diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 503857b5..3ad880eb 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -17,12 +17,9 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        version:
-          - '1'
-        os:
-          - ubuntu-latest
-        arch:
-          - x64
+        include:
+          - {version: '1', os: ubuntu-latest, arch: x64}
+          - {version: '1', os: macos-14,      arch: aarch64}
     steps:
       - uses: actions/checkout@v2
       - uses: julia-actions/setup-julia@v2
@@ -42,12 +39,9 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        version:
-          - '1'
-        os:
-          - ubuntu-latest
-        arch:
-          - x64
+        include:
+          - {version: '1', os: ubuntu-latest, arch: x64}
+          - {version: '1', os: macos-14,      arch: aarch64}
     steps:
       - uses: actions/checkout@v2
       - uses: julia-actions/setup-julia@v2
@@ -67,12 +61,9 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        version:
-          - '1'
-        os:
-          - ubuntu-latest
-        arch:
-          - x64
+        include:
+          - {version: '1', os: ubuntu-latest, arch: x64}
+          - {version: '1', os: macos-14,      arch: aarch64}
     steps:
       - uses: actions/checkout@v2
       - uses: julia-actions/setup-julia@v2
diff --git a/src/block_arrays.jl b/src/block_arrays.jl
index dfc47807..dfe737eb 100644
--- a/src/block_arrays.jl
+++ b/src/block_arrays.jl
@@ -297,10 +297,8 @@ function distance_eval_body(d,a::BVector,b::BVector)
     partials = map(blocks(a),blocks(b)) do ai,bi
         distance_eval_body(d,ai,bi)
     end
-    s = reduce((i,j)->Distances.eval_reduce(d,i,j),
-               partials,
-               init=Distances.eval_start(d, a, b))
-    s
+    init_val = Distances.eval_start(d, a, b)
+    reduce_for_distance(d, partials, init_val)
 end
 
 struct BroadcastedBArray{A}
diff --git a/src/p_vector.jl b/src/p_vector.jl
index dbc0cdf7..9ffffd97 100644
--- a/src/p_vector.jl
+++ b/src/p_vector.jl
@@ -1,6 +1,30 @@
 
 function local_values end
 
+function reduce_for_distance(d, partials, init)
+    reduce((i, j) -> Distances.eval_reduce(d, i, j), partials; init)
+end
+
+reduce_for_distance(::Distances.Chebyshev, partials, init) =
+    reduce(max, partials; init)
+    
+for T in (Distances.Euclidean,
+          Distances.SqEuclidean,
+          Distances.PeriodicEuclidean,
+          Distances.Cityblock,
+          Distances.TotalVariation,
+          Distances.Minkowski,
+          Distances.Hamming,
+          Distances.ChiSqDist,
+          Distances.KLDivergence,
+          Distances.GenKLDivergence)
+    # all Distances.metrics except Jaccard, RogersTanimoto, CosineDist, 
+    # RenyiDivergence, BrayCurtis, SpanNormDist, since these either operate with
+    # tuples or have a complex reduction logic.
+    @eval reduce_for_distance(::$(T), partials, init) =
+        reduce(+, partials; init)
+end
+
 function own_values end
 
 function ghost_values end
@@ -1328,10 +1352,8 @@ function distance_eval_body(d,a::PVector,b::PVector)
             return s
         end
     end
-    s = reduce((i,j)->Distances.eval_reduce(d,i,j),
-               partials,
-               init=Distances.eval_start(d, a, b))
-    s
+    init_val = Distances.eval_start(d, a, b)
+    reduce_for_distance(d, partials, init_val)
 end
 
 # New stuff
diff --git a/src/primitives.jl b/src/primitives.jl
index 86726137..a4183186 100644
--- a/src/primitives.jl
+++ b/src/primitives.jl
@@ -806,13 +806,10 @@ function ExchangeGraph_impl_with_neighbors(snd_ids,neighbors::ExchangeGraph)
         is_included_all=map(snd_ids_a, snd_ids_b) do snd_ids_a, snd_ids_b
             all(i->i in snd_ids_b,snd_ids_a)
         end
-        result=false
-        and(a,b)=a && b
-        is_included_all=reduction(and,is_included_all,destination=:all,init=one(eltype(is_included_all)))
-        map(is_included_all) do is_included_all
-            result = is_included_all
-        end
-        result
+        # perform a global AND using built-in & operator (static MPI-op)
+        is_included_all = reduction(&, is_included_all; destination=:all, init=one(eltype(is_included_all)))
+        # extract the single boolean result
+        collect(is_included_all)[1]
     end 
     @boundscheck is_included(snd_ids,neighbors.snd) || error("snd_ids must be a subset of neighbors.snd")
     rank = linear_indices(snd_ids)