diff --git a/ortools/glop/BUILD.bazel b/ortools/glop/BUILD.bazel
index da89eef6ffe..da97e5d2a8b 100644
--- a/ortools/glop/BUILD.bazel
+++ b/ortools/glop/BUILD.bazel
@@ -265,6 +265,7 @@ cc_library(
         "//ortools/lp_data:lp_utils",
         "//ortools/lp_data:scattered_vector",
         "//ortools/util:stats",
+        "//ortools/util:time_limit",
     ],
 )
 
@@ -330,6 +331,7 @@ cc_library(
         "//ortools/lp_data:lp_utils",
         "//ortools/lp_data:scattered_vector",
         "//ortools/util:stats",
+        "//ortools/util:time_limit",
     ],
 )
 
diff --git a/ortools/glop/dual_edge_norms.cc b/ortools/glop/dual_edge_norms.cc
index 4e81385ddf5..db19bc3d0ae 100644
--- a/ortools/glop/dual_edge_norms.cc
+++ b/ortools/glop/dual_edge_norms.cc
@@ -105,7 +105,7 @@ void DualEdgeNorms::UpdateBeforeBasisPivot(
 
     // Avoid 0.0 norms (The 1e-4 is the value used by Koberstein).
     // TODO(user): use a more precise lower bound depending on the column norm?
-    // We can do that with Cauchy-Swartz inequality:
+    // We can do that with Cauchy-Schwarz inequality:
     //   (edge . leaving_column)^2 = 1.0 < ||edge||^2 * ||leaving_column||^2
     const Fractional kLowerBound = 1e-4;
     if (output[e.row()] < kLowerBound) {
@@ -121,13 +121,23 @@ void DualEdgeNorms::UpdateBeforeBasisPivot(
 void DualEdgeNorms::ComputeEdgeSquaredNorms() {
   SCOPED_TIME_STAT(&stats_);
 
+  // time_limit_->LimitReached() can be costly sometimes, so we only do that
+  // if we feel this will be slow anyway.
+  const bool test_limit = (time_limit_ != nullptr) &&
+                          basis_factorization_.NumberOfEntriesInLU() > 10'000;
+
   // Since we will do a lot of inversions, it is better to be as efficient and
   // precise as possible by having a refactorized basis.
   DCHECK(basis_factorization_.IsRefactorized());
   const RowIndex num_rows = basis_factorization_.GetNumberOfRows();
-  edge_squared_norms_.resize(num_rows, 0.0);
+  edge_squared_norms_.resize(num_rows, 1.0);
   for (RowIndex row(0); row < num_rows; ++row) {
     edge_squared_norms_[row] = basis_factorization_.DualEdgeSquaredNorm(row);
+
+    // This operation can be costly, and we abort if we are stuck here.
+    // Note that we still mark edges as "recomputed" otherwise we can runs into
+    // some DCHECK before we actually abort the solve.
+    if (test_limit && time_limit_->LimitReached()) break;
   }
   recompute_edge_squared_norms_ = false;
 }
diff --git a/ortools/glop/dual_edge_norms.h b/ortools/glop/dual_edge_norms.h
index bf5ba3f7833..e0cbc7d795b 100644
--- a/ortools/glop/dual_edge_norms.h
+++ b/ortools/glop/dual_edge_norms.h
@@ -23,6 +23,7 @@
 #include "ortools/lp_data/permutation.h"
 #include "ortools/lp_data/scattered_vector.h"
 #include "ortools/util/stats.h"
+#include "ortools/util/time_limit.h"
 
 namespace operations_research {
 namespace glop {
@@ -102,6 +103,8 @@ class DualEdgeNorms {
     parameters_ = parameters;
   }
 
+  void SetTimeLimit(TimeLimit* time_limit) { time_limit_ = time_limit; }
+
   // Stats related functions.
   std::string StatString() const { return stats_.StatString(); }
 
@@ -130,6 +133,7 @@ class DualEdgeNorms {
 
   // Parameters.
   GlopParameters parameters_;
+  TimeLimit* time_limit_ = nullptr;
 
   // Problem data that should be updated from outside.
   const BasisFactorization& basis_factorization_;
diff --git a/ortools/glop/revised_simplex.cc b/ortools/glop/revised_simplex.cc
index 9961a5e4742..0ed83b1e366 100644
--- a/ortools/glop/revised_simplex.cc
+++ b/ortools/glop/revised_simplex.cc
@@ -221,6 +221,8 @@ ABSL_MUST_USE_RESULT Status RevisedSimplex::SolveInternal(
 
   SOLVER_LOG(logger_, "");
   primal_edge_norms_.SetTimeLimit(time_limit);
+  dual_edge_norms_.SetTimeLimit(time_limit);
+
   if (logger_->LoggingIsEnabled()) {
     DisplayBasicVariableStatistics();
   }