong results(different evaluation error)

src/tree/updater_quantile_hist.cc

src/tree/updater_quantile_hist.h

SmirnovEgorRu · 2020-04-30T04:28:01Z

src/tree/updater_quantile_hist.cc

+    common::ParallelFor2d(space, this->nthread_, [&](size_t node, common::Range1d r) {
+      const auto entry = nodes_for_explicit_hist_build_[node];
+      if (!((*p_tree)[entry.nid].IsLeftChild())) {
+        auto this_hist = hist_[entry.nid];

-      SubtractionHist(sibling_hist, parent_hist, this_hist, r.begin(), r.end());
-    }
-  });
+        if (!(*p_tree)[entry.nid].IsRoot() && entry.sibling_nid > -1) {
+          auto parent_hist = hist_[(*p_tree)[entry.nid].Parent()];
+          auto sibling_hist = hist_[entry.sibling_nid];
+          SubtractionHist(this_hist, parent_hist, sibling_hist, r.begin(), r.end());
+        }
+      }
+    });

-  if (isDistributed) {
-    this->histred_.Allreduce(hist_[starting_index].data(), hist_builder_.GetNumBins() * sync_count);
-    // use Subtraction Trick
-    for (auto const& node : nodes_for_subtraction_trick_) {
-      SubtractionTrick(hist_[node.nid], hist_[node.sibling_nid],
-                       hist_[(*p_tree)[node.nid].Parent()]);
-    }
+    common::BlockedSpace2d space2(nodes_for_subtraction_trick_.size(), [&](size_t node) {
+      return nbins;
+    }, 1024);
+
+    common::ParallelFor2d(space2, this->nthread_, [&](size_t node, common::Range1d r) {
+      const auto entry = nodes_for_subtraction_trick_[node];
+      if (!((*p_tree)[entry.nid].IsLeftChild())) {
+        auto this_hist = hist_[entry.nid];
+
+        if (!(*p_tree)[entry.nid].IsRoot() && entry.sibling_nid > -1) {
+          auto parent_hist = hist_[(*p_tree)[entry.nid].Parent()];
+          auto sibling_hist = hist_[entry.sibling_nid];
+          SubtractionHist(this_hist, parent_hist, sibling_hist, r.begin(), r.end());
+        }
+      }
+    });


So, you just do Subtraction Trick for left nodes from 2 different queues, other things are the same.
You can merge them into one simply or extract to a new method. Anyway, we should reduce "copy-past" code.

method ParallelSubtractionHist was added

trivialfis

Please don't merge until #5557 (comment) is resolved. Also can we bring some refactoring like #5557 (comment) before merging any further optimization ?

src/tree/updater_quantile_hist.cc

ShvetsKS · 2020-05-08T18:31:10Z

Please don't merge until #5557 (comment) is resolved. Also can we bring some refactoring like #5557 (comment) before merging any further optimization ?

@trivialfis Thanks for your suggestion.
Some refactoring was done, I hope it's more clear now. (Not canonical strategy pattern was implemented for synchronization and adding hist rows. I think the role of 'Context' class is covered by 'QuantileHistMaker::Builder' already. )

Unit tests for new functionality were also done.

ShvetsKS · 2020-05-13T15:30:21Z

@trivialfis , @SmirnovEgorRu seems all major comments were applied, could you take a look ?

thvasilo · 2020-05-15T17:38:55Z

Hello @ShvetsKS since these change aim at improving distributed code, have you tested against the master in a distributed (not local cluster) environment.

My concern with previous changes being that they introduced many synchronization points. That's not happening any more I think, but I'd like to see the differences when run say over dask or Spark (or MPI) with network communication happening.

To stress test I'd suggest using the URL dataset which should introduce large amounts of communication.

trivialfis · 2020-05-17T16:36:32Z

@thvasilo We have specialized optimization for super sparse dataset, the opt is not enabled on distributed environment. If @ShvetsKS is interested I can point out the details.

trivialfis · 2020-05-17T16:48:00Z

src/tree/updater_quantile_hist.h

+                                       HistSynchronizer(builder) {}
+
+  void SyncHistograms(int starting_index,
+                              int sync_count,


trivialfis · 2020-05-17T16:52:14Z

src/tree/updater_quantile_hist.h

@@ -105,6 +109,12 @@ class QuantileHistMaker: public TreeUpdater {
  }

 protected:
+  friend class HistSynchronizer;


Do you think using a visitor pattern is suitable here? https://en.wikipedia.org/wiki/Visitor_pattern

These classes are defined as friends as it requires access to protected fields. Usage of visitor pattern does't eliminate this problem (get methods should be added, or AddHistRows and SyncHistograms methods should accept all required fields).

trivialfis · 2020-05-17T16:55:52Z

src/tree/updater_prune.cc

@@ -14,7 +14,7 @@
 #include "xgboost/json.h"
 #include "./param.h"
 #include "../common/io.h"
-
+#include "../common/timer.h"


If you want you can put the pruning into pre-pruning (don't grow the tree if criteria is not met) in the future to get rid of this pruner.

I only wanted to track 'hidden' time.

trivialfis · 2020-05-17T16:59:58Z

src/tree/updater_quantile_hist.h

+  }
+
+ protected:
+  QuantileHistMaker::Builder* builder_;


If you use the visitor pattern than these interleaved pointers might not be needed.

This pointer was moved to SyncHistograms and AddHistRows methods.
Sorry, I don't see benefits from applying visitor pattern (#5557 (comment)).

ShvetsKS · 2020-05-17T18:44:41Z

@thvasilo We have specialized optimization for super sparse dataset, the opt is not enabled on distributed environment. If @ShvetsKS is interested I can point out the details.

As far as I understand changes in this PR also should introduce some impact for sparse datasets (we have BuildHist specialized for sparse data, and Partition kernel). @trivialfis Could you share the details please?

@thvasilo thanks for suggestions, measurements are in progress.

trivialfis · 2020-05-17T19:34:53Z

@ShvetsKS This line:

xgboost/src/common/hist_util.cc

Line 129 in 535479e

if (sparsity < kSparsityThreshold && (!rabit::IsDistributed())) {

If you run URL without the SparseCuts, the memory usage will blow up (by blow up I mean we don't know how much it needs, no one has the machine with enough memory to try). But it's not enabled for distributed environment as we need to sync the number of required entries, I wasn't familiar with distributed system back then ...

ShvetsKS · 2020-05-18T06:31:53Z

@thvasilo
please see measurements (over Spark)
mortgage data set was used.

606M Rows on 4 SKL8180 nodes	depthwise, s	lossguide, s
Master	473.64	749.62
This PR	303.51	399.67

SmirnovEgorRu · 2020-05-18T18:27:43Z

src/tree/updater_quantile_hist.cc

+    (*starting_index) = std::min(nid, (*starting_index));
  }
+  (*sync_count) = builder->nodes_for_explicit_hist_build_.size();


Minor comment: there is no need to initialize starting_index and sync_count by special value in cases of single-node mode. Since it's used only in distributed mode.

SmirnovEgorRu

My major comments are applied
Tests are added
New code was isolated in separate classes
Performance speedup on multi-node is checked

So, I see no reason not to merge this.

thvasilo · 2020-05-18T20:39:33Z

@ShvetsKS This line:

xgboost/src/common/hist_util.cc

Line 129 in 535479e

if (sparsity < kSparsityThreshold && (!rabit::IsDistributed())) {

If you run URL without the SparseCuts, the memory usage will blow up (by blow up I mean we don't know how much it needs, no one has the machine with enough memory to try). But it's not enabled for distributed environment as we need to sync the number of required entries, I wasn't familiar with distributed system back then ...

That will be interesting to investigate in a future PR.

Thanks @ShvetsKS for the measurements, great improvements!

trivialfis

Nice work!

@RAMitchell

* [dask] Accept other inputs for prediction. (dmlc#5428) * Returns a series when input is dataframe. * Merge assert client. * [R-package] changed FindLibR to take advantage of CMake cache (dmlc#5427) * Support pandas SparseArray. (dmlc#5431) * [R-package] fixed uses of class() (dmlc#5426) Thank you a lot. Good catch! * [dask] Fix missing value for scikit-learn interface. (dmlc#5435) * Ranking metric acceleration on the gpu (dmlc#5398) * Add link to GPU documentation (dmlc#5437) * Add Accelerated Failure Time loss for survival analysis task (dmlc#4763) * [WIP] Add lower and upper bounds on the label for survival analysis * Update test MetaInfo.SaveLoadBinary to account for extra two fields * Don't clear qids_ for version 2 of MetaInfo * Add SetInfo() and GetInfo() method for lower and upper bounds * changes to aft * Add parameter class for AFT; use enum's to represent distribution and event type * Add AFT metric * changes to neg grad to grad * changes to binomial loss * changes to overflow * changes to eps * changes to code refactoring * changes to code refactoring * changes to code refactoring * Re-factor survival analysis * Remove aft namespace * Move function bodies out of AFTNormal and AFTLogistic, to reduce clutter * Move function bodies out of AFTLoss, to reduce clutter * Use smart pointer to store AFTDistribution and AFTLoss * Rename AFTNoiseDistribution enum to AFTDistributionType for clarity The enum class was not a distribution itself but a distribution type * Add AFTDistribution::Create() method for convenience * changes to extreme distribution * changes to extreme distribution * changes to extreme * changes to extreme distribution * changes to left censored * deleted cout * changes to x,mu and sd and code refactoring * changes to print * changes to hessian formula in censored and uncensored * changes to variable names and pow * changes to Logistic Pdf * changes to parameter * Expose lower and upper bound labels to R package * Use example weights; normalize log likelihood metric * changes to CHECK * changes to logistic hessian to standard formula * changes to logistic formula * Comply with coding style guideline * Revert back Rabit submodule * Revert dmlc-core submodule * Comply with coding style guideline (clang-tidy) * Fix an error in AFTLoss::Gradient() * Add missing files to amalgamation * Address @RAMitchell's comment: minimize future change in MetaInfo interface * Fix lint * Fix compilation error on 32-bit target, when size_t == bst_uint * Allocate sufficient memory to hold extra label info * Use OpenMP to speed up * Fix compilation on Windows * Address reviewer's feedback * Add unit tests for probability distributions * Make Metric subclass of Configurable * Address reviewer's feedback: Configure() AFT metric * Add a dummy test for AFT metric configuration * Complete AFT configuration test; remove debugging print * Rename AFT parameters * Clarify test comment * Add a dummy test for AFT loss for uncensored case * Fix a bug in AFT loss for uncensored labels * Complete unit test for AFT loss metric * Simplify unit tests for AFT metric * Add unit test to verify aggregate output from AFT metric * Use EXPECT_* instead of ASSERT_*, so that we run all unit tests * Use aft_loss_param when serializing AFTObj This is to be consistent with AFT metric * Add unit tests for AFT Objective * Fix OpenMP bug; clarify semantics for shared variables used in OpenMP loops * Add comments * Remove AFT prefix from probability distribution; put probability distribution in separate source file * Add comments * Define kPI and kEulerMascheroni in probability_distribution.h * Add probability_distribution.cc to amalgamation * Remove unnecessary diff * Address reviewer's feedback: define variables where they're used * Eliminate all INFs and NANs from AFT loss and gradient * Add demo * Add tutorial * Fix lint * Use 'survival:aft' to be consistent with 'survival:cox' * Move sample data to demo/data * Add visual demo with 1D toy data * Add Python tests Co-authored-by: Philip Cho <chohyu01@cs.washington.edu> * Force compressed buffer to be 4 bytes aligned. (dmlc#5441) * Refactor tests with data generator. (dmlc#5439) * Resolve travis failure. (dmlc#5445) * Install dependencies by pip. * Device dmatrix (dmlc#5420) * Reducing memory consumption for 'hist' method on CPU (dmlc#5334) * [R-package] fixed inconsistency in R -e calls in FindLibR.cmake (dmlc#5438) * Thread safe, inplace prediction. (dmlc#5389) Normal prediction with DMatrix is now thread safe with locks. Added inplace prediction is lock free thread safe. When data is on device (cupy, cudf), the returned data is also on device. * Implementation for numpy, csr, cudf and cupy. * Implementation for dask. * Remove sync in simple dmatrix. * Add support for dlpack, expose python docs for DeviceQuantileDMatrix (dmlc#5465) * Reduce span check overhead. (dmlc#5464) * Update dmlc-core. (dmlc#5466) * Copy dmlc travis script to XGBoost. * Prevent copying SimpleDMatrix. (dmlc#5453) * Set default dtor for SimpleDMatrix to initialize default copy ctor, which is deleted due to unique ptr. * Remove commented code. * Remove warning for calling host function (std::max). * Remove warning for initialization order. * Remove warning for unused variables. * Remove silent parameter. (dmlc#5476) * Enable parameter validation for skl. (dmlc#5477) * Split up test helpers header. (dmlc#5455) * Implement host span. (dmlc#5459) * Accept other gradient types for split entry. (dmlc#5467) * Implement robust regularization in 'survival:aft' objective (dmlc#5473) * Robust regularization of AFT gradient and hessian * Fix AFT doc; expose it to tutorial TOC * Apply robust regularization to uncensored case too * Revise unit test slightly * Fix lint * Update test_survival.py * Use GradientPairPrecise * Remove unused variables * Fix dump model. (dmlc#5485) * Small updates to GPU documentation (dmlc#5483) * Add R code to AFT tutorial [skip ci] (dmlc#5486) * Upgrade clang-tidy on CI. (dmlc#5469) * Correct all clang-tidy errors. * Upgrade clang-tidy to 10 on CI. Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * corrected spelling of 'list' (dmlc#5482) * Edits on tutorial for XGBoost job on Kubernetes (dmlc#5487) * add reference to gpu external memory (dmlc#5490) * Fix out-of-bound array access in WQSummary::SetPrune() (dmlc#5493) * [jvm-packages]add feature size for LabelPoint and DataBatch (dmlc#5303) * fix type error * Validate number of features. * resolve comments * add feature size for LabelPoint and DataBatch * pass the feature size to native * move feature size validating tests into a separate suite * resolve comments Co-authored-by: fis <jm.yuan@outlook.com> * Use ellpack for prediction only when sparsepage doesn't exist. (dmlc#5504) * Fix checking booster. (dmlc#5505) * Use `get_params()` instead of `getattr` intrinsic. * Requires setting leaf stat when expanding tree. (dmlc#5501) * Fix GPU Hist feature importance. * Remove distcol updater. (dmlc#5507) Closes dmlc#5498. * Unify max nodes. (dmlc#5497) * Fix github merge. (dmlc#5509) * Update doc for parameter validation. (dmlc#5508) * Update doc for parameter validation. * Fix github rebase. * Serialise booster after training to reset state (dmlc#5484) * Serialise booster after training to reset state * Prevent process_type being set on load * Check for correct updater sequence * Remove makefiles. (dmlc#5513) * [R] R raw serialization. (dmlc#5123) * Add bindings for serialization. * Change `xgb.save.raw' into full serialization instead of simple model. * Add `xgb.load.raw' for unserialization. * Run devtools. * [CI] Use devtoolset-6 because devtoolset-4 is EOL and no longer available (dmlc#5506) * Use devtoolset-6. * [CI] Use devtoolset-6 because devtoolset-4 is EOL and no longer available * CUDA 9.0 doesn't work with devtoolset-6; use devtoolset-4 for GPU build only Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * fix typo "customized" (dmlc#5515) * Ensure that configured dmlc/build_config.h is picked up by Rabit and XGBoost (dmlc#5514) * Ensure that configured header (build_config.h) from dmlc-core is picked up by Rabit and XGBoost * Check which Rabit target is being used * Use CMake 3.13 in all Jenkins tests * Upgrade CMake in Travis CI * Install CMake using Kitware installer * Remove existing CMake (3.12.4) * Update Python doc. [skip ci] (dmlc#5517) * Update doc for copying booster. [skip ci] The issue is resolved in dmlc#5312 . * Add version for new APIs. [skip ci] * Add Neptune and Optuna to list of examples (dmlc#5528) * [jvm-packages] [CI] Create a Maven repository to host SNAPSHOT JARs (dmlc#5533) * Write binary header. (dmlc#5532) * Purge device_helpers.cuh (dmlc#5534) * Simplifications with caching_device_vector * Purge device helpers * [dask] dask cudf inplace prediction. (dmlc#5512) * Add inplace prediction for dask-cudf. * Remove Dockerfile.release, since it's not used anywhere * Use Conda exclusively in CUDF and GPU containers * Improve cupy memory copying. * Add skip marks to tests. * Add mgpu-cudf category on the CI to run all distributed tests. Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * [CI] Use Ubuntu 18.04 LTS in JVM CI, because 19.04 is EOL (dmlc#5537) * [jvm-packages] [CI] Publish XGBoost4J JARs with Scala 2.11 and 2.12 (dmlc#5539) * Fix CLI model IO. (dmlc#5535) * Add test for comparing Python and CLI training result. * Fix uninitialized value bug in xgboost callback (dmlc#5463) Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Use thrust functions instead of custom functions (dmlc#5544) * Optimizations for RNG in InitData kernel (dmlc#5522) * optimizations for subsampling in InitData * optimizations for subsampling in InitData Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com> * Add missing aft parameters. [skip ci] (dmlc#5553) * Don't use uint for threads. (dmlc#5542) * Fix skl nan tag. (dmlc#5538) * Assert matching length of evaluation inputs. (dmlc#5540) * Fix r interaction constraints (dmlc#5543) * Unify the parsing code. * Cleanup. * Fix slice and get info. (dmlc#5552) * gpu_hist performance fixes (dmlc#5558) * Remove unnecessary cuda API calls * Fix histogram memory growth * Use non-synchronising scan (dmlc#5560) * Fix non-openmp build. (dmlc#5566) * Add test to Jenkins. * Fix threading utils tests. * Require thread library. * Don't set seed on CLI interface. (dmlc#5563) * [jvm-packages] XGBoost Spark should deal with NaN when parsing evaluation output (dmlc#5546) * Group aware GPU sketching. (dmlc#5551) * Group aware GPU weighted sketching. * Distribute group weights to each data point. * Relax the test. * Validate input meta info. * Fix metainfo copy ctor. * Fix configuration I load model. (dmlc#5562) * [Breaking] Set output margin to True for custom objective. (dmlc#5564) * Set output margin to True for custom objective in Python and R. * Add a demo for writing multi-class custom objective function. * Run tests on selected demos. * For histograms, opting into maximum shared memory available per block. (dmlc#5491) * Use cudaDeviceGetAttribute instead of cudaGetDeviceProperties (dmlc#5570) * Restore attributes in complete. (dmlc#5573) * Enable parameter validation for R. (dmlc#5569) * Enable parameter validation for R. * Add test. * Update document. (dmlc#5572) * Port R compatibility patches from 1.0.0 release branch (dmlc#5577) * Don't use memset to set struct when compiling for R * Support 32-bit Solaris target for R package * [CI] Use Vault repository to re-gain access to devtoolset-4 (dmlc#5589) * [CI] Use Vault repository to re-gain access to devtoolset-4 * Use manylinux2010 tag * Update Dockerfile.jvm * Fix rename_whl.py * Upgrade Pip, to handle manylinux2010 tag * Update insert_vcomp140.py * Update test_python.sh * Avoid rabit calls in learner configuration (dmlc#5581) * Hide C++ symbols in libxgboost.so when building Python wheel (dmlc#5590) * Hide C++ symbols in libxgboost.so when building Python wheel * Update Jenkinsfile * Add test * Upgrade rabit * Add setup.py option. Co-authored-by: fis <jm.yuan@outlook.com> * Set device in device dmatrix. (dmlc#5596) * Fix compilation on Mac OSX High Sierra (10.13) (dmlc#5597) * Fix compilation on Mac OSX High Sierra * [CI] Build Mac OSX binary wheel using Travis CI * [CI] Grant public read access to Mac OSX wheels (dmlc#5602) * [R] Address warnings to comply with CRAN submission policy (dmlc#5600) * [R] Address warnings to comply with CRAN submission policy * Include <xgboost/logging.h> * Instruct Mac users to install libomp (dmlc#5606) * Clarify meaning of `training` parameter in XGBoosterPredict() (dmlc#5604) Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> Co-authored-by: Jiaming Yuan <jm.yuan@outlook.com> * Better message when no GPU is found. (dmlc#5594) * Refactor the CLI. (dmlc#5574) * Enable parameter validation. * Enable JSON. * Catch `dmlc::Error`. * Show help message. * Move dask tutorial closer other distributed tutorials (dmlc#5613) * Refactor gpu_hist split evaluation (dmlc#5610) * Refactor * Rewrite evaluate splits * Add more tests * Fix build on big endian CPUs (dmlc#5617) * Fix build on big endian CPUs * Clang-tidy * Remove dead code. (dmlc#5635) * Move device dmatrix construction code into ellpack. (dmlc#5623) * Enhance nvtx support. (dmlc#5636) * Support 64bit seed. (dmlc#5643) * Resolve vector<bool>::iterator crash (dmlc#5642) * Reduce device synchronisation (dmlc#5631) * Reduce device synchronisation * Initialise pinned memory * Upgrade to CUDA 10.0 (dmlc#5649) (dmlc#5652) Co-authored-by: fis <jm.yuan@outlook.com> Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * skip missing lookup if nothing is missing in CPU hist partition kernel. (dmlc#5644) * [xgboost] skip missing lookup if nothing is missing * Update Python demos with tests. (dmlc#5651) * Remove GPU memory usage demo. * Add tests for demos. * Remove `silent`. * Remove shebang as it's not portable. * Add JSON schema to model dump. (dmlc#5660) * Pseudo-huber loss metric added (dmlc#5647) - Add pseudo huber loss objective. - Add pseudo huber loss metric. Co-authored-by: Reetz <s02reetz@iavgroup.local> * [JVM Packages] Catch dmlc error by ref. (dmlc#5678) * Remove silent from R demos. (dmlc#5675) * Remove silent from R demos. * Vignettes. * add pointers to the gpu external memory paper (dmlc#5684) * Distributed optimizations for 'hist' method with CPUs (dmlc#5557) Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com> * Document more objective parameters in R package (dmlc#5682) * C++14 for xgboost (dmlc#5664) * Implement Python data handler. (dmlc#5689) * Define data handlers for DMatrix. * Throw ValueError in scikit learn interface. * [R-package] Reduce duplication in configure.ac (dmlc#5693) * updated configure * Remove redundant sketching. (dmlc#5700) * [R] Fix duplicated libomp.dylib error on Mac OSX (dmlc#5701) * Fix IsDense. (dmlc#5702) * Let XGBoostError inherit ValueError. (dmlc#5696) * Define _CRT_SECURE_NO_WARNINGS to remove unneeded warnings in MSVC (dmlc#5434) * Changed build.rst (binary wheels are supported for macOS also) (dmlc#5711) * [CI] Remove CUDA 9.0 from Windows CI. (dmlc#5674) * Remove CUDA 9.0 on Windows CI. * Require cuda10 tag, to differentiate Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Require CUDA 10.0+ in CMake build (dmlc#5718) * Require Python 3.6+; drop Python 3.5 from CI (dmlc#5715) * [dask] Return GPU Series when input is from cuDF. (dmlc#5710) * Refactor predict function. * [Doc] Fix typos in AFT tutorial (dmlc#5716) * gpu_hist performance tweaks (dmlc#5707) * Remove device vectors * Remove allreduce synchronize * Remove double buffer * Allow pass fmap to importance plot (dmlc#5719) Co-authored-by: Peter Jung <peter.jung@heureka.cz> Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * Fix release degradation (dmlc#5720) * fix release degradation, related to 5666 * less resizes Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com> * Fix loading old model. (dmlc#5724) * Add test. * Bump version to 1.2.0 snapshot in master (dmlc#5733) * Add swift package reference (dmlc#5728) Co-authored-by: Peter Jung <peter.jung@heureka.cz> * Don't use mask in array interface. (dmlc#5730) * Bump version in header. (dmlc#5742) * [CI] Remove CUDA 9.0 from CI (dmlc#5745) * Add pkgconfig to cmake (dmlc#5744) * Add pkgconfig to cmake * Move xgboost.pc.in to cmake/ Co-authored-by: Peter Jung <peter.jung@heureka.cz> Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * Expose device sketching in header. (dmlc#5747) * Add Python binding for rabit ops. (dmlc#5743) * Add float32 histogram (dmlc#5624) * new single_precision_histogram param was added. Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com> Co-authored-by: fis <jm.yuan@outlook.com> * Reorder includes. (dmlc#5749) * Reorder includes. * R. * Remove `max.depth` in R gblinear example. (dmlc#5753) * Speed up python test (dmlc#5752) * Speed up tests * Prevent DeviceQuantileDMatrix initialisation with numpy * Use joblib.memory * Use RandomState * Add helper for generating batches of data. (dmlc#5756) * Add helper for generating batches of data. * VC keyword clash. * Another clash. * Remove column major specialization. (dmlc#5755) Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * Document addition of new committer @SmirnovEgorRu (dmlc#5762) * Add release note for 1.1.0 in NEWS.md (dmlc#5763) * Add release note for 1.1.0 in NEWS.md * Address reviewer's feedback * Revert "Reorder includes. (dmlc#5749)" (dmlc#5771) This reverts commit d3a0efb. * [python-package] remove unused imports (dmlc#5776) * Added conda environment file for building docs (dmlc#5773) * [R] replace uses of T and F with TRUE and FALSE (dmlc#5778) * [R-package] replace uses of T and F with TRUE and FALSE * enable linting * Remove skip Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Implement weighted sketching for adapter. (dmlc#5760) * Bounded memory tests. * Fixed memory estimation. * Avoid including `c_api.h` in header files. (dmlc#5782) * Implement `Empty` method for host device vector. (dmlc#5781) * Fix accessing nullptr. * Bump com.esotericsoftware to 4.0.2 (dmlc#5690) Co-authored-by: Antti Saukko <antti.saukko@verizonmedia.com> * [DOC] Mention dask blog post in doc. [skip ci] (dmlc#5789) * [R] Remove dependency on gendef for Visual Studio builds (fixes dmlc#5608) (dmlc#5764) * [R-package] Remove dependency on gendef for Visual Studio builds (fixes dmlc#5608) * clarify docs * removed debugging print statement * Make R CMake install more robust * Fix doc format; add ToC * Update build.rst * Fix AppVeyor Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * Add new skl model attribute for number of features (dmlc#5780) * Fix exception causes all over the codebase (dmlc#5787) * Use hypothesis (dmlc#5759) * Use hypothesis * Allow int64 array interface for groups * Add packages to Windows CI * Add to travis * Make sure device index is set correctly * Fix dask-cudf test * appveyor * Accept string for ArrayInterface constructor. * Revert "Accept string for ArrayInterface constructor." This reverts commit e8ecafb. * Implement fast number serialization routines. (dmlc#5772) * Implement ryu algorithm. * Implement integer printing. * Full coverage roundtrip test. * Add cupy to Windows CI (dmlc#5797) * Add cupy to Windows CI * Update Jenkinsfile-win64 Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Update Jenkinsfile-win64 Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Update tests/python-gpu/test_gpu_prediction.py Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Add an option to run brute-force test for JSON round-trip (dmlc#5804) * Add an option to run brute-force test for JSON round-trip * Apply reviewer's feedback * Remove unneeded objects * Parallel run. * Max. * Use signed 64-bit loop var, to support MSVC * Add exhaustive test to CI * Run JSON test in Win build worker * Revert "Run JSON test in Win build worker" This reverts commit c97b2c7. * Revert "Add exhaustive test to CI" This reverts commit c149c2c. Co-authored-by: fis <jm.yuan@outlook.com> * [CI] Fix cuDF install; merge 'gpu' and 'cudf' test suite (dmlc#5814) * Implement extend method for meta info. (dmlc#5800) * Implement extend for host device vector. * Update rabit. (dmlc#5680) * Update document for model dump. (dmlc#5818) * Clarify the relationship between dump and save. * Mention the schema. * [Doc] Fix rendering of Markdown docs, e.g. R doc (dmlc#5821) * Remove unweighted GK quantile. (dmlc#5816) * Rename Ant Financial to Ant Group (dmlc#5827) * Accept string for ArrayInterface constructor. (dmlc#5799) * Implement a DMatrix Proxy. (dmlc#5803) * Relax test for shotgun. (dmlc#5835) * Relax linear test. (dmlc#5849) * Increased error in coordinate is mostly due to floating point error. * Shotgun uses Hogwild!, which is non-deterministic and can have even greater floating point error. * Implement iterative DMatrix. (dmlc#5837) * Ensure that LoadSequentialFile() actually read the whole file (dmlc#5831) * Add c-api-demo to .gitignore (dmlc#5855) * Use dmlc stream when URI protocol is not local file. (dmlc#5857) * Move feature names and types of DMatrix from Python to C++. (dmlc#5858) * Add thread local return entry for DMatrix. * Save feature name and feature type in binary file. Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Split Features into Groups to Compute Histograms in Shared Memory (dmlc#5795) * Implement GK sketching on GPU. (dmlc#5846) * Implement GK sketching on GPU. * Strong tests on quantile building. * Handle sparse dataset by binary searching the column index. * Hypothesis test on dask. * Accept iterator in device dmatrix. (dmlc#5783) * Remove Device DMatrix. * Remove print. (dmlc#5867) * fix device sketch with weights in external memory mode (dmlc#5870) * [Doc] Document that CUDA 10.0 is required [skip ci] (dmlc#5872) * [CI] Simplify CMake build with modern CMake techniques (dmlc#5871) * [CI] Simplify CMake build * Make sure that plugins can be built * [CI] Install lz4 on Mac * Add new parameter singlePrecisionHistogram to xgboost4j-spark (dmlc#5811) Expose the existing 'singlePrecisionHistogram' param to the Spark layer. * Upgrade Rabit (dmlc#5876) * [jvm-packages] update spark dependency to 3.0.0 (dmlc#5836) * Cleanup on device sketch. (dmlc#5874) * Remove old functions. * Merge weighted and un-weighted into a common interface. * [CI] Enforce daily budget in Jenkins CI (dmlc#5884) * [CI] Throttle Jenkins CI * Don't use Jenkins master instance * Add XGBoosterGetNumFeature (dmlc#5856) - add GetNumFeature to Learner - add XGBoosterGetNumFeature to C API - update c-api-demo accordingly * Fix NDK Build. (dmlc#5886) * Explicit cast for slice. * [CI] Reduce load on Windows CI pipeline (dmlc#5892) * Fix R package build with CMake 3.13 (dmlc#5895) * Fix R package build with CMake 3.13 * Require OpenMP for xgboost-r target * Simplify the data backends. (dmlc#5893) * [CI] update spark version to 3.0.0 (dmlc#5890) * [CI] update spark version to 3.0.0 * Update Dockerfile.jvm_cross Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Fix sketch size calculation. (dmlc#5898) * Dask device dmatrix (dmlc#5901) * Fix softprob with empty dmatrix. * GPU implementation of AFT survival objective and metric (dmlc#5714) * Add interval accuracy * De-virtualize AFT functions * Lint * Refactor AFT metric using GPU-CPU reducer * Fix R build * Fix build on Windows * Fix copyright header * Clang-tidy * Fix crashing demo * Fix typos in comment; explain GPU ID * Remove unnecessary #include * Add C++ test for interval accuracy * Fix a bug in accuracy metric: use log pred * Refactor AFT objective using GPU-CPU Transform * Lint * Fix lint * Use Ninja to speed up build * Use time, not /usr/bin/time * Add cpu_build worker class, with concurrency = 1 * Use concurrency = 1 only for CUDA build * concurrency = 1 for clang-tidy * Address reviewer's feedback * Update link to AFT paper * Fix Windows 2016 build. (dmlc#5902) * Further improvements and savings in Jenkins pipeline (dmlc#5904) * Publish artifacts only on the master and release branches * Build CUDA only for Compute Capability 7.5 when building PRs * Run all Windows jobs in a single worker image * Build nightly XGBoost4J SNAPSHOT JARs with Scala 2.12 only * Show skipped Python tests on Windows * Make Graphviz optional for Python tests * Add back C++ tests * Unstash xgboost_cpp_tests * Fix label to CUDA 10.1 * Install cuPy for CUDA 10.1 * Install jsonschema * Address reviewer's feedback * Support building XGBoost with CUDA 11 (dmlc#5808) * Change serialization test. * Add CUDA 11 tests on Linux CI. Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * Add Github Action for R. (dmlc#5911) * Fix lintr errors. * Fix typo in CI. [skip ci] (dmlc#5919) * [Doc] Document new objectives and metrics available on GPUs (dmlc#5909) * Fix mingw build with R. (dmlc#5918) * Add option to enable all compiler warnings in GCC/Clang (dmlc#5897) * Add option to enable all compiler warnings in GCC/Clang * Fix -Wall for CUDA sources * Make -Wall private req for xgboost-r * Setup github action. (dmlc#5917) * Remove R and JVM from appveyor. (dmlc#5922) * Fix r early stop with custom objective. (dmlc#5923) * Specify `ntreelimit`. * Add explicit template specialization for portability (dmlc#5921) * Add explicit template specializations * Adding Specialization for FileAdapterBatch * Cache dependencies on Github Action. (dmlc#5928) * Use `cudaOccupancyMaxPotentialBlockSize` to calculate the block size. (dmlc#5926) * [BLOCKING] Handle empty rows in data iterators correctly (dmlc#5929) * [jvm-packages] Handle empty rows in data iterators correctly * Fix clang-tidy error * last empty row * Add comments [skip ci] Co-authored-by: Nan Zhu <nanzhu@uber.com> * [CI] Make Python model compatibility test runnable locally (dmlc#5941) * [BLOCKING] Remove to_string. (dmlc#5934) * [R] Add a compatibility layer to load Booster object from an old RDS file (dmlc#5940) * [R] Add a compatibility layer to load Booster from an old RDS * Modify QuantileHistMaker::LoadConfig() to be backward compatible with 1.1.x * Add a big warning about compatibility in QuantileHistMaker::LoadConfig() * Add testing suite * Discourage use of saveRDS() in CRAN doc * [R] Enable weighted learning to rank (dmlc#5945) * [R] enable weighted learning to rank * Add R unit test for ranking * Fix lint * [BLOCKING] [jvm-packages] add gpu_hist and enable gpu scheduling (dmlc#5171) * [jvm-packages] add gpu_hist tree method * change updater hist to grow_quantile_histmaker * add gpu scheduling * pass correct parameters to xgboost library * remove debug info * add use.cuda for pom * add CI for gpu_hist for jvm * add gpu unit tests * use gpu node to build jvm * use nvidia-docker * Add CLI interface to create_jni.py using argparse Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * [CI] Improve R linter script (dmlc#5944) * [CI] Move lint to a separate script * [CI] Improved lintr launcher * Add lintr as a separate action * Add custom parsing logic to print out logs * Fix lintr issues in demos * Run R demos * Fix CRAN checks * Install XGBoost into R env before running lintr * Install devtools (needed to run demos) * Fix prediction heuristic (dmlc#5955) * Relax check for prediction. * Relax test in spark test. * Add tests in C++. * [Breaking] Fix custom metric for multi output. (dmlc#5954) * Set output margin to true for custom metric. This fixes only R and Python. * Disable feature validation on sklearn predict prob. (dmlc#5953) * Fix issue when scikit learn interface receives transformed inputs. * [CI] Fix broken Docker container 'cpu' (dmlc#5956) * Fix evaluate root split. (dmlc#5948) * [Dask] Asyncio support. (dmlc#5862) * Thread-safe prediction by making the prediction cache thread-local. (dmlc#5853) Co-authored-by: Jiaming Yuan <jm.yuan@outlook.com> * Force colored output for ninja build. (dmlc#5959) * Update XGBoost + Dask overview documentation (dmlc#5961) * Add imports to code snippet * Better writing. * Add CMake flag to log C API invocations, to aid debugging (dmlc#5925) * Add CMake flag to log C API invocations, to aid debugging * Remove unnecessary parentheses * [CI] Assign larger /dev/shm to NCCL (dmlc#5966) * [CI] Assign larger /dev/shm to NCCL * Use 10.2 artifact to run multi-GPU Python tests * Add CUDA 10.0 -> 11.0 cross-version test; remove CUDA 10.0 target * Add missing Pytest marks to AsyncIO unit test (dmlc#5968) * [R] Provide better guidance for persisting XGBoost model (dmlc#5964) * [R] Provide better guidance for persisting XGBoost model * Update saving_model.rst * Add a paragraph about xgb.serialize() * [jvm-packages] Fix wrong method name `setAllowZeroForMissingValue`. (dmlc#5740) * Allow non-zero for missing value when training. * Fix wrong method names. * Add a unit test * Move the getter/setter unit test to MissingValueHandlingSuite Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * Export DaskDeviceQuantileDMatrix in doc. [skip ci] (dmlc#5975) * Fix sklearn doc. (dmlc#5980) * Update Python custom objective demo. (dmlc#5981) * Update JSON schema. (dmlc#5982) * Update JSON schema for pseudo huber. * Update JSON model schema. * Fix missing data warning. (dmlc#5969) * Fix data warning. * Add numpy/scipy test. * Enforce tree order in JSON. (dmlc#5974) * Make JSON model IO more future proof by using tree id in model loading. * Fix dask predict shape infer. (dmlc#5989) * [R] fix uses of 1:length(x) and other small things (dmlc#5992) * Fix typo in tracker logging (dmlc#5994) * Introducing DPC++-based plugin (predictor, objective function) supporting oneAPI programming model (dmlc#5825) * Added plugin with DPC++-based predictor and objective function * Update CMakeLists.txt * Update regression_obj_oneapi.cc * Added README.md for OneAPI plugin * Added OneAPI predictor support to gbtree * Update README.md * Merged kernels in gradient computation. Enabled multiple loss functions with DPC++ backend * Aligned plugin CMake files with latest master changes. Fixed whitespace typos * Removed debug output * [CI] Make oneapi_plugin a CMake target * Added tests for OneAPI plugin for predictor and obj. functions * Temporarily switched to default selector for device dispacthing in OneAPI plugin to enable execution in environments without gpus * Updated readme file. * Fixed USM usage in predictor * Removed workaround with explicit templated names for DPC++ kernels * Fixed warnings in plugin tests * Fix CMake build of gtest Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu> * Remove skmaker. (dmlc#5971) * Rabit update. (dmlc#5978) * Remove parameter on JVM Packages. * Move warning about empty dataset. (dmlc#5998) * [Breaking] Fix .predict() method and add .predict_proba() in xgboost.dask.DaskXGBClassifier (dmlc#5986) * Unify CPU hist sketching (dmlc#5880) * Fix nightly build doc. [skip ci] (dmlc#6004) * Fix nightly build doc. [skip ci] * Fix title too short. [skip ci] * RMM integration plugin (dmlc#5873) * [CI] Add RMM as an optional dependency * Replace caching allocator with pool allocator from RMM * Revert "Replace caching allocator with pool allocator from RMM" This reverts commit e15845d. * Use rmm::mr::get_default_resource() * Try setting default resource (doesn't work yet) * Allocate pool_mr in the heap * Prevent leaking pool_mr handle * Separate EXPECT_DEATH() in separate test suite suffixed DeathTest * Turn off death tests for RMM * Address reviewer's feedback * Prevent leaking of cuda_mr * Fix Jenkinsfile syntax * Remove unnecessary function in Jenkinsfile * [CI] Install NCCL into RMM container * Run Python tests * Try building with RMM, CUDA 10.0 * Do not use RMM for CUDA 10.0 target * Actually test for test_rmm flag * Fix TestPythonGPU * Use CNMeM allocator, since pool allocator doesn't yet support multiGPU * Use 10.0 container to build RMM-enabled XGBoost * Revert "Use 10.0 container to build RMM-enabled XGBoost" This reverts commit 789021f. * Fix Jenkinsfile * [CI] Assign larger /dev/shm to NCCL * Use 10.2 artifact to run multi-GPU Python tests * Add CUDA 10.0 -> 11.0 cross-version test; remove CUDA 10.0 target * Rename Conda env rmm_test -> gpu_test * Use env var to opt into CNMeM pool for C++ tests * Use identical CUDA version for RMM builds and tests * Use Pytest fixtures to enable RMM pool in Python tests * Move RMM to plugin/CMakeLists.txt; use PLUGIN_RMM * Use per-device MR; use command arg in gtest * Set CMake prefix path to use Conda env * Use 0.15 nightly version of RMM * Remove unnecessary header * Fix a unit test when cudf is missing * Add RMM demos * Remove print() * Use HostDeviceVector in GPU predictor * Simplify pytest setup; use LocalCUDACluster fixture * Address reviewers' commments Co-authored-by: Hyunsu Cho <chohyu01@cs.wasshington.edu> Co-authored-by: Jiaming Yuan <jm.yuan@outlook.com> Co-authored-by: James Lamb <jaylamb20@gmail.com> Co-authored-by: sriramch <33358417+sriramch@users.noreply.github.com> Co-authored-by: Rory Mitchell <r.a.mitchell.nz@gmail.com> Co-authored-by: Avinash Barnwal <avinashbarnwal123@gmail.com> Co-authored-by: Philip Cho <chohyu01@cs.washington.edu> Co-authored-by: ShvetsKS <33296480+ShvetsKS@users.noreply.github.com> Co-authored-by: Paul Kaefer <2408155+paulkaefer@users.noreply.github.com> Co-authored-by: Yuan Tang <terrytangyuan@gmail.com> Co-authored-by: Rong Ou <rong.ou@gmail.com> Co-authored-by: Zhang Zhang <zhang.zhang@intel.com> Co-authored-by: Bobby Wang <wbo4958@gmail.com> Co-authored-by: Liang-Chi Hsieh <viirya@gmail.com> Co-authored-by: Nicolas Scozzaro <nscozzaro@gmail.com> Co-authored-by: Kamil A. Kaczmarek <kamil.kaczmarek@neptune.ai> Co-authored-by: Melissa Kohl <mjkohl32@gmail.com> Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com> Co-authored-by: Liang-Chi Hsieh <liangchi@uber.com> Co-authored-by: Andy Adinets <aadinets@nvidia.com> Co-authored-by: Jason E. Aten, Ph.D <j.e.aten@gmail.com> Co-authored-by: Oleksandr Kuvshynov <661042+okuvshynov@users.noreply.github.com> Co-authored-by: LionOrCatThatIsTheQuestion <44895499+LionOrCatThatIsTheQuestion@users.noreply.github.com> Co-authored-by: Reetz <s02reetz@iavgroup.local> Co-authored-by: Lorenz Walthert <lorenz.walthert@icloud.com> Co-authored-by: Dmitry Mottl <dmitry.mottl@gmail.com> Co-authored-by: Peter Jung <peter@jung.ninja> Co-authored-by: Peter Jung <peter.jung@heureka.cz> Co-authored-by: Elliot Hershberg <eahershberg@gmail.com> Co-authored-by: anttisaukko <antti.saukko@gmail.com> Co-authored-by: Antti Saukko <antti.saukko@verizonmedia.com> Co-authored-by: Alex <wozn0001@e.ntu.edu.sg> Co-authored-by: Ram Rachum <ram@rachum.com> Co-authored-by: Alexander Gugel <alexander.gugel@gmail.com> Co-authored-by: Nan Zhu <nanzhu@uber.com> Co-authored-by: boxdot <d@zerovolt.org> Co-authored-by: James Bourbeau <jrbourbeau@users.noreply.github.com> Co-authored-by: Shaochen Shi <shishaochen_ha@sina.com> Co-authored-by: Anthony D'Amato <anthony.damato@hotmail.fr> Co-authored-by: Vladislav Epifanov <vepifanov92@gmail.com> Co-authored-by: jameskrach <69264125+jameskrach@users.noreply.github.com> Co-authored-by: Hyunsu Cho <chohyu01@cs.wasshington.edu>

SmirnovEgorRu reviewed Apr 23, 2020

View reviewed changes

src/tree/updater_quantile_hist.cc Outdated Show resolved Hide resolved

src/tree/updater_quantile_hist.cc Outdated Show resolved Hide resolved

src/tree/updater_quantile_hist.cc Outdated Show resolved Hide resolved

ShvetsKS marked this pull request as ready for review April 24, 2020 16:12

ShvetsKS changed the title ~~[WIP] Distributed optimizations~~ Distributed optimizations Apr 27, 2020

ShvetsKS force-pushed the master_distiruted branch from f2b92ca to 77734fe Compare April 29, 2020 08:15

SmirnovEgorRu requested changes Apr 30, 2020

View reviewed changes

ShvetsKS force-pushed the master_distiruted branch from 4088972 to 4d6b921 Compare May 6, 2020 12:13

ShvetsKS requested a review from SmirnovEgorRu May 6, 2020 12:15

ShvetsKS force-pushed the master_distiruted branch from 4d6b921 to 4d4b11e Compare May 6, 2020 12:27

trivialfis requested changes May 7, 2020

View reviewed changes

SmirnovEgorRu reviewed May 7, 2020

View reviewed changes

src/tree/updater_quantile_hist.cc Outdated Show resolved Hide resolved

ShvetsKS requested review from trivialfis and SmirnovEgorRu May 8, 2020 18:31

ShvetsKS force-pushed the master_distiruted branch 5 times, most recently from 4e1f301 to f36e5cf Compare May 11, 2020 19:11

SHVETS, KIRILL added 8 commits May 12, 2020 15:43

distributed

c768395

distributed

977a402

distributed

80f1649

distributed

53d41ae

distributed lossguide

f5002ab

extra allreduce was deleted

2e17f8a

cleanup

c37e4cf

cleanup

ecb5f10

SHVETS, KIRILL added 4 commits May 12, 2020 15:43

fix comments, and reduced memory for all reduce

a54153c

comments were applied

b6726b5

strategy pattern was applied

dbe9c1f

tests were added and found bug was fixed(min->max)

605008f

ShvetsKS force-pushed the master_distiruted branch from f36e5cf to 605008f Compare May 12, 2020 12:46

trivialfis reviewed May 17, 2020

View reviewed changes

comments were applied

a9f16f2

ShvetsKS requested a review from trivialfis May 18, 2020 14:02

SmirnovEgorRu reviewed May 18, 2020

View reviewed changes

SmirnovEgorRu approved these changes May 18, 2020

View reviewed changes

trivialfis approved these changes May 19, 2020

View reviewed changes

SmirnovEgorRu merged commit dd01e4b into dmlc:master May 20, 2020

SmirnovEgorRu mentioned this pull request Jun 14, 2020

Low performance with dask #5785

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Distributed optimizations #5557

Distributed optimizations #5557

Distributed optimizations #5557

Distributed optimizations #5557

Conversation

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment