diff --git a/distributed/distributed.yaml b/distributed/distributed.yaml
index 3cf75a23298..2aa291cfd4f 100644
--- a/distributed/distributed.yaml
+++ b/distributed/distributed.yaml
@@ -164,8 +164,13 @@ distributed:
   nanny:
     preload: []             # Run custom modules with Nanny
     preload-argv: []        # See https://docs.dask.org/en/latest/how-to/customize-initialization.html
+
+    # Override environment variables. Read important caveats at
+    # https://distributed.dask.org/en/stable/worker.html#nanny.
     environ:
+      # See https://distributed.dask.org/en/stable/worker-memory.html#automatically-trim-memory
       MALLOC_TRIM_THRESHOLD_: 65536
+      # Numpy configuration
       OMP_NUM_THREADS: 1
       MKL_NUM_THREADS: 1
 
diff --git a/distributed/nanny.py b/distributed/nanny.py
index 1e3304eaad9..5d5a12e6573 100644
--- a/distributed/nanny.py
+++ b/distributed/nanny.py
@@ -87,6 +87,19 @@ class Nanny(ServerNode):
             2. Existing environment variables
             3. Dask configuration
 
+        Note
+        ----
+        Some environment variables, like ``OMP_NUM_THREADS``, must be set before
+        importing numpy to have effect. Others, like ``MALLOC_TRIM_THRESHOLD_`` (see
+        :ref:`memtrim`), must be set before starting the Linux process. So we need to
+        set them before spawning the subprocess, even if this means poisoning the
+        process running the Nanny.
+
+        For the same reason, be warned that changing
+        ``distributed.worker.multiprocessing-method`` from ``spawn`` to ``fork`` or
+        ``forkserver`` may inhibit some environment variables; if you do, you should
+        set the variables yourself in the shell before you start ``dask-worker``.
+
     See Also
     --------
     Worker
@@ -666,6 +679,10 @@ async def start(self) -> Status:
         self.stopped = asyncio.Event()
         self.status = Status.starting
 
+        # Must set env variables before spawning the subprocess.
+        # See note in Nanny docstring.
+        os.environ.update(self.env)
+
         try:
             await self.process.start()
         except OSError:
@@ -811,6 +828,9 @@ def _run(
         Worker,
     ):  # pragma: no cover
         try:
+            # Set the environment variables again. This is to avoid race conditions
+            # where different nannies in the same process set different environment
+            # variables.
             os.environ.update(env)
             dask.config.set(config)
 
diff --git a/distributed/tests/test_nanny.py b/distributed/tests/test_nanny.py
index b908236c4ab..f3d87cfbb1c 100644
--- a/distributed/tests/test_nanny.py
+++ b/distributed/tests/test_nanny.py
@@ -8,7 +8,6 @@
 import random
 import sys
 from contextlib import suppress
-from unittest import mock
 
 import psutil
 import pytest
@@ -275,42 +274,50 @@ async def test_wait_for_scheduler():
     assert "restart" not in log.lower(), log
 
 
-@gen_cluster(nthreads=[], client=True)
+@gen_cluster(client=True, nthreads=[])
 async def test_environment_variable(c, s):
-    a = Nanny(s.address, memory_limit=0, env={"FOO": "123"})
-    b = Nanny(s.address, memory_limit=0, env={"FOO": "456"})
-    await asyncio.gather(a, b)
-    results = await c.run(lambda: os.environ["FOO"])
-    assert results == {a.worker_address: "123", b.worker_address: "456"}
-    await asyncio.gather(a.close(), b.close())
-
-
-@gen_cluster(nthreads=[], client=True)
-async def test_environment_variable_by_config(c, s, monkeypatch):
-
     with dask.config.set({"distributed.nanny.environ": "456"}):
         with pytest.raises(TypeError, match="configuration must be of type dict"):
-            Nanny(s.address, memory_limit=0)
-
-    with dask.config.set({"distributed.nanny.environ": {"FOO": "456"}}):
-
-        # precedence
-        # kwargs > env var > config
+            Nanny(s.address)
+
+    try:
+        # precedence: kwargs > env var > config
+        with dask.config.set({"distributed.nanny.environ": {"FOO": "123"}}):
+            os.environ["FOO"] = "456"
+            async with Nanny(s.address, env={"FOO": "789"}) as n:
+                results = await c.run(lambda: os.environ["FOO"])
+                assert results == {n.worker_address: "789"}
+
+            # Nanny sets the env variables in the Nanny's own process.
+            # Need to manually clean up os.environ afterwards.
+            os.environ["FOO"] = "456"
+            async with Nanny(s.address) as n:
+                results = await c.run(lambda: os.environ["FOO"])
+                assert results == {n.worker_address: "456"}
 
-        with mock.patch.dict(os.environ, {"FOO": "BAR"}, clear=True):
-            a = Nanny(s.address, memory_limit=0, env={"FOO": "123"})
-            x = Nanny(s.address, memory_limit=0)
+            del os.environ["FOO"]
+            async with Nanny(s.address) as n:
+                results = await c.run(lambda: os.environ["FOO"])
+                assert results == {n.worker_address: "123"}
+    finally:
+        del os.environ["FOO"]
 
-        b = Nanny(s.address, memory_limit=0)
 
-        await asyncio.gather(a, b, x)
+@gen_cluster(client=True, nthreads=[])
+async def test_environment_variable_race_condition(c, s):
+    """Nannies need to set the environment variables in their own process so that
+    special variables such as MALLOC_TRIM_THRESHOLD_ are picked up when the Worker's
+    interpreter is started. However, this creates a race condition with multiple Nannies
+    in the same process with different variables. Test that the environment variables
+    are set again inside their own process."""
+    nannies = [Nanny(s.address, env={"FOO": str(i)}) for i in range(8)]
+    try:
+        await asyncio.gather(*nannies)
         results = await c.run(lambda: os.environ["FOO"])
-        assert results == {
-            a.worker_address: "123",
-            b.worker_address: "456",
-            x.worker_address: "BAR",
-        }
-        await asyncio.gather(a.close(), b.close(), x.close())
+        assert sorted(results.values()) == [str(i) for i in range(8)]
+        await asyncio.gather(*(n.close() for n in nannies))
+    finally:
+        del os.environ["FOO"]
 
 
 @gen_cluster(
@@ -585,3 +592,53 @@ async def test_scheduler_crash_doesnt_restart(s, a):
     await a.finished()
     assert a.status == Status.closed
     assert a.process is None
+
+
+@pytest.mark.slow
+@pytest.mark.skipif(not LINUX, reason="Requires GNU libc")
+@gen_cluster(
+    client=True,
+    Worker=Nanny,
+    nthreads=[("", 2)],
+    worker_kwargs={"memory_limit": "2GiB"},
+)
+async def test_malloc_trim_threshold(c, s, a):
+    """Test that the nanny sets the MALLOC_TRIM_THRESHOLD_ environment variable before
+    starting the worker process.
+
+    This test relies on these settings to work:
+
+        distributed.nanny.environ.MALLOC_TRIM_THRESHOLD_: 65536
+        distributed.worker.multiprocessing-method: spawn
+
+    We're deliberately not setting them explicitly in @gen_cluster above, as we want
+    this test to trip if somebody changes distributed.yaml.
+
+    Note
+    ----
+    This test may start failing in a future Python version if CPython switches to
+    using mimalloc by default. If it does, a thorough benchmarking exercise is needed.
+    """
+    da = pytest.importorskip("dask.array")
+
+    a = da.random.random(
+        2**29 // 8,  # 0.5 GiB,
+        chunks=160 * 2**10 // 8,  # 160 kiB
+    ).persist()
+    await wait(a)
+    # Wait for heartbeat
+    while s.memory.process < 2**29:
+        await asyncio.sleep(0.01)
+    del a
+
+    # This is the delicate bit, as it relies on
+    # 1. PyMem_Free() to be quick to invoke glibc free() when memory becomes available
+    # 2. glibc free() to be quick to invoke the kernel's sbrk() when the same happens
+    #
+    # At the moment of writing, the readings are:
+    # - 122 MiB after starting a new worker
+    # - 139 MiB after computing a trivial dask.array collection
+    # - 185 MiB at the end of this test, with MALLOC_TRIM_THRESHOLD_=65536
+    # - 698 MiB at the end of this test, without MALLOC_TRIM_THRESHOLD_
+    while s.memory.process > 250 * 2**20:
+        await asyncio.sleep(0.01)