Unconstrained Refinement #160

N-Maas · 2023-09-13T13:12:27Z

This PR adds the new unconstrained refinement algorithms from our most recent paper (Parallel Unconstrained Local Search for Partitioning Irregular Graphs). Details can be found there.

Summary:

Unconstrained LP: Performs moves without considering the balance constraint and rebalances after each round.
Unconstrained FM: Computes an approximate penalty for unconstrained moves, rebalancing happens after the localized search. The rebalancing moves are then interleaved with the FM moves to create a global move sequence where balance is immediately restored after each move, thereby maximizing the number of valid prefixes for the global rollback.
High-Quality Rebalancer: Based on a central concurrent priority queue where the priority is computed as the (negative) gain of a move divided by the node weight.

With regards to the code architecture, the new algorithms require the introduction of a new rebalancer interface as well as an interface for the FM strategy. Furthermore, the data for computing approximate penalties is added to the shared FM data.

Results on the medium hypergraph set (k <= 32):

Note that the quality improvement is much more significant on highly irregular graphs (see paper). Specifically, unconstrained has much better quality there than n-level while being almost as fast as the current default configuration. I plan to investigate the results on hypergraphs in more detail within the next week, but I think it makes sense to already open the PR.

Unconstrained has only minimal running time overhead in the multilevel configuration. Consequently, I've added it to the default and quality configurations.

In the n-level setting, the overhead is much more significant. This is because unconstrained refinement has some operations (initialization of penalty estimation and rebalancing) which require scanning the whole graph. The consequence is potentially quadratic total work since this can happen for every batch uncontraction. Therefore, the PR does not change the highest_quality config. A possible solution is to use unconstrained refinement only for the global refinement rounds. This would probably need some significant changes to the n-level code.

Additional notes:

We haven't tested unconstrained refinement during initial partitioning yet
The PR adds tools for converting hmetis/mtx to metis format
~~The PR removes the old rebalancer. There seems little point in keeping it, since the new rebalancer outpeforms it~~
The diff size is somewhat blown up because the PR includes a typo fix (SyncronizedEdgeUpdate -> SynchronizedEdgeUpdate). We could also merge this separately if it simplifies review

This reverts commit a972b3b.

…n whole graph)

codecov · 2023-09-13T13:54:03Z

Codecov Report

Merging #160 (e0a5aed) into master (858c9f0) will increase coverage by 0.72%.
The diff coverage is 91.28%.

❗ Current head e0a5aed differs from pull request most recent head 41003aa. Consider uploading reports for the commit 41003aa to get more accurate results

@@            Coverage Diff             @@
##           master     #160      +/-   ##
==========================================
+ Coverage   77.94%   78.66%   +0.72%     
==========================================
  Files         192      202      +10     
  Lines       18699    19626     +927     
  Branches     7589     8015     +426     
==========================================
+ Hits        14575    15439     +864     
- Misses       4124     4187      +63

Files Changed	Coverage Δ
...ar/partition/coarsening/multilevel_uncoarsener.cpp	`61.53% <ø> (ø)`
...ahypar/partition/coarsening/nlevel_uncoarsener.cpp	`78.00% <ø> (ø)`
...-kahypar/partition/coarsening/nlevel_uncoarsener.h	`100.00% <ø> (ø)`
.../initial_partitioning/greedy_initial_partitioner.h	`100.00% <ø> (ø)`
...rtitioning/label_propagation_initial_partitioner.h	`100.00% <ø> (ø)`
...ar/partition/refinement/gains/cut/cut_gain_cache.h	`78.09% <0.00%> (-1.52%)`	⬇️
...gains/cut_for_graphs/cut_gain_cache_for_graphs.cpp	`93.93% <0.00%> (ø)`
...t/gains/cut_for_graphs/cut_gain_cache_for_graphs.h	`71.62% <0.00%> (+1.75%)`	⬆️
...artition/refinement/gains/soed/soed_gain_cache.cpp	`75.28% <0.00%> (ø)`
.../partition/refinement/gains/soed/soed_gain_cache.h	`77.04% <0.00%> (-1.29%)`	⬇️
... and 49 more

... and 8 files with indirect coverage changes

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

kittobi1992

Lgtm. Only some minor changes. If these issues are addressed, feel free to merge (after fixing the windows build) ;-)

kittobi1992 · 2023-09-18T06:06:48Z

config/default_preset.ini

 # main -> refinement -> fm
-r-fm-type=kway_fm
+r-fm-type=unconstrained


I would rename this to "unconstrained_fm"

kittobi1992 · 2023-09-18T06:08:19Z

config/old_configs/default_preset.ini

@@ -57,7 +57,7 @@ i-r-fm-iter-moves-on-recalc=true
 # main -> initial_partitioning -> refinement -> flows
 i-r-flow-algo=do_nothing
 # main -> refinement
-r-rebalancer-type=simple_rebalancer
+r-rebalancer-type=rebalancer


did we remove the "simple_rebalancer"? These configs should be not modified as they refer to the state in our dissertations.

I updated the PR so that it doesn't remove the simple rebalancer anymore

kittobi1992 · 2023-09-18T06:14:44Z

tests/partition/refinement/rebalance_test.cc

@@ -1,123 +0,0 @@
-/*******************************************************************************
- * MIT License


Is there another test for the rebalancer?

kittobi1992 · 2023-09-18T06:25:10Z

mt-kahypar/datastructures/sparse_map.h

@@ -119,6 +119,14 @@ class SparseMapBase {
    return _dense[index].value;
  }

+  Value operator[] (const Key key) const {


This is a confusing semantic of the []-operator. Can we use at(key) function instead?

using getOrDefault(key) now

kittobi1992 · 2023-09-18T06:32:58Z

mt-kahypar/partition/refinement/fm/fm_commons.h

+// Contains data required for unconstrained FM: We group non-border nodes in buckets based on their
+// incident weight to node weight ratio. This allows to give a (pessimistic) estimate of the effective
+// gain for moves that violate the balance constraint
+struct UnconstrainedFMData {


this looks to me more like a class rather than a struct.

kittobi1992 · 2023-09-18T06:54:01Z

mt-kahypar/partition/refinement/fm/multitry_kway_fm.h

+                                                  const std::vector<HypernodeWeight>& max_part_weights,
+                                                  vec<vec<Move>>& rebalancing_moves_by_part);
+
+  void insertMovesToBalancePart(const PartitionedHypergraph& phg,


I prefer the term "block" over "part"

kittobi1992 · 2023-09-18T06:56:18Z

mt-kahypar/partition/refinement/fm/strategies/gain_cache_strategy.h

   *
   */

-class GainCacheStrategy {
+class LocalGainCacheStrategy {


move this in a separate file

kittobi1992 · 2023-09-18T06:57:57Z

mt-kahypar/partition/refinement/i_refiner.h

@@ -72,4 +72,42 @@ class IRefiner {
                          const double time_limit) = 0;
 };

+class IRebalancer: public IRefiner {


move this into "i_rebalancer.h"

kittobi1992 · 2023-09-18T07:01:22Z

mt-kahypar/partition/refinement/rebalancing/rebalancer_v2.h

+
+
+template <typename TypeTraits, typename GainTypes>
+class RebalancerV2 final : public IRebalancer {


I do not like the naming. Suggestion: "AdvancedRebalancer". Also I would still keep the old rebalancer for our old configurations.

kittobi1992 · 2023-09-18T07:02:54Z

mt-kahypar/partition/refinement/fm/strategies/unconstrained_strategy.h

+
+
+template<typename TypeTraits, typename GainTypes>
+class UnconstrainedStrategy: public IFMStrategy {


would split this into two separate files "unconstrained_fm_strategy.h" and "local_unconstrained_fm_strategy.h"

+ simplification in fm strategy

This reverts commit facb8be.

N-Maas added 30 commits August 29, 2023 16:26

refactor gain computation for use by jet refiner

a972b3b

jet context and implementation first phase

856da31

command line / factories integration

6caf0fb

rebalancing and rollback

d8810e7

support for randomized vertex locking

33c0f46

rework jet refiner to use multiple internal rounds

8046690

jet refiner fixes

a6f58be

use factory for rebalancer

c998bdc

jet rebalancer initial implementation and tests

3d17dcc

register jet balancer, use global rebalancer for jet refiner

b9483b6

improved GraphToHgr tool

ef19c63

initial implementation of unconstrained fm strategy

35e9cf5

fixes for merge to master

c33a672

unconstrained FM strategy and integration to kway FM

af10979

refactoring for shared data

a359d32

edge case

b64a361

add even more functionality to unconstrained data

c7eb45c

implement global rollback for unconstrained FM

3253601

minor

2324016

optimization: precompute incident weights

1fa9cf4

minor

20d6ee8

introduce IRebalancer interface and implement it for JetRebalancer

882c9ce

add option to allow part weight scaling as in global rollback

1603327

implement interleaving of move sequence

c46d04e

apply strategy for interleaved rebalancing moves in kway fm

bb690c2

jet rebalancer return moves

273265b

add cooling fm algorithm (unconstrained with increasing penalty)

61de936

add options for including border nodes

e8229b9

JET: option for fixed number of rounds and rollback after each iteration

62b656e

special case for penalty 0 in unconstrained FM

704047b

N-Maas added 16 commits September 5, 2023 14:23

[LP] fix gain cache updates & refactoring

0781e22

fix FM test

28884a8

use gain cache constant to determine whether entries must be recomputed

01445f7

minor

314c74b

minor

ac76973

Revert "refactor gain computation for use by jet refiner"

d4e6a0a

This reverts commit a972b3b.

update includes/license

b30c8a6

[rebalancer] allow to return moves in single vector

73ba8ca

[LP] final refactoring: compile times

8810fa5

remove old rebalancer

facb8be

test fixes

ac66406

[LP] use lazy initialization for n-level/localized variant (don't sca…

9966303

…n whole graph)

cleanup TODOs

b915a25

use unconstrained refinement in default and quality preset

755ff05

minor fixes

0f1cb8e

update readme

4097024

fix windows build

14ee858

larsgottesbueren requested a review from kittobi1992 September 13, 2023 17:41

kittobi1992 requested changes Sep 18, 2023

View reviewed changes

N-Maas added 8 commits September 18, 2023 11:12

small changes from review

b42dce3

split files for fm strategies and i_rebalancer

6a6804c

+ simplification in fm strategy

restore old rebalancer

43a0a0f

This reverts commit facb8be.

rename rebalancers

a28240c

add tests for new rebalancer

f1197ae

fix

cd719d6

minor

4460e6e

[rebalancer] don't assume phg.k() == context.partition.k

41003aa

N-Maas merged commit 9bbc98b into master Sep 19, 2023
15 checks passed

N-Maas deleted the ufm-cleanup branch November 30, 2023 12:33

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Unconstrained Refinement #160

Unconstrained Refinement #160

N-Maas commented Sep 13, 2023 •

edited

Loading

codecov bot commented Sep 13, 2023 •

edited

Loading

kittobi1992 left a comment •

edited

Loading

kittobi1992 Sep 18, 2023

kittobi1992 Sep 18, 2023

N-Maas Sep 18, 2023

kittobi1992 Sep 18, 2023

N-Maas Sep 18, 2023

kittobi1992 Sep 18, 2023

N-Maas Sep 18, 2023

kittobi1992 Sep 18, 2023

kittobi1992 Sep 18, 2023

kittobi1992 Sep 18, 2023

kittobi1992 Sep 18, 2023

kittobi1992 Sep 18, 2023 •

edited

Loading

kittobi1992 Sep 18, 2023

		@@ -1,123 +0,0 @@
		/*******************************************************************************
		* MIT License



		template <typename TypeTraits, typename GainTypes>
		class RebalancerV2 final : public IRebalancer {



		template<typename TypeTraits, typename GainTypes>
		class UnconstrainedStrategy: public IFMStrategy {

Unconstrained Refinement #160

Unconstrained Refinement #160

Conversation

N-Maas commented Sep 13, 2023 • edited Loading

codecov bot commented Sep 13, 2023 • edited Loading

Codecov Report

kittobi1992 left a comment • edited Loading

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

kittobi1992 Sep 18, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

N-Maas commented Sep 13, 2023 •

edited

Loading

codecov bot commented Sep 13, 2023 •

edited

Loading

kittobi1992 left a comment •

edited

Loading

kittobi1992 Sep 18, 2023 •

edited

Loading